Niiice HEDT! Mind you, optical audio is (very) limited in bandwidth, hdmi does more channels, higher bitrates and does so uncompressed. You could get RGB support from an internal add-on box, from razer for instance, sacrificing an intenal usb connector.
Note - the LGA 4677 motherboards are not easily found at a reasonable price. Chinese Ebay sellers sell them at a premium compared to local - if you can find them that is.
Yeah Gigabyte MHBs were going for about $500 before 4th gen boom with AI. Now they're up $700 and being resold at $1200 now. But 3647 and 4189 aren't really worth it anymore, but still better than Broadwell Es at least.
I doubt they pay that much. NVIDIA charges $30k for an H100, but the lowest price on average I've seen for customers that buy hundreds of them or even thousands of them is $3,000-$4,000 per GPU. The steepest discount I have ever seen on an Intel CPU for a single customer was for a pair of Xeon Platinum 8380s, which being $20k for said pair, were being offered at an 80% discount to just $4k. Of course, this discount was offered through a workstation retailer/vendor to a client that was likely well connected. You don't need connections to get an engineering sample CPU like the one shown in this video.
@@NUCLEARARMAMENT I used to cost such things like getting quotes for 100s of Xeons based servers. Any relatively large corp can easily get 60% discount off list. No one pays list. Anyone agreeing to list price should be fired.
I've always been wary of engineering samples. Do all the PCIe lanes work? You could maybe test that by putting the GPU in different slots. Does AVX-512 work?
Great video ! Could you please provide llama.cpp inference speeds for CPU alone for some llama 70B model e.g. Q4 and for example mistral 7B Q4 or maybe you already posted the results somewhere ?
Wow thx for being in my feed! 600 bucks US for a low clocked 60 core! Might be a good match for Folding at home or BOINC scientific computing however how cheaply can 1 get a motherboard for it if i don't need any PCI express lanes?
nice server system. I have a dual xeon 1366 system with a few sets of cpus. My modern ryzen beats it though by a lot.The 6 channel memory is still faster than most things today.
Why you need this, how many users on this system ? Running what ? If you have thousands and thousands of users, you will benefit from form 6 channels ! all needing the same resources, speed is essential !
@@lucasrem I bought it years ago. When you have simple low powered systems for years you get some money and buy something crazy. Thats what I did and it was cheap. I run windows and use benchmarks for fun. Now I use a ryzen and have that system in the corner of my room unused. Sometimes i play with it like any nerd would.
Hi! That's exciting. I have been looking at 2nd gen Epyc and 4 Socket Broadwell recently for a single-node cluster solution, but this would also tick the box. Can you tell me what all core turbo you can get? I imagine much more than the 1.7Ghz base? And do you have a passmark score by any chance so I can compare it to let's say 2x 7D12 Epycs?
Hey this might be an odd request but I wanted to ask if you could maybe try to compile chromium as a benchmark, I'm looking for a new work station with that as my main benchmark since it's the majority of work I do and would really love to get an idea of what I can expect with these CPUs.
You're definitely going to want to invest in a case with a fan config you're confident will comfortably cool at least a pair of passive 2-slot accelerators. Even if you don't plan on going nVidia, the meta when it comes to accelerators is high-density capable cards (those don't come with fans), especially when it comes to used cards.
@@itsmorefuntocompute3442 Let's see... Just eyeballing it, if you installed it in the 901's inverted layout (PCIe slots on top, CPU towards the bottom) you should be able to just fashion an air guide towards the cards, and a manual fan controller for the topmost fan. That'll give you the room and cooling to fit 2 accelerators, no matter who you choose. Be it nVidia Tesla, Radeon Instinct, Intel IA-series, or (if you're lucky enough to have that kind of access) Google's Coral Server series accelerators. Just remember to bump your fan speed if either card hits 80c, most will self-throttle at that point to protect the card, but won't complain or crash. P.S. If you do choose nVidia Tesla or Intel IA, DON'T use an 8-pin PCIe power cable on the accelerator. Tesla cards from Pascal onwards, and most Intel IA cards, use 8-pin EPS Power (sometimes labeled "CPU power" on PSUs) which looks identical to 8-pin PCIe, but have an extra 12v rail and an inverted pinout. If you attach 8-pin PCIe instead of 8-pin EPS you will blow the card and likely your PSU too. You can get PCIe to EPS adapters if your PSU doesn't have the extra EPS cables, just make sure you get one with a thick wire.
