Neither of those are issues with AMD CPUs. Pins on mobo and are covered till you are prepared to install. Marked corners on CPU to line up with marked corner on mobo.
My Slot 1 P3 was my favorite PC. Started as a P3 450, ended as a P3 933 that I acquired at a flea market for $15 Took a bit of BIOS tweaking for my motherboard to even support the damn thing, but it was a massive upgrade.
I didn't particularly like the slot/module form factors. Because they could get wiggly after enough insert/remove cycles. They seemed to still work fine but I don't believe wiggly parts are truly reliable.
I can confirm that running a P55C at 120x2.5 makes for a very fast 300 MHz Pentium. It beat my 300 MHz Klamath running 66 MHz bus at pretty much every metric in dosbench.
That's quite interesting. That would mean a 166MHz non-MMX Pentium overclocked to 120FSB, I wonder how stable that is, and how stable the ISA, PCI buses are. What's the heat output like?
@@thebayandurpoghosyanshow no it's an MMX 233 with a lowered multiplier. At 120x2.5 it wasn't perfectly stable but 100x3 seemed stable. Motherboard is Gigabyte 5AA so Super Socket 7 and quite capable. Heat is manageable with a reasonably large SS7 heatsink and fan.
@@agsel that makes more sense, because while the CPU might be able to run at 300MHz internally (to my knowledge, the laptop market actually got 300MHZ P1MMX CPUs), running the FSB at 120MHz, well, I shudder at the thought of how many more variables it introduces. It would be cool if you made a thread on Vogons with your findings :))
Yes, the slot 1 cpu's. I remember playing with those. I had a socket370 to slot 1 converter. I had an Asus P4b-F, with the venerable 440BX chipset. Originaly it was fitted with a PII 350 mhz with a 100mhz FSB. Later, I upgraded it to a PIII 800mhz. I remember searching high and low for the 100mhz FSB version. So that I could install it in my P4b-F. I also had a Thermaltake super orb CPU cooler installed, becouse other coolers would be to large to fit and hit components on the motherboard. If memory serves, the slot-converter had special brackets, so that you could mount it properly and wouldn't put to much stress on your socket/motherboard. Unfortunetly I don't have that setup anymore. It's hard to believe it's almost 25 years since I had it.
My first slot 1 CPU was a 233Mhz Pentium II that came in an IBM Aptiva tower, but my first system I built myself was a Celeron 300A that was overclockable to 450Mhz out of the box with no additional cooling required. Those were the days!
@@rootbeer666 At one time I ran a ABIT BH6, Pentium II 450, 384MB PC-100, and a Voodoo 3 3000. Later on I ran a ASUS P3B, a Celeron 1Ghz on a slocket adapter, 1GB PC-100 with a Radeon 8500. (I had hopes that ATI would resolve the driver issues a lot faster than they actually did.)
Wow! I knew the answer vaguely involved cache, and now I have a much stronger grasp on why CPU hardware was so weird for a few years before settling back down into sockets again. I’d enjoy more videos like this as you want to make them - they contextualize and deepen my appreciation for the benchmarks, and they are good fodder for future videos. I really look forward to the slotted Pentium/Xeon banchmark video. I’d also be interested to see a comparison video showcasing the performance changes between Intel’s and AMD’s last-gen slot and first-gen socketed CPUs, and a comparison video of the Pentium Pro, Slot 1, and Socket 370 Pentiums (compensating for speed/architecture differences as much as possible) showing the performance delta that cache packaging alone makes.
Ah... I had been so focused on the cache characteristics of CPUs with the same packaging, that I didn't even think about slot 1 vs. socket 370 or slot A vs. socket A. I would guess that the main difference there would be changes in chipset or memory technology... but it's definitely worth checking that assumption. I may just fold that in with the other comparisons.
vid is halfway over, subscribing right now. this may be dry, and maybe i just ate my intellectual-bandwidth-Wheaties this morning or something...but I'm glued.
In highschool I used to work at a company called Microway that built servers and workstations. I remember when Digital went out of business the owners came back from an auction with a bunch of those slot socketed alphas. I remember being so intrigued by them. The Owners were a couple. The wife was all business, but the husband was a huge tech nerd. I remember him being super excited because he had gotten all the technical and design documents and schematics for the Alpha processor, along with the sales rights from API.
Your videos are like the Great Library of Alexandria, but, thank God they are not burnt. I love them. The amount of info and specific details you are sharing is GOLD. Thank you so much.
The question now is that considering the late 90'$ are back in the forms of "chiplets" (multi chips 2D dyes) CPUs, are we bottlenecking again development wise?
yes and no the reason we can do advanced packaging now is that complex packaging technology has been advanced so it has the yields and density to be fast, efficient, and cheap enough to be worth it in many cases vs making bigger chips, as at the same time the price per transistor has been stagnating or even going up for advanced process nodes.
I just bumped into your channel but I'm really impressed with your depth of knowledge and storytelling ability. I learned things I didn't know I didn't know so I hope you'll keep doing contents like this!
Just stumbled on your channel from my UA-cam recommendations and while I haven’t seen your other videos as a point of comparison I really liked this conversational format!
I liked Slots probably because I used other systems with slot-like CPUs many years before Intel ever used them. The Acorn Archimedes A540 (1990) used slots for it's CPU and RAM which made them easily replaceable/upgradable even by non-technical people. This continued on in the Risc PC (1994) where it came with 2 CPU slots and you could use 2 ARM CPUs or an ARM and an Intel x86 CPU, and there were multiple speeds of ARM and then later StrongARM CPUs available, all very easily upgradable due to the slot-like cards they were on. They just weren't packaged in big lumps of plastic to hide what was inside.
Hi I am an IT-Technician in training and I'd love to hear you talk more indepth about cache and its many mysteries. Great channel, very glad I found it :D Thoughts: This Format doesn't work for everyone but something about your (enthusiasm or knowledge or both) makes it really fly by. More Please !
I loved overclocking the slot Athlons. It took some fine soldering skills, but it was fun. This channel is invaluable for younger enthusiasts who want to learn about how things developed. Awesome channel.
I had known a little about them having on cpu board cache being an advantage. I am glad I stuck around for an hour to watch this entire video. Thank you for being fairly in depth with all of it. I had wondered why they didn't continue further. Still, I kind of miss the era of computing and being able to use old hardware. It was interesting times indeed. I wish I had a slot one or preferrably a slot A system with a 1ghz cpu. I don't have the time, or the room to keep them around sadly. I do miss the era. I remember having over 1ghz systems and could pick up these older computers for cheaper at thrift stores. I remember having old systems where I could install 95, 98, 2000 and xp on them and they were still functional. Had a 733mhz slot 1 pentium 3 to use and I forget what my brother had, 1.4ghz pentium 3? It wasn't my main system, but I could go over to my brothers and we could play age of empires or other games on these systems. I just sort of miss the days of when you could switch between windows versions and systems and they would still be functional, you could still browse the web and use them all as daily drivers. Now you just have windows 10 or 11 and they try to kill off old hardware as much as possible. Still, it's why I love linux and it mixes with old hardware. If I have a laptop, it surely is getting linux on it.
I’m new to this channel and I absolutely loved the format. At first I was apprehensive about the length but when everything about cache speeds clicked it was so worth it :p I’d love to hear more tales like this
THIS WAS SUPER FUN. I remember being really young and my dad would bring home wiped computers from his job at Polyclad. it was cheaper to give them away to employees than to scrap them at that time and i got a bunch of stuff to mess with. old servers were a complete mystery to me and we didnt have internet or access to server os in any way so i just took them apart and was blown away by the internals, but i remember getting some of these slot cpu based pcs and getting them running just to play roller coaster tycoon but i always wondered why the cpu was different in those machines. Now I know! thanks so much i loved the in depth educational content and look forward to more from you.
Cool video, and yes I subscribed immediately, but I'm surprised you didn't really mention the big issue with a wider bus: namely that all the traces have to be the exactly the same length, and the tolerances get tighter as speeds increase.
Wasn't the video long enough?!? 🤣 I had forgotten about that aspect. Routing an increasing number of traces at all becomes absurdly difficult. Routing them well and within the other limits... forget about it.
First time seeing your channel, awesome retrospective. Having not seen the inside of a PC until my teens, I remember wondering what the heck these were when I saw them on the internet. Your level of detail is immaculate!
Intel had an idea for the slot package to integrate separate DSP accelerators, but that idea died with MMX and SSE, and the form factor was only useful until the L2 cache could be economically put on the CPU die. AMD's short adventure with the slot format had some theoretical possibility to cram up to 8MB of L2 clocked at 1/3 CPU speed, but they were also in a hurry to get rid of the expensive packaging. All this happened alongside a very fast progress in silicon manufacturing progress -- the migration to 300mm wafers and the steady transistor node shrinking made monolithic die integration the only economic option.
it would an interesting video, why memory is progressing so slow compared to CPUs and GPUs and why are we still using DDR memory technology that was invented in the late 90s over other memory technology like rambus.
RAMBUS vs everything else is quite another topic. My understanding is that RAMBUS has a couple issues. It has a huge amount of bandwidth, but this comes at the expensive of increased latency. In addition, the way the circuits work internally, a lot of more of the chip is active during memory transactions causing a lot more heat to be generated. RDIMMs had heat spreaders *years* before other RAM modules did.
Honestly memory is so dumb. I remember when I learned how dynamic ram actually worked and I just facepalmed so hard. It's literally a giant array of buckets with holes in them and you need to have someone run around non stop and refill any bucket that still has water in it to prevent data corruption. Hence the need for all the massive timing delays in modern memory configurations. So much time is just being spent refilling buckets of water. You would think we would have found a better way to do it by now but nope. We have (sram) but nothing cheap enough anyways.
@@JathraDH It comes down to cell size and component count. A DRAM cell needs only a couple of transistors and a capacitor. An SRAM cell needs anywhere from 5 to 9 transistors apiece. Multiply that by however many millions or billions of bits you want to store. If there was anything simpler that worked at that scale, it's not been found, yet.
great stuff, this would have been right before i started using computers as a young one and its really a treat to learn more intimately about what happened in the 90s with computing from someone who seems to really have a firm grasp and an affinity for explaining things clearly.
Hey, I'm a first time viewer so I can't say much about the format. I must say I am kinda impressed because I have ADHD and usually can't stay on one video for longer than 20 mins unless it's very interesting! So there you go mate. I was a bit intrigued when you mentioned that you wanted slot adapters. I just made a fast search and ebay has some. I live in Sweden and I'd love to help you out if there is something I can do to help you out. The prices for the slot adapters seem very reasonable to me. Have a great one!
Socket 370 to slot 1 adapters for Celerons and some Pentium III CPUs are pretty common, and I have several. What is very rare are socket 8 to slot 1 adapters to use a Pentium Pro on a slot 1 motherboard. I've never seen one IRL.
