Це відео не доступне.
Перепрошуємо.
Milan X: 768mb L3, What's it Good For?
Вставка
- Опубліковано 20 бер 2022
- **********************************
Check us out online at the following places:
+ Website: level1techs.com/
+ Forums: forum.level1techs.com/
+ Store: store.level1techs.com/
+ Patreon: / level1
+ L1 Twitter: / level1techs
+ L1/PGP Streaming: / teampgp
+ Business Inquiries/Brand Integrations: Queries@level1techs.com
IMPORTANT Any email lacking “level1techs.com” should be ignored and immediately reported to Queries@level1techs.com.
-------------------------------------------------------------------------------------------------------------
Intro music: "Earth Bound" by Slynk
Other music: "Drifting at 432Hz" Unicorn Heads
Outro music: "Follow Her" by Zeeky Beats
That's not a washing machine, it's a simulation of a mixing vessel. This is incredibly important in chemical reactions at industrial scale as you want your reagents, whatever they might be, to be uniformly mixed to get the reaction to run at maximum efficiency. Ran some of those in the past and they're super interesting as they combine CFD with mechanical aspects of the vessel and "propeller" that's in there in a way that makes standard airplane propeller scenario look simple in comparison.
I concur.
Chemical reactor vessel with pitch-blade turbine and baffles, not a washing machine.
Working in chemical engineering and evolving close to people doing computational fluid dynamics simulations, Milan-X looks like tremendous news!
In fairness there was a belt sander in the mix, too. :D
@@Level1Techs To be fair, a wash machine is a "mixing vessel".
While I agree, it didn’t look like a washing machine, I do think it likely that the big manufacturers do model washing mashines these days - my washing machine today is very different than the one I grew up with
It’s also irrelevant what it is- it’s some fluid dynamics sim is the point
Holy crap. This is astonishing, even though I have no one to geek out and get as excited about it as I am!
I think it would be interesting to see this 3d-stacking tech trickle down into AMD's APU lineup. Having a large cache in-package could really improve integrated-graphics performance (à la Crystal Well).
Wendel has me all excited for things I barely understand
Wendell should wear some ear protection there 😂
As an ANSYS CFD user, this is the EPYC release I’ve been waiting for!
ANSYS has been optimized for AVX-512 to help feed the cpu in the past and the lack of support from AMD always left doubt in my mind about whether EPYC could keep up with Intel for my workloads. However this seems like an innovative alternate to AVX-512…nice work AMD!
I heard Genoa will support AVX-512
@@rezaramadea7574 Only partially. It will do some parts on 2 x 256 and some parts on 512 (AMD claims that they get around 90% of Intel efficiency with that. I'm not sure whether they mean per core or per CPU). I think it's going to be Zen 5 EPYC (Turin) that will have full AVX-512 support.
I am seriously considering these for my companies ESXi server hardware refresh
I wish they'd do a 5950x with 3d stacked cache, or at least a threadripper line with 3d cache. I'm a grad student, so broke af, but I do simulations of simple reinforcement learning that are cache limited (like after I increase an array size to a certain point I all of a sudden get a 10x increase in how long it takes to run the simulation, so I''m assuming that's a cache issue). The 5800x3D is not enough threads for me. And yes, I know I'm in a very niche scenario and AMD has an obligation to go after the highest profit margins which are going to be server CPUs, but I just needed to vent😮💨
Threadripper for sure
Finally, a processor with nearly sufficient cache for Dwarf Fortress.
It has more cache than my first PC had ram, and my first PC was a pentium 4 running windows XP
That’s the thing - this is just natural progression of the chip industry - nothing godly so to speak. Tangential to that,Wendell mentioned 0.5 TB of RAM - I’m sure that itself will become an everyday thing even for desktop processors in a few years. This launch however should be a nice kick to the other chipmakers!
Dunno if you could even find a application processor these days with less cache than my first computer had RAM, a whole 64k (and a tape drive!). Even if you just counted my first IBM compatible - that was a whole 640k RAM with a massive 64k graphics RAM lmao. Even the colossal 42MB HDD is less than one core of cache on this beast, or the cache on my 3900X.
My first PC had a 360MB hard drive! :D
Great video! This is another video that demonstrates absolutely conclusively that of the world did not already have a Wendell , we would have to build one from scratch atom by atom!
Thanks for the mention, Wendell!
Please could you do some RAMCACHE and RAMDISK benchmarks using PostgreSQL, MySQL, MongoDB for Web Servers IIS, Apache, and Nginx.
It would be great to see build time, response time, TTFB in ms and any other caching options to see the best caching option and hit rate.
IBM has had monster cache thats L2/L3 shared between cores, but they are so niche I would never see one.
Great Scott!
Not a clue what you are talking about and don't see a time where I will ever be able to use the information. But I just watched the whole video and found myself enjoying every minute of it. Thank you sir!
He pretty much said "The company what was laying off even it's cleaning personnel in 2016 to avoid bankruptcy is hitting Intel with yet another big problem they have no immediate answer to" :P
At some point this year, L1techs became my favourite tech channel.
superb presentation!
Amazing video Level1!
Glad you mentioned containers - I am looking for the ideal platform to run Kubernetes - 40 to 80 containers. - Good tip about L3 cache miss logs, but I dont have a platform yet. - I wonder if you can do some tests with Kubernetes? - I guess the threadripper pro system you have featured previously might be good enough.
Glad you're plugging Brendan Gregg, he's a systems administration god !
Hey thats in Austin! Love that campus.