@@scouzi7201 Don't waste your time with air cooling. I built a dual socket custom cooling loop for LGA-4677 (Socket E, which is what these Sapphire & Emerald Rapids CPUs use), and the price for the D5 pump and reservoir combo ($80), two Byski LGA-4677 waterblocks ($100), EPDM tubing (15 ft, about $50), 16 G 1/4" BSPP threaded soft-tube compression fittings ($50), and a couple of 420 mm Hardware Labs GTS 420s (30 mm thick/420 mm long radiators, $150), all came out to about $430, which is nothing for a system like this that will easily dissipate 360 W per CPU or 720 W for both chips at full load without any overclocking. 1,000+ W dissipation for two of these [60-cores] even mildly overclocked (10%-20%) is not unheard of in heavy-handed in either vectorized or multithreaded workloads. In fact the Xeon W-3495X with 56 cores at stock manages 700+ W heat loads in AVX-512 full-width vectorized workloads, so two of them would hit 1,400+ W at stock. The air cooler (a Noctua UH14S meant for LGA-4677) was hitting 90-100+ C whereas the max junction temperature for this CPU is less than 90 C, and you don't want to operate above that 24/7 especially when you need to run mission critical workloads for extended periods of time. The heatsink is also $125 each, so it's cheaper than the $430 watercooling setup I specified earlier, but your temperatures will be at least 30-40 C worse for the wattage heat dissipation. Trust me, spare yourself the headache and invest in watercooling, and enjoy an extra decade or more in lifespan versus air cooling; not to mention you never have to worry about case or DIMM clearance issues again if you go with watercooling.
The board layout is super weird. The first few PCIe slots will be constrained by the ram sticks. Probably all slots are x16 (this platform has many PCIe lanes) so you can use GPUs on the lower slots
Nice im also working on a epyc system built its a Supermicro H11DSI dual-socket motherboard im using the epyc 7551 and the cooler TR4 sp3 i and for the ram 8gbx8 64gb so far
Was reading the specs sheet for this. Surprised to see 16 High Priority Cores @ 2.0ghz + 44 Low Priority Cores @ 1.60Ghz. However it says the “Base Frequency” is 1.90Ghz, how do they get that number, an average. They claim Max Turbo Boost is 3.50Ghz! YEAH RIGHT!
First of all, How? unless its a sample even if used.. you know what Iam glad for you brother! it totally demolishes threadripper which is 500USD I mean the ram & motherboard are expensive as is.. What is the power draw on that beast under load?
No sane person would build a Threadripper if they knew about the existence of these CPUs and realized the performance to price ratio, unless they were a financially illiterate AMD fanboy.
@@NUCLEARARMAMENT Yeah; No sane person would want a cpu that’s guaranteed to work, has a warranty, full motherboard compatibility, draws less power and is far more reliable…. I know you lack common sense but this is a EARLY ENGINEERING SAMPLE, hence why a 1 year old data center cpu is selling for 1/40th of its “retail price”….. You’ll have to be incredibly stupid to use this in a application that requires high availability!
22927 difference with the Ryzen. This is only for companies and people with money, great acquisition. it's time to hunt for offers on these machines. Thanks for share the power.
Thanks for taking the time to share this. I was looking for a similar purchase , this sample seems to be working quite well. Do you have the exact parts number or eBay reference ?
llama_print_timings: load time = 13387.07 ms llama_print_timings: sample time = 74.30 ms / 400 runs ( 0.19 ms per token, 5383.36 tokens per second) llama_print_timings: prompt eval time = 308.04 ms / 19 tokens ( 16.21 ms per token, 61.68 tokens per second) llama_print_timings: eval time = 91794.32 ms / 399 runs ( 230.06 ms per token, 4.35 tokens per second) llama_print_timings: total time = 92486.77 ms / 418 tokens Log end@@adamstewarton
Very interesting to see an ES working ! do you think this kind of cpu can works on w790 chipset ? The bords with w790 are estheticaly nicer as you mentioned in the video.
@@timramich Hello, i did some research about server xeons on workstation motherboard. It seem it can work on some board (Asus's, gigabytes's and some asrocks's) if the microcode for the D0 stepping of engeneering sample CPU is still present in the bios. Intel only seperated their xeon lineup in 2 separate brandings (Workstation and scalable) at the E0 stepping . Just to prepare the consumer samples we can buy. So your 8490h ES could work on a w790 mb if it is a D0 (probably since it is the most common).