There is an audio glitch at 13:33 in the video almost sounds like a skipping DVD or CD. Other then that great video loved the information and subscribed
I started with Pentium 2, 233Mhz that a friend built for me. A couple years later, he put together an Athlon 850Mhz system with a CPU like that. However that system had a crashing problem. After about an hour of use, or 1-2 minutes of use if I ran a game that used the TNT graphics card in 3D mode, it would instantly crash to a red screen, and I'd have to manually power it off and reboot. Then it would be fine for about another hour (or 1-2 minutes of game time). I eventually stopped using it and went back to the P2, until I got a P4. I remember that CPU unit had a small card edge connector on top of it, and my friend had plugged a small board into it to over-clock the CPU. Yes, I tried it without that board and it still crashed.
I've got on the PC of my mum back then, it was the MMX 233MHz then years later I've got a P3 866MHz (non-slot Coppermine) PC from my uncle - I was mindblown HAHA XD
I remember those slot CPU days. Good to hear why it was done. It would be interesting to hear about how this applies to GPUs, which have their wide memory busses and memory on the same daughterboard.
Owner of the FIRST 1 Ghz AMD Thunderbird Slot A processor back in the day, it was my first own computer and semi-built (bought a kit from Tigermax, haha), so many great memories with that computer, it never died and I upgraded from it some years later. Thanks for the trip down memory lane man!
Nice! My brain capacity has mightily increaseth, many complex things from the past have been explaineth and many praise must be given! Also, the format is great. I re-formatted my RAID-6 floppy disk array according to your specs and performance is great - mostly because this array exists only on the outer fringes of my imagination, but probably for some other reason too. Instant subscribe, much recommended ultra-niche-(cache!)-nerd-stuff, more please!
One of my students was asking about similar things and I was explaining about CPU structure. At the time, I didn't even want to touch on slots... 😕 I always had more issues with them than any socket board. At the time I chocked it up to some people messing up the pins on the CPU chips, whereas slot CPUs didn't have that issue. I may have been completely wrong - obviously - but that's my thinking at the time. Many things had changed (some good, some bad). Older systems had all the I/O and processor/memory on cards and only the bus with slots as the "motherboard". Then came primitive motherboards with just enough to boot the system. Then I/O was gradually added to the motherboards. Floppy drives mysteriously went from up to eight down to two connections. Hard drives went from various settings and RLL complications to IDE making things very simple (so surprising). IDE also added tape backup and CD/DVD/BD. Parallel and serial yielded to USB. During this time we went from 16-bit to 32-bit to 64-bit in basic home systems and laptops. SATA and NvME. Somehow we lost decent sound cards and picked up USB sound "cards" and devices. All manner of USB devices over the USB connection that can then pass to the appropriate device driver (like USB HDDs/USB sticks that pass to the IDE drivers to act as IDE mapped devices). Along the way, devices became so different, processors so different, mobos so different that older OS platforms don't work on modern hardware without virtualization. Yet, this knowledge is disappearing daily as schools churn out students without any hardware understanding and even placing students into programming courses with no Structured Programming preliminary education. Students want to graduate immediately and expect they'll be ready to work with almost no experience. If one of my students expresses an interest beyond the limited aspect of class and wants to know why things are the way they are, I take the time to explain it and to send them to channels like yours for even more info. Thanks for taking the time making/editing this video!
When I was briefly teaching computer graphics, I noticed that nobody has the faintest idea how hardware works. When I had to explain bit shifting and bit operations to the students, I almost had to cry myself to sleep. :(
Back in the 1990's I had a job as the in-house computer tech at a multi-million dollar store. I quickly cleared the large (several dozen) backlog of computers needing repair and started taking on more responsibility: Ordering, on-site repair and sales (we had an in store salesman and the large company/government sales, network install and his assistant) and eventually some of the smaller on-site networking. The in-store salesman handled invoices and first interviews. We had a policy to always interview. One day a Guy came in for a computer tech job, the in-store salesman punted to me. I asked the Guy to tell me about himself. He'd just graduated college with a computer tech "degree" from ITT. So I had him come to my work area where I opened up a computer. I explained that I was going to start simple and get more complex to get an idea of his knowledge. He was very excited and said that thus is what he trained for. I asked him to point out the CPU and I shit you not his response was "Can you give me the schematic diagram?". I figured that the inside salesman had set me up. This Guy could pick out the CPU, memory, hard drive, I/O board, video card and had no idea what ISA stood for. After he left the inside salesman wasn't laughing and said he was on the phone the whole time. Then a few weeks later another ITT graduate came in. I shit you not, it was the same. Being pretty confident that I wasn't being setup I asked "What do you need a schematic for to find the CPU?" He responded "So I can trace out where the CPU is." Huh? "How many motherboards have you seen?" He responded "We used about 20 motherboard schematics!" I said "How many actual motherboards have you seen?" He replied "This one, but I just need the schematic". I said "Touch the motherboard." He couldn't. Feeling bad for the Guy, I told him "We might have a few dozen schematic diagrams for motherboards, most of which are several year old Compaq servers. It is very rare that we repair circuit boards because of how long it takes to get schematic diagrams and how cheap computer parts have become. The store used to have a board level repair tech but it was un-profitable. He would spend hours to 'repair' a component. The store charges $40/hr for that work, but we can't charge for his research time or for him to source parts. A new, quality PC motherboard is about $100 retail. Even if we only charge $80 labor plus parts to repair does that make sense? It is likely the new motherboard will have better features and include a warrenty. We warrenty the repair for 30 days and a good portion had additional 'hidden' problems that have to be fixed at no cost. The reason I know this is that we contract that tech for specific repairs but only for expensive equipment or industrial equipment that can't be replaced." "Video cards are generally between $30 and $80 retail, I/O cards $25. It costs more money to diagnose the electrical issues for these components than to buy a new, upgraded component. The cost to repair the component is usually several times the cost to replace with new." "This means a computer tech must be able to immediately identify the components of a computer and have the skill to diagnose which parts need replacement in order to quickly give a customer a detailed list and a price quote." His answer was "That's not what they trained us." I said "Then find a job where they repair components but don't waste your time at computer stores. You need to find an industrial facility and get an electronics repair tech job." He says "I'be got a degree for computer repair, that's what I want to do and you have no idea what you are talking about!" Every few months another ITT graduate would come in without ANY knowledge about computers and always request a schematic. A couple years later I took a class at the local community college. I spoke to the professor about ITT. He laughed and said "yeah, we get ITT grads coming to get a real degree. They are exactly as you describe and usually drop out because we recommend that they get an electronics degree because of their insistence on needing a schematic for everything."
I suspect that a lot of people knew about workstation or Power Mac CPU cards without making the connection. I never thought about it until recently, and that's mostly because I've been working with workstation CPU boards quite a bit over the last couple years. :)
I knew all this going in but still great video! Very entertaining and subscribed. Celeron 300A is truly the most legendary CPU Intel ever made LOL. I still have both of mine and their ABIT overclocking boards in my closet. Thank you for the trip down memory lane. Honestly I kind of miss the slot processors, they were quirky and felt special somehow. Oh, one MINOR nitpick on your bandwidth analogy. Doubling the diameter of a tube does not increase the throughput by 2x, it increases it by 4x as circular area is PI * R^2, 1x1 = 1, 2x2 = 4, 4x4 = 16, etc. Doubling the diameter won't increase it by 4x but its still much more than 2x. Your point got across just fine but it did still trigger my OCD a bit.
So great hearing you explain this. Hopefully some of the younger generation will watch this and realize what we had to go through back in the 90s. I started with an XT back in the mid 80s, a clone.. but the first system I build for myself was a 386sx25 100mb hdd 4mb ram.. yeah it was a dog but SO much better than the XT. Then I put together a 486dx4 100, that thing was light years ahead of the 386. I ran that for so long I NEVER built myself a Pentium.. my buddy grabbed the Pentium 233mhz slot, or I think it was maybe a 266 it was SO long ago.. but man I ran that 486 for so long then and kinda got busy in life, didn't really mess with it for a while, then I took a giant leap and built an Athlon XP 1.3ghz system. That thing was like light years ahead of the 486. Years too, that was probably 2003 when I built that.. crazy how far they've come in SO little time.. since 2019 I've been on AM4, went from a 2600x to a 5800x3d (staying with am4, I already have 32gb ram and all the other fun stuff I need, no reason to go ddr5, why? PCIE4 should be enough for the next few years if not longer) anyway.. man we have come a long way.. I can't wait to see what the next evolution of CPU structure is.. we've come to Moore's law now, I think we're at 3nm on CPUs, Intel isn't even measuring by nanometers anymore.. I'm curious.. I know chiplets, 3d architecture, organic material.. who knows but I know it's going to be big!!
Very interesting video who has been gaming since the Atari 2600. Love hearing about how computers have evolved over the years. To be completely honest my first serious home computer was around 2008 when my broher-in-law got me playing World of Warcraft. Up until then I used my consoles (NES, SNES, Genesis, Dreamcast, PS1, & PS2) for gaming and my computer was for everything else. I've never looked back. About every 5 years I upgrade my PC with somethign that keep me pretty close to the "bleedign edge" so I can play modern titles at max, or near max, setting and not have unbearable lag or latency. Thank you for your video. Going to check out the rest of your catalog. Btw, I do also enjoy computer repair videos and by what you said about your usual format at the conclusion of this video I suspect I will be watching more of your video. Like, subscrived, AND commented. Have a good day.
Hi Ian. I enjoyed every minute of this video, although I watched it for three days, as it is not a video to do "something else" while watching, but attending, understanding and visualizing in (my own) memory... It brought to my memory some weirdnesses on caches from my past, where the only way to optimize some math crunching processes was turning down the caches off, although you couldn't modify L1 always. I will definitely keep this video and revisit it when playing with caches again, as it brought me nice (and bitter) memories. A video on cache performance running on "weird stuff" (with weird code to force cache substitution, if possible) would be much appreciated!!!
My dad upgraded from his 1040STF to a Performa 475. The software "MagiC Mac" was the main reason that he switched. We still have a LC-PDS interface with serial and parallel ports that was patched into MagiC Mac to work directly with original printers and other hardware. I have lost the driver/patch floppy, though and cannot get the card to work anymore.
I loved slot based CPUs, they were super easy to assemble as long as you kept the stock cooler, I still have the fondest of memories of these platforms.
Nice, refreshing vidoe! I only can say good things about slot1 CPUs, my 266MHz celeron used to operate at 450 to 500MHz which was quite fine overclock for any day standards and considering the price of PII... Still to this day I have a drawer full of older CPUs, including a few slot1s. Brings memories.
WOW! 1 hour explanation for slot CPU! Insane! But appreciate your dedication! Just way way too long for someone just wanted to know why slot CPU and nothing more. But this is masterclass in CPU architecture and goes far beyond the video title.