AGI is artificial general intelligence. There are different definitions, I would define it as an AI which has human expert abilities in a wide spectrum of tasks. I think humans would have called GPT4 an "AGI" 20 years ago.
@@itsmorefuntocompute3442I'd probably have called ChatGPT4 (which I do use myself) a set of smart algorithms. However: ChatGPT only reacts to something - no matter what it is. An AGI should be able to proactively act instead of just reacting. So given it knows how the world works, I'd expect it to want to improve / change it (for the better or for worse). But as you said, there are different definitions of what an AGI supposedly is.
these things sip power regardless.... funny how over the last 25 years i seen things go from OMFGURD this is alien tech is hardly draws any power !!!1 and the SAME EXACT CPU 10 years later, everyone things its a "nuclear power plant"...LOL Fucking morons.
this is llama 2 70b: llama_print_timings: load time = 44755.73 ms llama_print_timings: sample time = 47.43 ms / 241 runs ( 0.20 ms per token, 5081.71 tokens per second) llama_print_timings: prompt eval time = 2945.46 ms / 21 tokens ( 140.26 ms per token, 7.13 tokens per second) llama_print_timings: eval time = 195141.13 ms / 240 runs ( 813.09 ms per token, 1.23 tokens per second) llama_print_timings: total time = 198337.28 ms / 261 tokens
60 cores 120 threads. This is overkill, maybe usefull in the 2030s for gaming lol The worst thing is that it's clocked at 1.9Ghz based with 350W and can overclock to around 3ghz. 112MB of cache, that's gianormous
@@trumpsextratesticle8590 How much is your time worth? For many people, a small investment in a modern PC, which you can build yourself, is better than spending lots of time and effort on a dusty old obsolete Xeon, that still has many compromises and issues in 2024. I admire Ford pickup trucks from the 1940's, but I wouldn't buy one and spend many many hours fixing it and refurbishing it! Xeon is the same way.
@@DerekDavis213 If someone is video editing or whatnot this Xeon build is just fine. Most people are not doing enterprise level shit or needing that security level. Most people would be fine on a Xeon, even gaming on almost all new titles.
@@itsmorefuntocompute3442 Thanks, then congrats to your new computer! But I have to tell you that designing a gerneral AI is probably not only a question of compute power 😉
"I mentioned 'execute,' but should have used the word 'inference.' The design and training of models nearing AGI capabilities are currently the privilege of a few large companies. However, I remain optimistic about the continual improvement of open-source models. @erikreimann5556"
If you see "confidential" on an Intel processor, that means it's an engineering sample(pre-release CPU). It's from some manufacturer/OEM and made it's way to ebay. It's not technically legal to sell(property of Intel), might not function correctly(early engineering samples of a product may be substantially different than the release product), and may have been intentionally mistreated(for example, if it went to be used in cooler design). There are tons out there, it's not like you're going to get in serious trouble owning one, but you're never going to get any support/warranty, and you have to hope you don't get a weird/early/fried one.
@@Grimmwolddsin my 30 years of building stuff and dealing with PCs in general, I have never had to rely on any kind of CPU warranty. In fact usually a CPU either works or it doesn't and in most cases as a mere mortal you will not get any kind of meaningful support except for if that particular die had issues. And even then you might've had to contact your dealer instead of Intel. Sure, for bigger enterprise companies, there are better support contracts and SLAs, but they usually are very expensive and don't make sense anyway. That said, buying used gear on e.g. eBay is always risky anyway.
"Genuine Intel CPU 0000%".. nothing smells more trustworthy than a CPU that doesn't even know what it itself is. I'm pretty sure that Chinese Engineering Sample is just a production reject that *should* have gone to recycling but didn't. I hope it stays stable enough for you, but I wouldn't be surprised if it didn't.
Congratulations on your new system! Thank you for sharing the experience.
And a backup whole house heating system!
Congratulations on ur purchase and new powerful computer. May u have a great journey for any research/learning on it.
Thank you so much!
Congrats on the new system! I am sure you are going to have a lot of fun putting it to work. :)
Grats to your power bill!
Mmm...that core count 🤤 I wonder if my neighbor would notice an extra power cable going to his fuse box 🤔
Niiice HEDT! Mind you, optical audio is (very) limited in bandwidth, hdmi does more channels, higher bitrates and does so uncompressed. You could get RGB support from an internal add-on box, from razer for instance, sacrificing an intenal usb connector.
Note - the LGA 4677 motherboards are not easily found at a reasonable price. Chinese Ebay sellers sell them at a premium compared to local - if you can find them that is.
agreed
Scouzi
Homemade boards, chips found on garbage boards ! Soldering it together boards !
Why you need this, render server ?
Why you need to find this ?
Yeah Gigabyte MHBs were going for about $500 before 4th gen boom with AI. Now they're up $700 and being resold at $1200 now. But 3647 and 4189 aren't really worth it anymore, but still better than Broadwell Es at least.
I'm from Brazil and I loved your video. I found this processor on Alibaba and I'm thinking about buying it... I'll analyze it a little more.
Fun fact, 20k price is meant to act as a scare for retail customers, yuge cloud providers pay right around 500$ for such cpus.
well thats becasue they buying 100s of them
@@ThatZenkiS14 it does not justify 95% discount. we've seen exact copies of a cpu retail vs xeon (lga1366) with xeon priced 5-10 times more...
I doubt they pay that much. NVIDIA charges $30k for an H100, but the lowest price on average I've seen for customers that buy hundreds of them or even thousands of them is $3,000-$4,000 per GPU. The steepest discount I have ever seen on an Intel CPU for a single customer was for a pair of Xeon Platinum 8380s, which being $20k for said pair, were being offered at an 80% discount to just $4k. Of course, this discount was offered through a workstation retailer/vendor to a client that was likely well connected. You don't need connections to get an engineering sample CPU like the one shown in this video.
@@NUCLEARARMAMENT I used to cost such things like getting quotes for 100s of Xeons based servers. Any relatively large corp can easily get 60% discount off list. No one pays list. Anyone agreeing to list price should be fired.
@@scouzi7201 I wholeheartedly agree, even someone buying a single workstation with one or two of these CPUs can get half-off the list price.
I've always been wary of engineering samples. Do all the PCIe lanes work? You could maybe test that by putting the GPU in different slots. Does AVX-512 work?
Avx 512 and every slot is working 😁
Home made from old trashed parts, resoldering !
work, hobby ?
Why he needs it ???????
Mad here !
RGB, optical......not going to be added to a board with more PCIe slots to add what you want and still not fall short of PCIe slots.
Great video ! Could you please provide llama.cpp inference speeds for CPU alone for some llama 70B model e.g. Q4 and for example mistral 7B Q4 or maybe you already posted the results somewhere ?
I will do a video about it and also about other deeplearning benchmarks
Wow thx for being in my feed! 600 bucks US for a low clocked 60 core! Might be a good match for Folding at home or BOINC scientific computing however how cheaply can 1 get a motherboard for it if i don't need any PCI express lanes?
The cheapest is from Supermicro for about $600
Great find sir
nice server system. I have a dual xeon 1366 system with a few sets of cpus. My modern ryzen beats it though by a lot.The 6 channel memory is still faster than most things today.
Why you need this, how many users on this system ?
Running what ?
If you have thousands and thousands of users, you will benefit from form 6 channels ! all needing the same resources, speed is essential !
@@lucasrem I bought it years ago. When you have simple low powered systems for years you get some money and buy something crazy. Thats what I did and it was cheap. I run windows and use benchmarks for fun. Now I use a ryzen and have that system in the corner of my room unused. Sometimes i play with it like any nerd would.
Hi! That's exciting. I have been looking at 2nd gen Epyc and 4 Socket Broadwell recently for a single-node cluster solution, but this would also tick the box. Can you tell me what all core turbo you can get? I imagine much more than the 1.7Ghz base? And do you have a passmark score by any chance so I can compare it to let's say 2x 7D12 Epycs?
well the Xeon Platinum 8490H which is 1.9ghz boost up to 3.5..
You are supposed to remove the brand sticker from the M.2 PRIOR to installing the cooler, that will make it worse if you leave it on
That is awesome. Nice build. Server chip sets are different then standard ones.
Thank you!