Going back to DEC, there are some systems with the Alpha 21264, like the DEC/Compaq AlphaServer DS20 that I have, which have a type of "slot based" CPU configuration, just like the Pentium II, Pentium III, and Athlon CPUs (Like you mentioned in the video). The CPU card in my AlphaServer contains the CPU soldered to the card, along with a VRM, cache memory, and a firmware ROM. The 500MHz CPU card that I have has 4MBytes of cache on the card. Because there is a VRM on the CPU card, it has a separate plug to provide +5VDC and Ground to the CPU card directly. I found it kind of interesting that they would include a firmware ROM on the CPU card too, rather than on the mainboard. My system supports two of these CPU cards, although I only have one installed currently. They are getting really difficult to find. One problem is that I can't just slot in one of the 833MHz CPU cards from the DS25 or DS40, for example. They made the CPU cards incompatible across systems. I think I recall a 633MHz Alpha is the fastest card I can put in my particular system. I'm not entirely sure why they did this, and why the CPUs were soldered right in to these CPU cards, but my guess is that it forces you to go to their newer model of AlphaServer if you want the faster CPUs.
That is interesting. I was not aware of compatibility issues with the slot B Alphas. My friend in the AlphaServer ES47 video (ua-cam.com/video/z658a8Js5qg/v-deo.html) has a slot B system, and I don't recall him ever mentioning that. It seems like usually when there's a restriction like that, there is some (perhaps weak) technical reason... but I can't guess at it what it might be in this case.
Greatest format in the world. Didn't have enough pins or they'd have kept it for more generations. I had a Slot A 512kb cache 650 Athlon and it rocked for most of a decade. Not just the lack of pins tho, the harmonics caused by a point contact in the middle of a copper contact pad were immensely troublesome. This was resolved a fair bit with PCIe but is still an issue. This package would still be valid today if it included a 2 slot system, one for the power board and the other for the IO
Very cool, yes please do more of that! Might be a bit weird to ask, but here goes: Do you happen to attend a conference about a certain graphics API next week ? I remember meeting you in Brussels in 2019, back when I was a mere masters student, that was cool!
I forgot to renew my passport in time, so I won't be at FOSDEM this year. I really want to go back, and... I kinda want to do a video about the conference.
52:30 I literally just got done replying to a comment about the Celeron 300A! It was the first system I ever built that was with one of these chips. Nothing like a quick and EZ upgrade to 450Mhz!
Good talk. FWIW, I was a level 3 support engineer at Sequent, focusing on performance and kernel crash dump analysis, in the 1990's (my login was "krader").
I was an early adopter of the slot A Athlon, a 650MHz part, running on an IronGate 751 chipset board (the better chipsets from VIA were still a few months away). Needless to say, I just had to try fiddling with the resistor combinations that set the multiplier. Getting the plastic casing off was very difficult; I ended up destroying it - but I also ended up getting a stable 850MHz clock, for free! While it was a faff and I voided the warranty within a week, there is something to be said about small CPU PCBs that allow you to swap resistors around as opposed to the unreliable pencil trick on the later socket A parts. I never did get a goldfinger device - I accidentally inserted a dimm improperly and ended up burning the memory slot and the dimm so the motherboard went into the trash :(
The weirder thing is that with the rise of liquid cooling the slotcart form factor hasn't returned when a cart would make it far easier to mount the cooler and keep extra heat off the mobo
That's because liquid cooling is still pretty much a limited market for specialist and enthousiast applications. Custom builders will use a liquid cooling setup, but there are no large-volume system integrators who would benefit from liquid cooling when compared to the disadvantages (e.g. more time needed to build each system and the service/repair risks being higher).
Nice video. I couldn't find any comments about the BEST feature of this era, or ecosystem of hardware. Slot 1 based motherboards (Pentium 2/3) all had both ISA, PCI and AGP slots, which is great for DOS gaming; good graphics card (AGP), the best sound cards (ISA), and a myriad of other cards on PCI. It is just a fantastic time period. And everything pretty much just works.
Apple's PDS wasn't normally for CPUs to be installed, they were for expansion cards to communicate directly with the CPU. In some instances this facilitated a CPU upgrade because the PDS card's CPU could override the onboard CPU, but this didn't last much past the 68k days (and even then there were exceptions where the PDS wasn't really a PDS). The CPU slot was just called a "CPU slot." The Old World Power Mac CPU cards were really just a way to entice people to buy a computer with an easy upgrade path and to use the same boards with multiple configurations (and later allowed the option of multiple CPUs on the same card). They didn't offer inline caches (they were all board-level L2, on the logic board) until the Mach V variants of the 604e, in the 300 and 350MHz 8600 and 9600s, and then with the backside caches of the PPC 750 and 7400 the rest is history. All of the Old World Power Mac CPU cards were also vertical (excepting some clones). Given the size of the heatsinks on those, I imagine that heat dissipation was the primary reason they were installed in that fashion. Since Pentium IIs were fairly hot, I figure that's also why Intel chose vertical (and I imagine edge connectors are cheaper to manufacture and harder to damage than multi-pin socket types).
My p2-333Mhz slot 1 is still my daily dos gaming / win98se desktop workhorse. I have every possible era of system from a xt class until a 56 core xeon and it amazes me how compatible and versatile the slot 1 generation was/is + it still fits a modern case and psu.
It's worth noting that the Athlon was a brand new platform based on a completely different bus architecture (the DEC EV6 bus, thats why there were DEC Alpha CPUs in the Slot A form factor), not just a new processor and form factor. It took longer to get to market than expected, were expensive, and OEMs had first dibs. The K6-3, then, was a way of buying time till Socket A and its processors came out. The slightly later K6-3+ and K6-2+ were a die shrink aimed at a mobile market hungry for something fast that could use a normal socket. The fact they were unlocked, could run in a lot of desktop boards, and supported on the fly multi adjustment (you could OC alot of them to 500+ MHZ) on the fly meant users with Socket 7 and Super Socket 7 rigs could eke even more use out of their systems. By the time the first Athlons were released I'm sure AMD knew Slot A was a dead platform (only lasting about a year as a consumer platform) and probably only released it due to OEM contracts (the Slot A T-Birds were OEM only, for example). It does make one wonder if AMD might have shifted some resources towards what would become the T-Bird and released the first Athlon in some sort of Socket compatible form factor (you could have a reference design that has the L2 close to the socket, and could use what would become the Irongate NB) for direct consumers instead of messing with slot CPUs had they known how late the Athlon would be to market. As a bonus, it would leave the option of bios updates for those early boards that would let them use newer on die L2 Athlons and use the mobo based L2 as L3. Would have made for some stupid fast systems.
I definitely loved the video. I initially felt a bit.... "bored" isn't the right word. I felt that direct mapping and some early parts were just tiny bit too long (or repetitive?), but as soon as we went to associativity - that was worth it. Especially given how this is among least understood topics. I would LOVE such video on cache, where you explain it all. I had Pentium II 233 MHz. Very decent processor that served me for over 5 years. But since I was mostly gaming, you can't exactly keep playing on it, when most CPU's in early 2000's were already pushing 2 GHz easily (my next CPU was Athlon XP 2200+ - so from memory it was actually 1.8 GHz. And Pentium 4's weren't that great). I remember helping to "service" Pentium III Xeon that was in large tower format in local university in 2010. It was necessary to keep it alive, because it was the only computer with motherboard that still had ISA (?) port. And spectrometer required that port. So some of them had pretty long lives. I did wonder why you didn't compare the caches of Zen 3/Zen 4 as an example of modern cache on CPU, but I guess it makes sense to compare Intel to Intel and differences aren't that staggering. Though when I saw first Zen being released and it's cache structure (from memory L1 - 64KiB [4-Way] Instruction and 32KiB [8-Way] Data Cache for 1 core, L2 - 512 KiB [8-Way] for 1 core, L3 - 8MiB [16-Way] - for CCX) on Zeppelin die, I was blown away by how much it would have to cost in transistor space (though I stopped following technology a decade earlier, so I guess my shock wasn't universal). And then they doubled L3 for Zen 2, and then they unified cache with Zen 3, and then L2 got doubled with Zen 4. And how they did it without gigantic hit to latency (there was some, but it was minor to what was expected) is still a mystery to me. And yes, that is why I'm interested in cache video - because I never understood associativity well, all I "knew" is that "more way better, but costly when it comes to transistors and if you don't use transistors for decoders then slower" - which is a bit simplistic and probably wrong ;) Intel also started to massively increase L2 and L3 cache after their "stagnation period" (to put it mildly). I only wondered whether it is "fair" to compare 1-core CPU transistor count to 8-core CPU transistor count, especially when L1 and L2 caches (depending on micro-architecture) are often counted "per core" (especially L1). Then again, it is valid comparison of 1 CPU to another 1 CPU. If you have more content like that (talking about history or micro-architecture) - I'd definitely watch it. Intel was quite annoyed about AMD getting 1 GHz and Itanium not panning out. I hope they manage to escape the problem that is similar to Boeing (too many managers talking about profits, not enough engineers at highest levels). They used to innovate so much that it is sad what is happening to them :(
39:13 as having worked in chip e-test, I don't see any reason why the cache chip and the cpu chip couldn't and wouldn't have been tested independently. It would be absolutely insane to bond multiple die into packages without testing each die independently. I can tell you that even today, individual pieces of designs are tested independently from each other, even if they're on the same die. Running e-test on a chip is not like running the chip in normal mode. There's a lot of fancy stuff that goes on that allows individual units to be powered up and down and tested independently. And of course pieces that fail testing can be permanently disabled, and the chips down binned. You'd be amazed at how much redundancy there is in the cache arrays even today, and how much extra cache is there that can be mapped in to any location to "repair" defects. All of this functionality is permanently disabled once the chips are tested and configured and binned, so no end user can ever see that.
A couple fun facts (as of my writing) The current chip with the most transistors is Wafer Scale Engine 2 by Cerebras, with a whopping 2.6 Trillion mosfets. The current commercially available CPU with the highest number of transistors is Apple's M2 Ultra, with 134 Billion.
It always seemed obvious to me why they existed. Upgrading your CPU as simply as your GPU, just by plugging in a new card, no need to fiddle with removing and installing a heatsink, and no need to worry about motherboard compatibility. I understand that there are many limitations to doing things this way and that's why it ultimately went away, but boy was it so much more convenient.