@@itsmorefuntocompute3442 I know you knew that... Nice build killed my 12 core threadripper
Hey this might be an odd request but I wanted to ask if you could maybe try to compile chromium as a benchmark, I'm looking for a new work station with that as my main benchmark since it's the majority of work I do and would really love to get an idea of what I can expect with these CPUs.
You're definitely going to want to invest in a case with a fan config you're confident will comfortably cool at least a pair of passive 2-slot accelerators. Even if you don't plan on going nVidia, the meta when it comes to accelerators is high-density capable cards (those don't come with fans), especially when it comes to used cards.
I put it in the Be Quiet Dark Base Pro 901
@@itsmorefuntocompute3442 Let's see... Just eyeballing it, if you installed it in the 901's inverted layout (PCIe slots on top, CPU towards the bottom) you should be able to just fashion an air guide towards the cards, and a manual fan controller for the topmost fan. That'll give you the room and cooling to fit 2 accelerators, no matter who you choose. Be it nVidia Tesla, Radeon Instinct, Intel IA-series, or (if you're lucky enough to have that kind of access) Google's Coral Server series accelerators. Just remember to bump your fan speed if either card hits 80c, most will self-throttle at that point to protect the card, but won't complain or crash.
P.S. If you do choose nVidia Tesla or Intel IA, DON'T use an 8-pin PCIe power cable on the accelerator. Tesla cards from Pascal onwards, and most Intel IA cards, use 8-pin EPS Power (sometimes labeled "CPU power" on PSUs) which looks identical to 8-pin PCIe, but have an extra 12v rail and an inverted pinout. If you attach 8-pin PCIe instead of 8-pin EPS you will blow the card and likely your PSU too. You can get PCIe to EPS adapters if your PSU doesn't have the extra EPS cables, just make sure you get one with a thick wire.
@@itsmorefuntocompute3442 What about the CPU cooler? Is it enough at full load vs liquid cool?
@@scouzi7201 Don't waste your time with air cooling. I built a dual socket custom cooling loop for LGA-4677 (Socket E, which is what these Sapphire & Emerald Rapids CPUs use), and the price for the D5 pump and reservoir combo ($80), two Byski LGA-4677 waterblocks ($100), EPDM tubing (15 ft, about $50), 16 G 1/4" BSPP threaded soft-tube compression fittings ($50), and a couple of 420 mm Hardware Labs GTS 420s (30 mm thick/420 mm long radiators, $150), all came out to about $430, which is nothing for a system like this that will easily dissipate 360 W per CPU or 720 W for both chips at full load without any overclocking.
1,000+ W dissipation for two of these [60-cores] even mildly overclocked (10%-20%) is not unheard of in heavy-handed in either vectorized or multithreaded workloads. In fact the Xeon W-3495X with 56 cores at stock manages 700+ W heat loads in AVX-512 full-width vectorized workloads, so two of them would hit 1,400+ W at stock. The air cooler (a Noctua UH14S meant for LGA-4677) was hitting 90-100+ C whereas the max junction temperature for this CPU is less than 90 C, and you don't want to operate above that 24/7 especially when you need to run mission critical workloads for extended periods of time. The heatsink is also $125 each, so it's cheaper than the $430 watercooling setup I specified earlier, but your temperatures will be at least 30-40 C worse for the wattage heat dissipation.
Trust me, spare yourself the headache and invest in watercooling, and enjoy an extra decade or more in lifespan versus air cooling; not to mention you never have to worry about case or DIMM clearance issues again if you go with watercooling.
@@scouzi7201 it's strong enough. But I would recommend the 140 millimeter version.
Which case did you use to house your system in?
Bw quiet dark base pro 901
Welches Computergehäuse haben Sie für Ihr System verwendet?
Keiner hat sich so über ein Recovery Menü gefreut!😊
How did you get that for 500? I see it for 5k still. NM, I found some for 580. Oh man this just got me thinking about a build to make.
Were?
Its because he bought a testing product, compared to buying the official concluded product thats currently for sale.
@@kodekorp2064 No shit as I said I found it for 580 aka an engineering sample. I think we already knew he didn't buy the non engineering sample.
How much watts does it consume in stress tests?
The board layout is super weird. The first few PCIe slots will be constrained by the ram sticks.
Probably all slots are x16 (this platform has many PCIe lanes) so you can use GPUs on the lower slots
Cool! May i know if the motherboard can do x4x4x4x4 bifurication in all slots?
Nice im also working on a epyc system built its a Supermicro H11DSI dual-socket motherboard im using the epyc 7551 and the cooler TR4 sp3 i and for the ram 8gbx8 64gb so far
ES models are a hit or miss.
Could you make a tutorial video on how to get started with gpus for deep learning, like cuda installation, tensorflow on windows etc
I will think about it!
Someone used an QYFR 8480+ could even blast off 62k with an w790 on serverthehome. Should you tunning the bios for performance?
How much the price of that cpu? This cpu just ES not OEM
Thanks. What case are using for this motherboard?
What was your idle power consumption at 0% load? Im thinking of buying a xeon 5th gen system soon.
Do all the accelerator engines work like the final version? AMX is the big one for AI workload.
AMX is working like at the final version. Only the base frequency is more low. AMX is a big leap!
I bet it could handle many open chrome tabs :D thx for sharing
Would run out of RAM first... it's the Chrome extensions that suck up all the RAM.
Turn up the lights when you do videos, just a thought....:)
Was reading the specs sheet for this. Surprised to see 16 High Priority Cores @ 2.0ghz + 44 Low Priority Cores @ 1.60Ghz. However it says the “Base Frequency” is 1.90Ghz, how do they get that number, an average. They claim Max Turbo Boost is 3.50Ghz! YEAH RIGHT!
First of all, How? unless its a sample even if used.. you know what Iam glad for you brother! it totally demolishes threadripper which is 500USD
I mean the ram & motherboard are expensive as is..
What is the power draw on that beast under load?
No sane person would build a Threadripper if they knew about the existence of these CPUs and realized the performance to price ratio, unless they were a financially illiterate AMD fanboy.
350w to 400w
@@NUCLEARARMAMENT Yeah; No sane person would want a cpu that’s guaranteed to work, has a warranty, full motherboard compatibility, draws less power and is far more reliable….
I know you lack common sense but this is a EARLY ENGINEERING SAMPLE, hence why a 1 year old data center cpu is selling for 1/40th of its “retail price”….. You’ll have to be incredibly stupid to use this in a application that requires high availability!
@@ThatGuyPal88 You don't know what you're talking about. Crawl back under your rock, Patrick.
Can we see single core performance?
good luck with the power bill
It's a 350W CPU and server parts run at TDP. They don't boost like desktop parts do. This 60c part probably consumes less power than a modern i7.
22927 difference with the Ryzen.
This is only for companies and people with money, great acquisition. it's time to hunt for offers on these machines. Thanks for share the power.
And 7 PCIe x16 lanes ;)
Nice to see an Intel ARC A770 GPU, trying to obtain one myself too :D
how does it score on xmr? how many H/s?
You dont need an rgb header to have rgb. RGB controllers attach via usb.
Thanks for taking the time to share this. I was looking for a similar purchase , this sample seems to be working quite well. Do you have the exact parts number or eBay reference ?
I bought it from the seller sinobright. Part number is difficult, because often the pictures don't match the final product
nice but don't like the cooler, its not suited unless there is not too much heat from cpu the ram should be ok, but still front to back never fails.
Thanx for sharing the video ! 🫡 I looking something similar for finite element analysis. What was all core boost frequency ?
@@Dmitrij_S All core frequency with a heavy workload was 1.7 GHz.
Link us to the store where you got that beast xeon!
sinobright on ebay
Supermicro actually offer a dual socket board for this CPU - just when you thought this one was insane!
I saw that! I went for the single socket, because WSL only supports one.
Awesome CPU for the price. But, what did the whole system (motherboard, memory, etc.) cost?
About 4000€
That is great. Congrats.
have you tried to run a local LLM with this setup ?
for llama2 7b :
prompt eval time = 265.29 ms / 18 tokens ( 14.74 ms per token, 67.85 tokens per second)
@@itsmorefuntocompute3442 and what is the eval time ? that's the value that show the token generation speed.
llama_print_timings: load time = 13387.07 ms
llama_print_timings: sample time = 74.30 ms / 400 runs ( 0.19 ms per token, 5383.36 tokens per second)
llama_print_timings: prompt eval time = 308.04 ms / 19 tokens ( 16.21 ms per token, 61.68 tokens per second)
llama_print_timings: eval time = 91794.32 ms / 399 runs ( 230.06 ms per token, 4.35 tokens per second)
llama_print_timings: total time = 92486.77 ms / 418 tokens
Log end@@adamstewarton
Get rufus and it asks if you want to remove secureboot/tpm requirement when you burn Win11 ISO to pendrive.