A couple of notes/comments: While mentioned, I'd emphasis more in the video that the external L2 cache of the Pentium Pro, Pentium 2 and early Pentium 3s were all on their own dedicated bus independent of the memory bus. This means additional wiring and board complexity. Thus if the SRAMs were placed on the motherboard, additional layers would have been necessary due to the need to keep the SRAM close to the CPU socket. Another factor of SRAM is that it is speed specific: the processor would have to have run it at a divisor of the CPU clock. As faster and faster processors were released, the SRAM divisor would become increasingly greater, resulting in lower efficacy of the cache vs. a natively higher clocked cache. Once it is on the motherboard, it has to support new processors throughout the life span of that platform. Speaking of motherboards, in this era Intel was not the only company making chipsets for their CPUs. By keeping the external L2 cache as part of the CPU package, it'd prevent 3rd party chipset makers form removing a key feature for the Pentium IIs performance. 39:40 - It may have been an artifact of the time but modern chiplet designs leveraging multiple dies in a package do have their individual dies tested before being assembled. I do not know off hand if this applied on the same scale but there are indeed some debug pins on the package that a partially assembled Pentium Pro could be tested. In essence you could test if either all the SRAM is good before adding the CPU core or vice versa. This would catch some of the yield issues early on in the assembly process but obviously not all of them. To your point at 40:00 about binning, yes it was possible. Some of the Pentium Pro 512 KB units were fused off to be Pentium Pro 256 KB model because of bad cache. I haven't seen them but conceptually it'd be possible to bin the 1 MB units with two dies down to 512 KB in two different ways: disable one bad 512 KB die or nerf both 512 KB dies down to 256 KB to make 512 KB total. Intel also had the option of just putting in one 512 KB die into the 1 MB package to simplify manufacturing. One other major difference with the Pentium Pro and modern multi-die designs is that the packaging itself is different: the Pentium Pro leverages ceramics. For chip bonding and wiring, organics is reportedly easier and cheaper to do. Intel started to move to organic packaging with the P55C. Though modern chip packaging effectively uses 'nails' to link two pieces of silicon together in some cases. 45:31 The reasoning for a slotted CPU vs more of a card that is parallel to the motherboard is that it mimics the ATX designs of the time that Intel was promoting. Not all implementations were the same but many vendors kept the CPU card slot parallel to the PCI (-X) slots of the mother board which helped with cooling. Apple also in this era would take IBM/Motorola PowerPC chips and put them onto cards. In Apple's case, it included a small VRM and even multiple processors since multiprocessing on those systems shared a common bus. A select few models of PowerMacs leveraged in-line cache which was also found on the processor card (though there was still a motherboard cache slot). 49:35 AMD didn't copy Intel, they copied DEC Alpha by licensing their bus protocols. This also explains why some Alphas were slotted because the IO bus between an Athlon was the same as the Alphas of the era. One of the ideas of the time is that a motherboard could support both an Athlon or Alpha, though I am only aware of one model ever being released. This also explains 55:45 why AMD released the Athlon being slotted vs. the K6-3 being socketed: the K6-3 was to be a consumer/mobile chip was Athlon (or Alpha!) were meant for workstations. The K6-3 also used AMD's own bus design.
Story about Pentium Pro made me think about how good an Architechture it was. So Good that the modern Intel Core microarchitechtures are still based in part on the Pentium Pro. P6 had legs.
39:18 im curious about what you said about Pentium Pro with "you couldn't test those chips without embedding them both into ceramic, and by that point, you reached the point of no return". Was it not possible to test them together in a BGA adapter harness? did their packaging requirements prevent them from being connectable to anything before soldering? Did BGA harnesses just not exist at that point?
That is a very good question. To be perfectly honest, I don't know. This is an oft repeated "fact," but I don't know of an authoritative source to support it. The Wikipedia article doesn't even cite anything. At that time, there weren't many MCMs (multichip modules), so the technology and best practices probably were not well established. Chips were connected to the pins of the package by soldering tiny wires... maybe it wasn't possible to put them in temporary packaging for testing, then put them in the production package afterwards.
Wouldn't have been feasible in those days - in fact, I'm not sure it's economically feasible today. The size of the contact points of the silicon wafers are so small that you couldn't use a pogo-pin (i.e. BGA) type assembly to test it. You would need flying probes to do this and at that size, wear and tear would be a significant issue assuming you could even have the precision (I doubt this was even possible at the time) to make it reliably repeatable. It's still cheaper to package and bin the CPUs in such large production runs than to constantly replace/ maintain the test vehicles for testing these individual dies. For context, I did some work for one of Seagate's contract manufacturers back in the day - their sister company made a multi-million dollar business out of manufacturing tiny plastic pieces (suction tips) that are used in the machines to pick and place HDD components for manufacturing. That should give you an idea of the kind of wear and tear involved in machines for large scale production like this.
24:14 Not to be confused with what is discussed in this video, I also recall "#-way" being used a few places to also describe how many things can access different locations of cache/buffer/register at the same time, often involving multiple address decoders for each parallel I/O and read/write circuits for each 'word' of the storage-logic array making for very complex circuits indeed. That was a 'very long time ago' and vastly different from methods of deciding what to put in the cache and how to catalog it.
Місяць тому
Serial buses won almost everywhere in both terms of bandwidth and latency because of clock skew (latency) of parallel buses. Slot processors limited the number of pins and were mechanically difficult to install and remove.
One benefit of the slot cpu was no bent pins for the explosion of windows home users as the hobby exploded, cooling was easy and like you said in so many words the foot print for what they were doing with the tech at the time would have took up a very large space on a mobo, I also think that mass production on say a dell assembly line would be less likely do damage a cpu built by hand. Just thinking out loud.
1024 bit memory bus is already available. It is called High Bandwith Memory, HBM for short. And for exactly the reasons you mentioned in the video, they only make it in the form of silicon dies to be connected directly through and interposer, not a normal PCB.
I've still got a PII and PIII here, and it's a fine looking thing - also have the unshrouded Celeron too (the overclockable one). The slot idea, for the time, for both Intel and AMD helped them with the issue of what they needed "in/on and around the cpu" and creating a mini-motherboard interface really solver that for them. Also, and remember using this during upgrade paths, the newer gen 370 cpus could be fitted to a slot motherboard through a slot adapter, very handy. Building with Slot 1s or As was a breeze - typically there were 4 to 8 jumpers to set, and that was it - much easier than the myriad of jumpers for the older isa boards. Beyond that, AMD really came to the fore with the Athlon XP series (loved using those).
The 8088 vs 8086. The 8088 had a smaller bus. This video also reminds me of the gaming benchmarks for the Ryzen _800X3D CPUs. For example, the Ryzen 5800X vs Ryzen 5800X3D. The right X3D CPUs have been dominating gaming benchmarks. They have a crazy amount of L3 cache.
I worked for Unisys in the UK from mid 90's to 2008. I was a Unix/Linux expert and for one of my tasks, I had to put together a little Sequent Symmetry development system for my team in the mid 90's).. My boss took the Unisys Sequent guyin Milton Keynes for a pint ... and I assembled a Sequent Symmetry + Sequent NUMA-Q from spare parts (and no I'm NOT Sequent trained!) and wrapped then up for shipping to Manchester + sneaked them out the back door. I had to lovingly cared for those little mini-computers for a couple of years. I've always wondered ... why the fugazi did the NUMA-Q console run Windows NT 4(!!!). I ran Cygwin/KDE on the console (for X-Windows) to install Oracle!
Addendum - I compiled SAMBA (+ other GNU stuff) from sources on DYNIX/ptx on both the Symmetry and NUMA-Q so the MIcrosofties could map Sequent disk space as 'drives' and could cope with writing Tuxedo/T modules + Oracle bindings as the back end for a Visual BASIC 6 front end (shudders!). They used a Unix/Linux compatible Windows based text editor but I can't remember which one. We wrote a TCP/IP multiplexor to frontend Tuxedo/T to ... erm .... reduce the number of licenses we needed. Neither Oracle nor Tuxedo were impressed but it worked really well and was perfectly legal - sort of like having an L2 cache really!
There were sun's and ibm's that used something like the slot configuration as well, like the Ultra Sparc II in the Sun Ultra 60 and the Power 3 in the rs/6000 43p
At the time didn't hear negativity towards CPU slots. Never had to worry about bent pins or putting the CPU in wrong.
And you could remove the CPU without having to removing the cooler. I like them at the time.
Neither of those are issues with AMD CPUs. Pins on mobo and are covered till you are prepared to install. Marked corners on CPU to line up with marked corner on mobo.
@@SupraSav I'm glad you had good experiences with them. Working in break/fix I saw my fair share of oopsies.
It made swapping CPUs much easier. Benchmarkers probably loved that.
@@SupraSav AMD just switched to LGA 5 minutes ago, they've been pinned for a long time.
I honestly did not expect a Master's degree level of lecture in a random UA-cam video I found today.
I very much enjoyed this format :p
me, too! no more busted knuckles!
i thought they were great. much easier for testing and troubleshooting
My Slot 1 P3 was my favorite PC.
Started as a P3 450, ended as a P3 933 that I acquired at a flea market for $15
Took a bit of BIOS tweaking for my motherboard to even support the damn thing, but it was a massive upgrade.
I didn't particularly like the slot/module form factors.
Because they could get wiggly after enough insert/remove cycles. They seemed to still work fine but I don't believe wiggly parts are truly reliable.
I meant the video content format, but i'm happy all of you had these cool anecdotes to share :p
This is perfect retropc content for just pure listening which is what I have been looking for.
I love playing a video like this in the background while doing laundry for the work week.
Really glad I didn't buy into this because the motherboards had as long a life as the idea.
@@livefreeprintguns I did some laundry too!!
@@randonkbay 😂👍
the format is absolutely perfect! we need more videos in that format :)
old people need it i guess, why talk all the time, keep it shorter please !
I can confirm that running a P55C at 120x2.5 makes for a very fast 300 MHz Pentium. It beat my 300 MHz Klamath running 66 MHz bus at pretty much every metric in dosbench.
That... is interesting. When I finally get around to doing the multiway benchmark battle, I may have to try to reproduce that result.
@@TalesofWeirdStuff I'd love to see if you get similar results!
That's quite interesting. That would mean a 166MHz non-MMX Pentium overclocked to 120FSB, I wonder how stable that is, and how stable the ISA, PCI buses are. What's the heat output like?
@@thebayandurpoghosyanshow no it's an MMX 233 with a lowered multiplier. At 120x2.5 it wasn't perfectly stable but 100x3 seemed stable. Motherboard is Gigabyte 5AA so Super Socket 7 and quite capable. Heat is manageable with a reasonably large SS7 heatsink and fan.
@@agsel that makes more sense, because while the CPU might be able to run at 300MHz internally (to my knowledge, the laptop market actually got 300MHZ P1MMX CPUs), running the FSB at 120MHz, well, I shudder at the thought of how many more variables it introduces.
It would be cool if you made a thread on Vogons with your findings :))
This was the first video of yours I've seen, and I love it. Gonna check out the other stuff next!