Excellent; what kind of DDR5 ECC (specs) are you using in your setup? Thanks
Officially, this CPU supports 4800mhz in 1 Dim per channel mode. But in the BIOS was a 5200mhz setting, which is working 😁
60 cores!
hell yeah
not bad, but Cinebench record is around 129 000, for waht tasks is this build?
For finetuning of LLMs
Cool build, have you seen those neuromorphic clusters? They look so cool.
Really cool! But they have to prove that they can utilize their raw performance
Bin innerlich etwas gestorben als du angefangen hast dir RGB und Audio für ein Server board zu wünschen 🤣
I've subscribed, i want to see where you take this beast!
wow you have bought very expensive this xeon gaming awesome :)
how much was on the ram?
Have you try mining on?
Very interesting to see an ES working ! do you think this kind of cpu can works on w790 chipset ? The bords with w790 are estheticaly nicer as you mentioned in the video.
No, those are workstation boards. Get a workstation chip for a workstation, and a server CPU for a server.
@@timramich Hello, i did some research about server xeons on workstation motherboard. It seem it can work on some board (Asus's, gigabytes's and some asrocks's) if the microcode for the D0 stepping of engeneering sample CPU is still present in the bios. Intel only seperated their xeon lineup in 2 separate brandings (Workstation and scalable) at the E0 stepping . Just to prepare the consumer samples we can buy. So your 8490h ES could work on a w790 mb if it is a D0 (probably since it is the most common).
looked like a bunch of bent pins near the triangle
What is "AGI" You talked about when you mentioned AI?
AGI is artificial general intelligence. There are different definitions, I would define it as an AI which has human expert abilities in a wide spectrum of tasks. I think humans would have called GPT4 an "AGI" 20 years ago.
@@itsmorefuntocompute3442I'd probably have called ChatGPT4 (which I do use myself) a set of smart algorithms. However: ChatGPT only reacts to something - no matter what it is.
An AGI should be able to proactively act instead of just reacting. So given it knows how the world works, I'd expect it to want to improve / change it (for the better or for worse).
But as you said, there are different definitions of what an AGI supposedly is.
how much did this all cost
Hi. I don't never heard this server cpu. On ebay i've found Intel Xeon Platinum 8490H ES, but what means ES?
Engineering sample. It's for manufacturers to test the mainboards and coolers they design.
Can you try mining a CPU-based crypto like XMR or Zeph on it? I am interested in the amount of hashes per second this thing gets.
Will not be as fast as my raspberry pi five.. 😎
Love the accent. Had to suscribe just because of it. Keep up the good work!
Thank you so much! I am really glad to hear that 😊
What happened to your dual AMD EPYC 7601 board ?
I sold it on eBay again. I like to use WSL and WSL only supports single socket systems.
Learned something new :--( Not good for my (refurbished) HP Z840.
@@iforth64 You should be able to use WSL by running Windows under KVM on Linux on a dual socket system, but I guess that'd be redundant. Ha!
350W based clock.
This thing, consumes bout 2.4x what my whole rig is eating up.
Damn, and it's just the processor
these things sip power regardless....
funny how over the last 25 years i seen things go from OMFGURD this is alien tech is hardly draws any power !!!1 and the SAME EXACT CPU 10 years later, everyone things its a "nuclear power plant"...LOL
Fucking morons.
This platform was doa because of AMD’s Epyc. Still a great find and one that gets my brain thinking about buying one for myself 😂
Can it run any LLMs?
this is llama 2 70b:
llama_print_timings: load time = 44755.73 ms
llama_print_timings: sample time = 47.43 ms / 241 runs ( 0.20 ms per token, 5081.71 tokens per second)
llama_print_timings: prompt eval time = 2945.46 ms / 21 tokens ( 140.26 ms per token, 7.13 tokens per second)
llama_print_timings: eval time = 195141.13 ms / 240 runs ( 813.09 ms per token, 1.23 tokens per second)
llama_print_timings: total time = 198337.28 ms / 261 tokens
Congrats.
What for are you going to use this machine ?
Finetuning of big LLMs (that's why I bought a CPU with AMX instructions) and compiling
Cpu or 2.5” hard drive?