Yes, the slot 1 cpu's. I remember playing with those. I had a socket370 to slot 1 converter. I had an Asus P4b-F, with the venerable 440BX chipset. Originaly it was fitted with a PII 350 mhz with a 100mhz FSB. Later, I upgraded it to a PIII 800mhz. I remember searching high and low for the 100mhz FSB version. So that I could install it in my P4b-F. I also had a Thermaltake super orb CPU cooler installed, becouse other coolers would be to large to fit and hit components on the motherboard.
If memory serves, the slot-converter had special brackets, so that you could mount it properly and wouldn't put to much stress on your socket/motherboard.
Unfortunetly I don't have that setup anymore. It's hard to believe it's almost 25 years since I had it.
I run the P4B-F + P3 500 + Voodoo3 3000 for my retro rig
My first slot 1 CPU was a 233Mhz Pentium II that came in an IBM Aptiva tower, but my first system I built myself was a Celeron 300A that was overclockable to 450Mhz out of the box with no additional cooling required. Those were the days!
Was that the Pentium II that was a slot CPU? I might still have that and the MB in a box somewhere. I totally forgot about those things!
@@rootbeer666 At one time I ran a ABIT BH6, Pentium II 450, 384MB PC-100, and a Voodoo 3 3000. Later on I ran a ASUS P3B, a Celeron 1Ghz on a slocket adapter, 1GB PC-100 with a Radeon 8500. (I had hopes that ATI would resolve the driver issues a lot faster than they actually did.)
And if you were really rich, you could get slot 2's!
Wow! I knew the answer vaguely involved cache, and now I have a much stronger grasp on why CPU hardware was so weird for a few years before settling back down into sockets again.
I’d enjoy more videos like this as you want to make them - they contextualize and deepen my appreciation for the benchmarks, and they are good fodder for future videos.
I really look forward to the slotted Pentium/Xeon banchmark video. I’d also be interested to see a comparison video showcasing the performance changes between Intel’s and AMD’s last-gen slot and first-gen socketed CPUs, and a comparison video of the Pentium Pro, Slot 1, and Socket 370 Pentiums (compensating for speed/architecture differences as much as possible) showing the performance delta that cache packaging alone makes.
Ah... I had been so focused on the cache characteristics of CPUs with the same packaging, that I didn't even think about slot 1 vs. socket 370 or slot A vs. socket A. I would guess that the main difference there would be changes in chipset or memory technology... but it's definitely worth checking that assumption. I may just fold that in with the other comparisons.
Nice video! Randomly was suggested your video and you've earnt yourself a new subscriber!
Yes I would be very interested in an in depth review of cache architectures; thank you for taking the time to explain the relevance of slot CPUs.
I loved the concept, no bent pins at all, being easy to slot in and pull out, made it easy to test and troubleshoot quickly.
vid is halfway over, subscribing right now. this may be dry, and maybe i just ate my intellectual-bandwidth-Wheaties this morning or something...but I'm glued.
Wanted to make an "apples to apples" comparison joke... We're just doing PC to PC LOL. Thanks for the memory lane stroll.
Lol! I may steal that joke when I get around to doing those videos. :)
DDR memory lane?
In highschool I used to work at a company called Microway that built servers and workstations. I remember when Digital went out of business the owners came back from an auction with a bunch of those slot socketed alphas. I remember being so intrigued by them. The Owners were a couple. The wife was all business, but the husband was a huge tech nerd. I remember him being super excited because he had gotten all the technical and design documents and schematics for the Alpha processor, along with the sales rights from API.
That's awesome! There are a couple Microway systems that I would love to add to my collection.
So informative, thank you! Really enjoyed this video.
Your videos are like the Great Library of Alexandria, but, thank God they are not burnt. I love them. The amount of info and specific details you are sharing is GOLD. Thank you so much.
This is the first video of yours I’ve seen but it was great. I’d love to hear more deep dives on stuff like cache hierarchies.
The question now is that considering the late 90'$ are back in the forms of "chiplets" (multi chips 2D dyes) CPUs, are we bottlenecking again development wise?
yes and no
the reason we can do advanced packaging now is that complex packaging technology has been advanced so it has the yields and density to be fast, efficient, and cheap enough to be worth it in many cases vs making bigger chips, as at the same time the price per transistor has been stagnating or even going up for advanced process nodes.
I just bumped into your channel but I'm really impressed with your depth of knowledge and storytelling ability. I learned things I didn't know I didn't know so I hope you'll keep doing contents like this!
Just stumbled on your channel from my UA-cam recommendations and while I haven’t seen your other videos as a point of comparison I really liked this conversational format!
I liked Slots probably because I used other systems with slot-like CPUs many years before Intel ever used them. The Acorn Archimedes A540 (1990) used slots for it's CPU and RAM which made them easily replaceable/upgradable even by non-technical people. This continued on in the Risc PC (1994) where it came with 2 CPU slots and you could use 2 ARM CPUs or an ARM and an Intel x86 CPU, and there were multiple speeds of ARM and then later StrongARM CPUs available, all very easily upgradable due to the slot-like cards they were on. They just weren't packaged in big lumps of plastic to hide what was inside.
Hi I am an IT-Technician in training and I'd love to hear you talk more indepth about cache and its many mysteries. Great channel, very glad I found it :D Thoughts: This Format doesn't work for everyone but something about your (enthusiasm or knowledge or both) makes it really fly by. More Please !
I loved overclocking the slot Athlons. It took some fine soldering skills, but it was fun. This channel is invaluable for younger enthusiasts who want to learn about how things developed. Awesome channel.
Thems were the days. [H]ard|OCP shitposting, Alpha/PAL Heatsinks and goldfinger devices.
I had known a little about them having on cpu board cache being an advantage. I am glad I stuck around for an hour to watch this entire video. Thank you for being fairly in depth with all of it. I had wondered why they didn't continue further. Still, I kind of miss the era of computing and being able to use old hardware. It was interesting times indeed. I wish I had a slot one or preferrably a slot A system with a 1ghz cpu. I don't have the time, or the room to keep them around sadly.
I do miss the era. I remember having over 1ghz systems and could pick up these older computers for cheaper at thrift stores. I remember having old systems where I could install 95, 98, 2000 and xp on them and they were still functional. Had a 733mhz slot 1 pentium 3 to use and I forget what my brother had, 1.4ghz pentium 3? It wasn't my main system, but I could go over to my brothers and we could play age of empires or other games on these systems. I just sort of miss the days of when you could switch between windows versions and systems and they would still be functional, you could still browse the web and use them all as daily drivers. Now you just have windows 10 or 11 and they try to kill off old hardware as much as possible. Still, it's why I love linux and it mixes with old hardware. If I have a laptop, it surely is getting linux on it.
Long, but the most authoritative video on the subject on YT. Thanks for sharing your impressive knowledge.
I’m new to this channel and I absolutely loved the format. At first I was apprehensive about the length but when everything about cache speeds clicked it was so worth it :p I’d love to hear more tales like this
This video was pure nostalgia for me, as I remember building one of these machines. I recall it made for a relatively easy build.
I'm surprised every video that you don't have more subscribers. Thank You For Your Time And Effort !!
feelings? about this history lesson?
I was in tears remembering the processors I grew up with
THIS WAS SUPER FUN. I remember being really young and my dad would bring home wiped computers from his job at Polyclad. it was cheaper to give them away to employees than to scrap them at that time and i got a bunch of stuff to mess with. old servers were a complete mystery to me and we didnt have internet or access to server os in any way so i just took them apart and was blown away by the internals, but i remember getting some of these slot cpu based pcs and getting them running just to play roller coaster tycoon but i always wondered why the cpu was different in those machines. Now I know! thanks so much i loved the in depth educational content and look forward to more from you.
You showed up in my recommendations. Great video, very informative. Used to hate this CPU form factor as well, great to hear about why they did it!
it wasen beause the indorlomnt modelig schene in hungariy waset repliring the reomin leyier
Stumbled across this channel subscribed immediately.
You've cleared up a previously muddy story for me. I have a theory that you can clear up your audio artifacts by placing your cell phone elsewhere.
This is the first video of yours I ever watched, and I really liked the deep dive stuff.
Cool video, and yes I subscribed immediately, but I'm surprised you didn't really mention the big issue with a wider bus: namely that all the traces have to be the exactly the same length, and the tolerances get tighter as speeds increase.
Wasn't the video long enough?!? 🤣 I had forgotten about that aspect. Routing an increasing number of traces at all becomes absurdly difficult. Routing them well and within the other limits... forget about it.
I enjoyed this longform description. Well-spoken and well-understood.
i really enjoyed that thank you for the upload
Awesome, the format is called awesome.
First time seeing your channel, awesome retrospective. Having not seen the inside of a PC until my teens, I remember wondering what the heck these were when I saw them on the internet. Your level of detail is immaculate!
Great episode, I have learn a lot about L1 and L2 cache. I have not known that these design have so great impact on the performance.
Intel had an idea for the slot package to integrate separate DSP accelerators, but that idea died with MMX and SSE, and the form factor was only useful until the L2 cache could be economically put on the CPU die. AMD's short adventure with the slot format had some theoretical possibility to cram up to 8MB of L2 clocked at 1/3 CPU speed, but they were also in a hurry to get rid of the expensive packaging. All this happened alongside a very fast progress in silicon manufacturing progress -- the migration to 300mm wafers and the steady transistor node shrinking made monolithic die integration the only economic option.
That is very interesting. I have never heard that before. Do you have any sources? I would really like to learn more.
it would an interesting video, why memory is progressing so slow compared to CPUs and GPUs and why are we still using DDR memory technology that was invented in the late 90s over other memory technology like rambus.
RAMBUS vs everything else is quite another topic. My understanding is that RAMBUS has a couple issues. It has a huge amount of bandwidth, but this comes at the expensive of increased latency. In addition, the way the circuits work internally, a lot of more of the chip is active during memory transactions causing a lot more heat to be generated. RDIMMs had heat spreaders *years* before other RAM modules did.
@@TalesofWeirdStuff also RAMBUS was patent encumbered which lead to it being more expensive so it fell out of failure fairly rapidly.
Honestly memory is so dumb. I remember when I learned how dynamic ram actually worked and I just facepalmed so hard.
It's literally a giant array of buckets with holes in them and you need to have someone run around non stop and refill any bucket that still has water in it to prevent data corruption.
Hence the need for all the massive timing delays in modern memory configurations. So much time is just being spent refilling buckets of water.
You would think we would have found a better way to do it by now but nope. We have (sram) but nothing cheap enough anyways.
@@JathraDH It comes down to cell size and component count. A DRAM cell needs only a couple of transistors and a capacitor. An SRAM cell needs anywhere from 5 to 9 transistors apiece.
Multiply that by however many millions or billions of bits you want to store.
If there was anything simpler that worked at that scale, it's not been found, yet.
@@watchm4ker I know, but its still just dumb lol.
great stuff, this would have been right before i started using computers as a young one and its really a treat to learn more intimately about what happened in the 90s with computing from someone who seems to really have a firm grasp and an affinity for explaining things clearly.