😂😂
Just need an $1,800 USD motherboard to go with it. Then add DDR5 RAM (non-ECC is cheaper..)
I paid 1070 for the Mainboard ;)
There is another noctua cooler that blows in the other direction
The noctua cooler in the video was too little and too loud. I swapped it against the 140 millimeter version.
8490H ES*
sergi ?
Why you dit buy that, running what ?
Render Server, you make animation movies ?
I need it for finetuning of LLMs
Jesus !! Motherboard above the static bag!! Don't do this most of they are more or less conductive and can damage the mobo.
8470 is arguably better due to base clock, no? 8490s are better to be found than 8480+s. Why's this?
why did it dropped ~8k?
60 cores 120 threads.
This is overkill, maybe usefull in the 2030s for gaming lol
The worst thing is that it's clocked at 1.9Ghz based with 350W and can overclock to around 3ghz.
112MB of cache, that's gianormous
nothing is overkill, for is good
And a modern 24-core processor will have much higher performance, and much lower power consumption.
Modern PC will also have much faster RAM and SSD.
@@DerekDavis213 but cost much much more.... Nobody is building these things for 100% top-of-the-line-performance...SMH
@@trumpsextratesticle8590 How much is your time worth? For many people, a small investment in a modern PC, which you can build yourself, is better than spending lots of time and effort on a dusty old obsolete Xeon, that still has many compromises and issues in 2024.
I admire Ford pickup trucks from the 1940's, but I wouldn't buy one and spend many many hours fixing it and refurbishing it! Xeon is the same way.
@@DerekDavis213 If someone is video editing or whatnot this Xeon build is just fine. Most people are not doing enterprise level shit or needing that security level.
Most people would be fine on a Xeon, even gaming on almost all new titles.
If you don’t mind what was the name of the seller on eBay?
sinobright
The motherboard costs between 1200 to 1800 on ebay
I paid about 1070. Not cheap 😢
Its a lot of ultra low-loss PCB material with a lot of PCIe retimers to support PCIe 5.0, these things aren't cheap
350 Watts baby....
Wow. I'm jealous.
sehr gut
How is it on Flux...? lol
very nice
How is it going in game bro ?,
where are getting the mobo from
When it is worth $20k and works - why do they sell it for 500 bucks?
It has a 200mhz lower base clock. And I think big companies and data centers can only buy the officially supported products.
@@itsmorefuntocompute3442 Thanks, then congrats to your new computer! But I have to tell you that designing a gerneral AI is probably not only a question of compute power 😉
"I mentioned 'execute,' but should have used the word 'inference.' The design and training of models nearing AGI capabilities are currently the privilege of a few large companies. However, I remain optimistic about the continual improvement of open-source models. @erikreimann5556"
If you see "confidential" on an Intel processor, that means it's an engineering sample(pre-release CPU). It's from some manufacturer/OEM and made it's way to ebay. It's not technically legal to sell(property of Intel), might not function correctly(early engineering samples of a product may be substantially different than the release product), and may have been intentionally mistreated(for example, if it went to be used in cooler design). There are tons out there, it's not like you're going to get in serious trouble owning one, but you're never going to get any support/warranty, and you have to hope you don't get a weird/early/fried one.
@@Grimmwolddsin my 30 years of building stuff and dealing with PCs in general, I have never had to rely on any kind of CPU warranty. In fact usually a CPU either works or it doesn't and in most cases as a mere mortal you will not get any kind of meaningful support except for if that particular die had issues. And even then you might've had to contact your dealer instead of Intel.
Sure, for bigger enterprise companies, there are better support contracts and SLAs, but they usually are very expensive and don't make sense anyway.
That said, buying used gear on e.g. eBay is always risky anyway.
Whoa!
Use Windows 10 for Workstations and run those tests again
RGB on a server board. Gamers really do think that the whole computing universe exists to serve their needs.
"Genuine Intel CPU 0000%".. nothing smells more trustworthy than a CPU that doesn't even know what it itself is. I'm pretty sure that Chinese Engineering Sample is just a production reject that *should* have gone to recycling but didn't. I hope it stays stable enough for you, but I wouldn't be surprised if it didn't.
I can set up steam sharing if you want to test some games :) 7 days to die loves memory bandwidth
20 fps at 2 chunks