Hey, I'm a first time viewer so I can't say much about the format. I must say I am kinda impressed because I have ADHD and usually can't stay on one video for longer than 20 mins unless it's very interesting! So there you go mate. I was a bit intrigued when you mentioned that you wanted slot adapters. I just made a fast search and ebay has some. I live in Sweden and I'd love to help you out if there is something I can do to help you out. The prices for the slot adapters seem very reasonable to me.
Have a great one!
Socket 370 to slot 1 adapters for Celerons and some Pentium III CPUs are pretty common, and I have several. What is very rare are socket 8 to slot 1 adapters to use a Pentium Pro on a slot 1 motherboard. I've never seen one IRL.
There is an audio glitch at 13:33 in the video almost sounds like a skipping DVD or CD. Other then that great video loved the information and subscribed
I started with Pentium 2, 233Mhz that a friend built for me. A couple years later, he put together an Athlon 850Mhz system with a CPU like that. However that system had a crashing problem. After about an hour of use, or 1-2 minutes of use if I ran a game that used the TNT graphics card in 3D mode, it would instantly crash to a red screen, and I'd have to manually power it off and reboot. Then it would be fine for about another hour (or 1-2 minutes of game time). I eventually stopped using it and went back to the P2, until I got a P4.
I remember that CPU unit had a small card edge connector on top of it, and my friend had plugged a small board into it to over-clock the CPU. Yes, I tried it without that board and it still crashed.
I've got on the PC of my mum back then, it was the MMX 233MHz then years later I've got a P3 866MHz (non-slot Coppermine) PC from my uncle - I was mindblown HAHA XD
I remember those slot CPU days. Good to hear why it was done. It would be interesting to hear about how this applies to GPUs, which have their wide memory busses and memory on the same daughterboard.
Owner of the FIRST 1 Ghz AMD Thunderbird Slot A processor back in the day, it was my first own computer and semi-built (bought a kit from Tigermax, haha), so many great memories with that computer, it never died and I upgraded from it some years later. Thanks for the trip down memory lane man!
Nice! My brain capacity has mightily increaseth, many complex things from the past have been explaineth and many praise must be given!
Also, the format is great. I re-formatted my RAID-6 floppy disk array according to your specs and performance is great - mostly because this array exists only on the outer fringes of my imagination, but probably for some other reason too.
Instant subscribe, much recommended ultra-niche-(cache!)-nerd-stuff, more please!
One of my students was asking about similar things and I was explaining about CPU structure. At the time, I didn't even want to touch on slots... 😕 I always had more issues with them than any socket board. At the time I chocked it up to some people messing up the pins on the CPU chips, whereas slot CPUs didn't have that issue. I may have been completely wrong - obviously - but that's my thinking at the time. Many things had changed (some good, some bad). Older systems had all the I/O and processor/memory on cards and only the bus with slots as the "motherboard". Then came primitive motherboards with just enough to boot the system. Then I/O was gradually added to the motherboards. Floppy drives mysteriously went from up to eight down to two connections. Hard drives went from various settings and RLL complications to IDE making things very simple (so surprising). IDE also added tape backup and CD/DVD/BD. Parallel and serial yielded to USB. During this time we went from 16-bit to 32-bit to 64-bit in basic home systems and laptops. SATA and NvME. Somehow we lost decent sound cards and picked up USB sound "cards" and devices. All manner of USB devices over the USB connection that can then pass to the appropriate device driver (like USB HDDs/USB sticks that pass to the IDE drivers to act as IDE mapped devices). Along the way, devices became so different, processors so different, mobos so different that older OS platforms don't work on modern hardware without virtualization. Yet, this knowledge is disappearing daily as schools churn out students without any hardware understanding and even placing students into programming courses with no Structured Programming preliminary education. Students want to graduate immediately and expect they'll be ready to work with almost no experience. If one of my students expresses an interest beyond the limited aspect of class and wants to know why things are the way they are, I take the time to explain it and to send them to channels like yours for even more info. Thanks for taking the time making/editing this video!
By the time courses are available it's a history of computer science class.😂
When I was briefly teaching computer graphics, I noticed that nobody has the faintest idea how hardware works. When I had to explain bit shifting and bit operations to the students, I almost had to cry myself to sleep. :(
Back in the 1990's I had a job as the in-house computer tech at a multi-million dollar store. I quickly cleared the large (several dozen) backlog of computers needing repair and started taking on more responsibility: Ordering, on-site repair and sales (we had an in store salesman and the large company/government sales, network install and his assistant) and eventually some of the smaller on-site networking.
The in-store salesman handled invoices and first interviews. We had a policy to always interview. One day a Guy came in for a computer tech job, the in-store salesman punted to me.
I asked the Guy to tell me about himself. He'd just graduated college with a computer tech "degree" from ITT.
So I had him come to my work area where I opened up a computer. I explained that I was going to start simple and get more complex to get an idea of his knowledge. He was very excited and said that thus is what he trained for.
I asked him to point out the CPU and I shit you not his response was "Can you give me the schematic diagram?". I figured that the inside salesman had set me up. This Guy could pick out the CPU, memory, hard drive, I/O board, video card and had no idea what ISA stood for. After he left the inside salesman wasn't laughing and said he was on the phone the whole time.
Then a few weeks later another ITT graduate came in. I shit you not, it was the same. Being pretty confident that I wasn't being setup I asked "What do you need a schematic for to find the CPU?" He responded "So I can trace out where the CPU is." Huh?
"How many motherboards have you seen?" He responded "We used about 20 motherboard schematics!" I said "How many actual motherboards have you seen?" He replied "This one, but I just need the schematic". I said "Touch the motherboard." He couldn't.
Feeling bad for the Guy, I told him "We might have a few dozen schematic diagrams for motherboards, most of which are several year old Compaq servers. It is very rare that we repair circuit boards because of how long it takes to get schematic diagrams and how cheap computer parts have become. The store used to have a board level repair tech but it was un-profitable. He would spend hours to 'repair' a component. The store charges $40/hr for that work, but we can't charge for his research time or for him to source parts. A new, quality PC motherboard is about $100 retail. Even if we only charge $80 labor plus parts to repair does that make sense? It is likely the new motherboard will have better features and include a warrenty. We warrenty the repair for 30 days and a good portion had additional 'hidden' problems that have to be fixed at no cost. The reason I know this is that we contract that tech for specific repairs but only for expensive equipment or industrial equipment that can't be replaced."
"Video cards are generally between $30 and $80 retail, I/O cards $25. It costs more money to diagnose the electrical issues for these components than to buy a new, upgraded component. The cost to repair the component is usually several times the cost to replace with new."
"This means a computer tech must be able to immediately identify the components of a computer and have the skill to diagnose which parts need replacement in order to quickly give a customer a detailed list and a price quote."
His answer was "That's not what they trained us." I said "Then find a job where they repair components but don't waste your time at computer stores. You need to find an industrial facility and get an electronics repair tech job." He says "I'be got a degree for computer repair, that's what I want to do and you have no idea what you are talking about!"
Every few months another ITT graduate would come in without ANY knowledge about computers and always request a schematic.
A couple years later I took a class at the local community college. I spoke to the professor about ITT. He laughed and said "yeah, we get ITT grads coming to get a real degree. They are exactly as you describe and usually drop out because we recommend that they get an electronics degree because of their insistence on needing a schematic for everything."
Good stuff, never thought about why slot CPU existed in this light, even though I always knew of daughter board CPUs in the workstations. _shrug_
I suspect that a lot of people knew about workstation or Power Mac CPU cards without making the connection. I never thought about it until recently, and that's mostly because I've been working with workstation CPU boards quite a bit over the last couple years. :)
Your explanations of latency and bandwidth are top notch.
I knew all this going in but still great video! Very entertaining and subscribed.
Celeron 300A is truly the most legendary CPU Intel ever made LOL. I still have both of mine and their ABIT overclocking boards in my closet. Thank you for the trip down memory lane.
Honestly I kind of miss the slot processors, they were quirky and felt special somehow.
Oh, one MINOR nitpick on your bandwidth analogy. Doubling the diameter of a tube does not increase the throughput by 2x, it increases it by 4x as circular area is PI * R^2, 1x1 = 1, 2x2 = 4, 4x4 = 16, etc. Doubling the diameter won't increase it by 4x but its still much more than 2x. Your point got across just fine but it did still trigger my OCD a bit.
So great hearing you explain this. Hopefully some of the younger generation will watch this and realize what we had to go through back in the 90s. I started with an XT back in the mid 80s, a clone.. but the first system I build for myself was a 386sx25 100mb hdd 4mb ram.. yeah it was a dog but SO much better than the XT. Then I put together a 486dx4 100, that thing was light years ahead of the 386. I ran that for so long I NEVER built myself a Pentium.. my buddy grabbed the Pentium 233mhz slot, or I think it was maybe a 266 it was SO long ago.. but man I ran that 486 for so long then and kinda got busy in life, didn't really mess with it for a while, then I took a giant leap and built an Athlon XP 1.3ghz system. That thing was like light years ahead of the 486. Years too, that was probably 2003 when I built that.. crazy how far they've come in SO little time.. since 2019 I've been on AM4, went from a 2600x to a 5800x3d (staying with am4, I already have 32gb ram and all the other fun stuff I need, no reason to go ddr5, why? PCIE4 should be enough for the next few years if not longer) anyway.. man we have come a long way.. I can't wait to see what the next evolution of CPU structure is.. we've come to Moore's law now, I think we're at 3nm on CPUs, Intel isn't even measuring by nanometers anymore.. I'm curious.. I know chiplets, 3d architecture, organic material.. who knows but I know it's going to be big!!
I still have my Athlon 650Mhz Slot A :D lets not forget was the first overclocked cpu reaching the 1Ghz mark :)
Very interesting video who has been gaming since the Atari 2600. Love hearing about how computers have evolved over the years. To be completely honest my first serious home computer was around 2008 when my broher-in-law got me playing World of Warcraft. Up until then I used my consoles (NES, SNES, Genesis, Dreamcast, PS1, & PS2) for gaming and my computer was for everything else. I've never looked back. About every 5 years I upgrade my PC with somethign that keep me pretty close to the "bleedign edge" so I can play modern titles at max, or near max, setting and not have unbearable lag or latency. Thank you for your video. Going to check out the rest of your catalog. Btw, I do also enjoy computer repair videos and by what you said about your usual format at the conclusion of this video I suspect I will be watching more of your video. Like, subscrived, AND commented. Have a good day.
Hi Ian. I enjoyed every minute of this video, although I watched it for three days, as it is not a video to do "something else" while watching, but attending, understanding and visualizing in (my own) memory... It brought to my memory some weirdnesses on caches from my past, where the only way to optimize some math crunching processes was turning down the caches off, although you couldn't modify L1 always. I will definitely keep this video and revisit it when playing with caches again, as it brought me nice (and bitter) memories. A video on cache performance running on "weird stuff" (with weird code to force cache substitution, if possible) would be much appreciated!!!
My dad upgraded from his 1040STF to a Performa 475. The software "MagiC Mac" was the main reason that he switched. We still have a LC-PDS interface with serial and parallel ports that was patched into MagiC Mac to work directly with original printers and other hardware. I have lost the driver/patch floppy, though and cannot get the card to work anymore.
I loved slot based CPUs, they were super easy to assemble as long as you kept the stock cooler, I still have the fondest of memories of these platforms.
Nice, refreshing vidoe! I only can say good things about slot1 CPUs, my 266MHz celeron used to operate at 450 to 500MHz which was quite fine overclock for any day standards and considering the price of PII... Still to this day I have a drawer full of older CPUs, including a few slot1s. Brings memories.
Very interesting! You should do more of these lecture style videos.
WOW! 1 hour explanation for slot CPU! Insane! But appreciate your dedication! Just way way too long for someone just wanted to know why slot CPU and nothing more. But this is masterclass in CPU architecture and goes far beyond the video title.
13:50 It's unfortunate that the graph ends at 2000. Around 2010 the memory controller was on-die and the memory got physically closer to the CPU.
There are updated versions of the graph in later editions of the book, and the same trend continues.
Going back to DEC, there are some systems with the Alpha 21264, like the DEC/Compaq AlphaServer DS20 that I have, which have a type of "slot based" CPU configuration, just like the Pentium II, Pentium III, and Athlon CPUs (Like you mentioned in the video). The CPU card in my AlphaServer contains the CPU soldered to the card, along with a VRM, cache memory, and a firmware ROM. The 500MHz CPU card that I have has 4MBytes of cache on the card. Because there is a VRM on the CPU card, it has a separate plug to provide +5VDC and Ground to the CPU card directly. I found it kind of interesting that they would include a firmware ROM on the CPU card too, rather than on the mainboard. My system supports two of these CPU cards, although I only have one installed currently. They are getting really difficult to find.
One problem is that I can't just slot in one of the 833MHz CPU cards from the DS25 or DS40, for example. They made the CPU cards incompatible across systems. I think I recall a 633MHz Alpha is the fastest card I can put in my particular system. I'm not entirely sure why they did this, and why the CPUs were soldered right in to these CPU cards, but my guess is that it forces you to go to their newer model of AlphaServer if you want the faster CPUs.
That is interesting. I was not aware of compatibility issues with the slot B Alphas. My friend in the AlphaServer ES47 video (ua-cam.com/video/z658a8Js5qg/v-deo.html) has a slot B system, and I don't recall him ever mentioning that. It seems like usually when there's a restriction like that, there is some (perhaps weak) technical reason... but I can't guess at it what it might be in this case.
Greatest format in the world. Didn't have enough pins or they'd have kept it for more generations. I had a Slot A 512kb cache 650 Athlon and it rocked for most of a decade. Not just the lack of pins tho, the harmonics caused by a point contact in the middle of a copper contact pad were immensely troublesome. This was resolved a fair bit with PCIe but is still an issue. This package would still be valid today if it included a 2 slot system, one for the power board and the other for the IO
Very cool, yes please do more of that!
Might be a bit weird to ask, but here goes: Do you happen to attend a conference about a certain graphics API next week ? I remember meeting you in Brussels in 2019, back when I was a mere masters student, that was cool!
I forgot to renew my passport in time, so I won't be at FOSDEM this year. I really want to go back, and... I kinda want to do a video about the conference.
@@TalesofWeirdStuff I'm also missing FOSDEM this year since I have to fly in the other direction. To be clear I meant Vulkanised :P
Ah. Yeah, I'm not going to that either. :)
52:30 I literally just got done replying to a comment about the Celeron 300A! It was the first system I ever built that was with one of these chips. Nothing like a quick and EZ upgrade to 450Mhz!
Good talk. FWIW, I was a level 3 support engineer at Sequent, focusing on performance and kernel crash dump analysis, in the 1990's (my login was "krader").
This was a fascinating listen. Thank you!
I really liked the slot design. It was easy to install/remove the CPU without having to worry about bending any pins.
I was an early adopter of the slot A Athlon, a 650MHz part, running on an IronGate 751 chipset board (the better chipsets from VIA were still a few months away). Needless to say, I just had to try fiddling with the resistor combinations that set the multiplier. Getting the plastic casing off was very difficult; I ended up destroying it - but I also ended up getting a stable 850MHz clock, for free! While it was a faff and I voided the warranty within a week, there is something to be said about small CPU PCBs that allow you to swap resistors around as opposed to the unreliable pencil trick on the later socket A parts. I never did get a goldfinger device - I accidentally inserted a dimm improperly and ended up burning the memory slot and the dimm so the motherboard went into the trash :(
The weirder thing is that with the rise of liquid cooling the slotcart form factor hasn't returned when a cart would make it far easier to mount the cooler and keep extra heat off the mobo
That's because liquid cooling is still pretty much a limited market for specialist and enthousiast applications. Custom builders will use a liquid cooling setup, but there are no large-volume system integrators who would benefit from liquid cooling when compared to the disadvantages (e.g. more time needed to build each system and the service/repair risks being higher).
It would also be pretty difficult to get up to 2000 contacts onto an edge connector or pin row
Wow, 10 minutes of interesting info relating to the subject in almost an hour video.
Nice video. I couldn't find any comments about the BEST feature of this era, or ecosystem of hardware. Slot 1 based motherboards (Pentium 2/3) all had both ISA, PCI and AGP slots, which is great for DOS gaming; good graphics card (AGP), the best sound cards (ISA), and a myriad of other cards on PCI. It is just a fantastic time period. And everything pretty much just works.
Apple's PDS wasn't normally for CPUs to be installed, they were for expansion cards to communicate directly with the CPU. In some instances this facilitated a CPU upgrade because the PDS card's CPU could override the onboard CPU, but this didn't last much past the 68k days (and even then there were exceptions where the PDS wasn't really a PDS). The CPU slot was just called a "CPU slot." The Old World Power Mac CPU cards were really just a way to entice people to buy a computer with an easy upgrade path and to use the same boards with multiple configurations (and later allowed the option of multiple CPUs on the same card). They didn't offer inline caches (they were all board-level L2, on the logic board) until the Mach V variants of the 604e, in the 300 and 350MHz 8600 and 9600s, and then with the backside caches of the PPC 750 and 7400 the rest is history. All of the Old World Power Mac CPU cards were also vertical (excepting some clones). Given the size of the heatsinks on those, I imagine that heat dissipation was the primary reason they were installed in that fashion. Since Pentium IIs were fairly hot, I figure that's also why Intel chose vertical (and I imagine edge connectors are cheaper to manufacture and harder to damage than multi-pin socket types).
For a second I thought you were going use a bong to make a bandwidth analogy.
That made me literally laugh out loud. That is, however, a different kind of water pipe. :) I wonder... would YT de-monetize a video for that?
My p2-333Mhz slot 1 is still my daily dos gaming / win98se desktop workhorse. I have every possible era of system from a xt class until a 56 core xeon and it amazes me how compatible and versatile the slot 1 generation was/is + it still fits a modern case and psu.
It's worth noting that the Athlon was a brand new platform based on a completely different bus architecture (the DEC EV6 bus, thats why there were DEC Alpha CPUs in the Slot A form factor), not just a new processor and form factor. It took longer to get to market than expected, were expensive, and OEMs had first dibs. The K6-3, then, was a way of buying time till Socket A and its processors came out. The slightly later K6-3+ and K6-2+ were a die shrink aimed at a mobile market hungry for something fast that could use a normal socket. The fact they were unlocked, could run in a lot of desktop boards, and supported on the fly multi adjustment (you could OC alot of them to 500+ MHZ) on the fly meant users with Socket 7 and Super Socket 7 rigs could eke even more use out of their systems. By the time the first Athlons were released I'm sure AMD knew Slot A was a dead platform (only lasting about a year as a consumer platform) and probably only released it due to OEM contracts (the Slot A T-Birds were OEM only, for example). It does make one wonder if AMD might have shifted some resources towards what would become the T-Bird and released the first Athlon in some sort of Socket compatible form factor (you could have a reference design that has the L2 close to the socket, and could use what would become the Irongate NB) for direct consumers instead of messing with slot CPUs had they known how late the Athlon would be to market. As a bonus, it would leave the option of bios updates for those early boards that would let them use newer on die L2 Athlons and use the mobo based L2 as L3. Would have made for some stupid fast systems.
I definitely loved the video. I initially felt a bit.... "bored" isn't the right word. I felt that direct mapping and some early parts were just tiny bit too long (or repetitive?), but as soon as we went to associativity - that was worth it. Especially given how this is among least understood topics. I would LOVE such video on cache, where you explain it all.
I had Pentium II 233 MHz. Very decent processor that served me for over 5 years. But since I was mostly gaming, you can't exactly keep playing on it, when most CPU's in early 2000's were already pushing 2 GHz easily (my next CPU was Athlon XP 2200+ - so from memory it was actually 1.8 GHz. And Pentium 4's weren't that great).
I remember helping to "service" Pentium III Xeon that was in large tower format in local university in 2010. It was necessary to keep it alive, because it was the only computer with motherboard that still had ISA (?) port. And spectrometer required that port. So some of them had pretty long lives.
I did wonder why you didn't compare the caches of Zen 3/Zen 4 as an example of modern cache on CPU, but I guess it makes sense to compare Intel to Intel and differences aren't that staggering.
Though when I saw first Zen being released and it's cache structure (from memory L1 - 64KiB [4-Way] Instruction and 32KiB [8-Way] Data Cache for 1 core, L2 - 512 KiB [8-Way] for 1 core, L3 - 8MiB [16-Way] - for CCX) on Zeppelin die, I was blown away by how much it would have to cost in transistor space (though I stopped following technology a decade earlier, so I guess my shock wasn't universal). And then they doubled L3 for Zen 2, and then they unified cache with Zen 3, and then L2 got doubled with Zen 4. And how they did it without gigantic hit to latency (there was some, but it was minor to what was expected) is still a mystery to me. And yes, that is why I'm interested in cache video - because I never understood associativity well, all I "knew" is that "more way better, but costly when it comes to transistors and if you don't use transistors for decoders then slower" - which is a bit simplistic and probably wrong ;)
Intel also started to massively increase L2 and L3 cache after their "stagnation period" (to put it mildly). I only wondered whether it is "fair" to compare 1-core CPU transistor count to 8-core CPU transistor count, especially when L1 and L2 caches (depending on micro-architecture) are often counted "per core" (especially L1). Then again, it is valid comparison of 1 CPU to another 1 CPU.
If you have more content like that (talking about history or micro-architecture) - I'd definitely watch it. Intel was quite annoyed about AMD getting 1 GHz and Itanium not panning out. I hope they manage to escape the problem that is similar to Boeing (too many managers talking about profits, not enough engineers at highest levels). They used to innovate so much that it is sad what is happening to them :(
Alpha Dec the extra 128bit of memory could be used by the surrounding CISC cpus. I have seen design specs with up to 5 Cisc processors.
There were "regular" SMP Pentium Pro systems with 6 CPUs and Xeon systems with up to 8 CPUs.
39:13 as having worked in chip e-test, I don't see any reason why the cache chip and the cpu chip couldn't and wouldn't have been tested independently. It would be absolutely insane to bond multiple die into packages without testing each die independently. I can tell you that even today, individual pieces of designs are tested independently from each other, even if they're on the same die. Running e-test on a chip is not like running the chip in normal mode. There's a lot of fancy stuff that goes on that allows individual units to be powered up and down and tested independently. And of course pieces that fail testing can be permanently disabled, and the chips down binned. You'd be amazed at how much redundancy there is in the cache arrays even today, and how much extra cache is there that can be mapped in to any location to "repair" defects. All of this functionality is permanently disabled once the chips are tested and configured and binned, so no end user can ever see that.
A couple fun facts (as of my writing) The current chip with the most transistors is Wafer Scale Engine 2 by Cerebras, with a whopping 2.6 Trillion mosfets.
The current commercially available CPU with the highest number of transistors is Apple's M2 Ultra, with 134 Billion.
It always seemed obvious to me why they existed. Upgrading your CPU as simply as your GPU, just by plugging in a new card, no need to fiddle with removing and installing a heatsink, and no need to worry about motherboard compatibility. I understand that there are many limitations to doing things this way and that's why it ultimately went away, but boy was it so much more convenient.
A couple of notes/comments:
While mentioned, I'd emphasis more in the video that the external L2 cache of the Pentium Pro, Pentium 2 and early Pentium 3s were all on their own dedicated bus independent of the memory bus. This means additional wiring and board complexity. Thus if the SRAMs were placed on the motherboard, additional layers would have been necessary due to the need to keep the SRAM close to the CPU socket. Another factor of SRAM is that it is speed specific: the processor would have to have run it at a divisor of the CPU clock. As faster and faster processors were released, the SRAM divisor would become increasingly greater, resulting in lower efficacy of the cache vs. a natively higher clocked cache. Once it is on the motherboard, it has to support new processors throughout the life span of that platform. Speaking of motherboards, in this era Intel was not the only company making chipsets for their CPUs. By keeping the external L2 cache as part of the CPU package, it'd prevent 3rd party chipset makers form removing a key feature for the Pentium IIs performance.
39:40 - It may have been an artifact of the time but modern chiplet designs leveraging multiple dies in a package do have their individual dies tested before being assembled. I do not know off hand if this applied on the same scale but there are indeed some debug pins on the package that a partially assembled Pentium Pro could be tested. In essence you could test if either all the SRAM is good before adding the CPU core or vice versa. This would catch some of the yield issues early on in the assembly process but obviously not all of them. To your point at 40:00 about binning, yes it was possible. Some of the Pentium Pro 512 KB units were fused off to be Pentium Pro 256 KB model because of bad cache. I haven't seen them but conceptually it'd be possible to bin the 1 MB units with two dies down to 512 KB in two different ways: disable one bad 512 KB die or nerf both 512 KB dies down to 256 KB to make 512 KB total. Intel also had the option of just putting in one 512 KB die into the 1 MB package to simplify manufacturing.
One other major difference with the Pentium Pro and modern multi-die designs is that the packaging itself is different: the Pentium Pro leverages ceramics. For chip bonding and wiring, organics is reportedly easier and cheaper to do. Intel started to move to organic packaging with the P55C. Though modern chip packaging effectively uses 'nails' to link two pieces of silicon together in some cases.
45:31 The reasoning for a slotted CPU vs more of a card that is parallel to the motherboard is that it mimics the ATX designs of the time that Intel was promoting. Not all implementations were the same but many vendors kept the CPU card slot parallel to the PCI (-X) slots of the mother board which helped with cooling. Apple also in this era would take IBM/Motorola PowerPC chips and put them onto cards. In Apple's case, it included a small VRM and even multiple processors since multiprocessing on those systems shared a common bus. A select few models of PowerMacs leveraged in-line cache which was also found on the processor card (though there was still a motherboard cache slot).
49:35 AMD didn't copy Intel, they copied DEC Alpha by licensing their bus protocols. This also explains why some Alphas were slotted because the IO bus between an Athlon was the same as the Alphas of the era. One of the ideas of the time is that a motherboard could support both an Athlon or Alpha, though I am only aware of one model ever being released. This also explains 55:45 why AMD released the Athlon being slotted vs. the K6-3 being socketed: the K6-3 was to be a consumer/mobile chip was Athlon (or Alpha!) were meant for workstations. The K6-3 also used AMD's own bus design.
Ah yes, Hennesey and Patterson, the 'cinder block' as I think of it. It is, very much, a tome.
They one I had back in the day was definitely the heaviest textbook of all my undergrad classes.
Story about Pentium Pro made me think about how good an Architechture it was. So Good that the modern Intel Core microarchitechtures are still based in part on the Pentium Pro. P6 had legs.
39:18 im curious about what you said about Pentium Pro with "you couldn't test those chips without embedding them both into ceramic, and by that point, you reached the point of no return". Was it not possible to test them together in a BGA adapter harness? did their packaging requirements prevent them from being connectable to anything before soldering? Did BGA harnesses just not exist at that point?
That is a very good question. To be perfectly honest, I don't know. This is an oft repeated "fact," but I don't know of an authoritative source to support it. The Wikipedia article doesn't even cite anything. At that time, there weren't many MCMs (multichip modules), so the technology and best practices probably were not well established. Chips were connected to the pins of the package by soldering tiny wires... maybe it wasn't possible to put them in temporary packaging for testing, then put them in the production package afterwards.
The similar statement in the Slot 1 article on Wikipedia is tagged "citation needed." :)
Wouldn't have been feasible in those days - in fact, I'm not sure it's economically feasible today. The size of the contact points of the silicon wafers are so small that you couldn't use a pogo-pin (i.e. BGA) type assembly to test it.
You would need flying probes to do this and at that size, wear and tear would be a significant issue assuming you could even have the precision (I doubt this was even possible at the time) to make it reliably repeatable.
It's still cheaper to package and bin the CPUs in such large production runs than to constantly replace/ maintain the test vehicles for testing these individual dies.
For context, I did some work for one of Seagate's contract manufacturers back in the day - their sister company made a multi-million dollar business out of manufacturing tiny plastic pieces (suction tips) that are used in the machines to pick and place HDD components for manufacturing.
That should give you an idea of the kind of wear and tear involved in machines for large scale production like this.
24:14 Not to be confused with what is discussed in this video, I also recall "#-way" being used a few places to also describe how many things can access different locations of cache/buffer/register at the same time, often involving multiple address decoders for each parallel I/O and read/write circuits for each 'word' of the storage-logic array making for very complex circuits indeed. That was a 'very long time ago' and vastly different from methods of deciding what to put in the cache and how to catalog it.
Serial buses won almost everywhere in both terms of bandwidth and latency because of clock skew (latency) of parallel buses.
Slot processors limited the number of pins and were mechanically difficult to install and remove.
One benefit of the slot cpu was no bent pins for the explosion of windows home users as the hobby exploded, cooling was easy and like you said in so many words the foot print for what they were doing with the tech at the time would have took up a very large space on a mobo, I also think that mass production on say a dell assembly line would be less likely do damage a cpu built by hand. Just thinking out loud.
1024 bit memory bus is already available. It is called High Bandwith Memory, HBM for short. And for exactly the reasons you mentioned in the video, they only make it in the form of silicon dies to be connected directly through and interposer, not a normal PCB.
Still have my pentium 2 cpu displayed in my room XD loved how strange it looked. Thx for the video always wonderd why!
I've still got a PII and PIII here, and it's a fine looking thing - also have the unshrouded Celeron too (the overclockable one). The slot idea, for the time, for both Intel and AMD helped them with the issue of what they needed "in/on and around the cpu" and creating a mini-motherboard interface really solver that for them. Also, and remember using this during upgrade paths, the newer gen 370 cpus could be fitted to a slot motherboard through a slot adapter, very handy. Building with Slot 1s or As was a breeze - typically there were 4 to 8 jumpers to set, and that was it - much easier than the myriad of jumpers for the older isa boards. Beyond that, AMD really came to the fore with the Athlon XP series (loved using those).
My first P3 build had a Tyan motherboard that had both. It started with a 233 or 333 slot CPU and ended up with a 933 "flip-chip" socketed CPU.
This certainly takes me back to memory lane.
The 8088 vs 8086. The 8088 had a smaller bus.
This video also reminds me of the gaming benchmarks for the Ryzen _800X3D CPUs. For example, the Ryzen 5800X vs Ryzen 5800X3D. The right X3D CPUs have been dominating gaming benchmarks. They have a crazy amount of L3 cache.
I worked for Unisys in the UK from mid 90's to 2008. I was a Unix/Linux expert and for one of my tasks, I had to put together a little Sequent Symmetry development system for my team in the mid 90's).. My boss took the Unisys Sequent guyin Milton Keynes for a pint ... and I assembled a Sequent Symmetry + Sequent NUMA-Q from spare parts (and no I'm NOT Sequent trained!) and wrapped then up for shipping to Manchester + sneaked them out the back door. I had to lovingly cared for those little mini-computers for a couple of years. I've always wondered ... why the fugazi did the NUMA-Q console run Windows NT 4(!!!). I ran Cygwin/KDE on the console (for X-Windows) to install Oracle!
Addendum - I compiled SAMBA (+ other GNU stuff) from sources on DYNIX/ptx on both the Symmetry and NUMA-Q so the MIcrosofties could map Sequent disk space as 'drives' and could cope with writing Tuxedo/T modules + Oracle bindings as the back end for a Visual BASIC 6 front end (shudders!). They used a Unix/Linux compatible Windows based text editor but I can't remember which one. We wrote a TCP/IP multiplexor to frontend Tuxedo/T to ... erm .... reduce the number of licenses we needed. Neither Oracle nor Tuxedo were impressed but it worked really well and was perfectly legal - sort of like having an L2 cache really!
I remember the slot processors, I skipped over that period. I went from Super Socket 7 with a K6-III 450Mhz to Socket 370, Pentium III 1.4GHz
There were sun's and ibm's that used something like the slot configuration as well, like the Ultra Sparc II in the Sun Ultra 60 and the Power 3 in the rs/6000 43p
I'm getting an Ultra 60 tonight, so I'll have to check that out. I _thought_ the used boards that were vaguely reminiscent of the old MBus boards.