FreeBSD's ports tree could be used to compilation. I tried to find a reference for how long it takes to do a build of all packages but couldn't find one, the time that comes to mind is a week. Poudriere and Synth are two tools that are used to build packages in bulk, doing a specified number of parallel ports builds and making sure that all prerequisites are built first.
These look awesome. Let's get them on into that Torrent case!
you can literally buy 2 different 3D CFD packages from Ansys and they give you different results - friend worked on Mercedes F1 CFD team 2018-2021 - vortex generation
A great video Wendell as always. Sitting in the same room with the major competing CAE vendors must have had an electrical atmosphere! Great progress from AMD for the CAE/CFD community... we need as much faster-memory access as we can have, it's never enough. Do you have any conclusions in relation to 64core CPU price/performance/scalability wrt to 32core CPU with this increased L3 cache?
Are the presentations from Ansys, Siemens and Altair you referred to in your video public? Thanks.
This was great to see some CFD and FEA talk. I definitely see this being a killer on clusters and blades. I have been using Ansys Fluent and StarCCM+ inside a CentOS docker on Fedora and its awesome. You can do some neat things in CFD, not only for prototype and testing, but for even research in more theoretical realm of science. Having access to Fluents UDF, or openfoam code or a custom solver code is mind blowing and having more cache is gonna be a killer. I got me wondering how much I save shave off openfoam vortex shedding over a box with more cache. I am also excited for a prosumer version of this. Only a student right now, so can't really afford an epyc yet lol
Not trying to crush your but, AMD wouldn't sell you a Milian, it's vendors only.
@@carl8790 waiting on tr pro with 3d cache if I really want one, but tr pro is expense af too
@@Infinit3Enigma true. I'm disappointed in AMD for discontinuing their non pro TR. I get why they're doing it, doesn't mean I support it. A lot of pro users with limited budgets appreciate the hell out of those chips.
This is so exciting!
Some people have asked why other CPUs havent tacked on this much or more cache. Its the same reasons that we have multiple cache levels.
You need to have a lookup table for the cache, the bigger the cache the bigger and slower the lookup table is. Also the bigger the cache the better your prediction algos need to be to utilise it well. There's also obviously power and heat considerations for very fast RAM as well.
Being able to tack this much cache on a processor isnt THAT impressive........ being able to use it WELL is.
I guess cost might be a major contributing factor as well lol
@@allesklarklaus147 especially in the context of per core licensing, the cost of the CPU isnt really a major factor in the lifetime cost of the sort of machine this is designed for
@@mycosys Well.. yes and no. I would say you still need to offer reasobale price/performance in comparison to your competitors. Tacking on half a gig of cache would be pretty.... interesting just a couple of years ago. And a couple of years really isn't that long when it comes to cpu development where the development cycle apparently is pretty long (like, 3-5 years or so)
hence the huge advantage of being able to stack in 3D - being able to locate all that cache in a small amount of space around the core, keeping latencies relatively low (albeit a tiny bit higher than before I believe)
Don't know what I'd do with it, but 🤤🤤.
Now we just need 64GB L4 cache!
YES! Only 20 seconds and we have a bananas already. This is going to be bananas.
I think this the mandiant 1U I used to have in my home lab was louder. I didn't run it often because it bothered my kids...who were upstairs.
6:50 Siemens reverse engineering Miele's Hydromatic (2022, colorized)
Didn't know they used simulation software at BSH to optimize Bosch and Siemens their appliances, interesting.
Can you try to do a chrome compile? Chrome compile is spending 1/2 of it’s time in linking and I assume that could be improved by the increase in cache.
It varies! Chrome out of the box compile wasn't helped much but running multiple or parallel jobs? Maybe. Still working on that
@@Level1Techs I'd like to know too. Chromium is a 50,000 source file compile, all C++ code. Theoretically the footprint is large enough that it should be able to take advantage of the additional cache.
(We run chromium compiles on our 3990X w/256GB of ram, along with other bulk package building, using TMPFS to hold all temporary files including extracted distribution files and object files during the builds).
Not that I could afford this particular baby. But it would be interesting to know how -j 128 parallelized C++ compiles do on it.
-Matt
@@junkerzn7312 Why would anyone compile Chrome?
@@patrikjankovics2113 You understand that before you, as a user, can run the chrome application, some vendor or project has to compile it up from sources for you, right?
All of the applications you use, and all of the other applications and libraries (tens of thousands of them, with millions of individual source files) have to be built in order to create a cohesive result for the end-users who just do an apt-get install or a pkg install or whatever to put something onto their computer.
Vendors, large companies, and open-source projects are constantly upgrading various component sources and effectively rebuilding all of the applications and libraries that depend on those components, often for multiple operating system and distribution targets. Even a single change to a single library can require ten thousand dependencies to be rebuilt.
Thus those of us who do this type of work care a lot about how well machines handle concurrent compiles. And chromium... is one of the biggest. 50,000 C++ source files alone, not even counting the 100+ dynamic libraries the binary depends on. (Chrome is probably one of the most bloated applications that has ever existed in open-source).
-Matt
@@junkerzn7312 Hey, thanks for the length reply.
Yes, of course I understand that. But at first sight I thought the compiling is done for in-house use, and for that it doesn't make a whole lot of sense not to use pre-compiled binaries.
But, now that I've read your comment, from my understanding you guys are the ones actually supplying the pre-built binaries?
Also, yeah.
No matter how much I love C++, I always end up having to acknowledge that Linus Torvald's is right about C++: It becomes hardly manageable over time, so is the case with Chromium. Although.. Well, it's a debate of it's own, so let's not get into it here.
an vlog style intro..... still took me by suprise, i'm still trained to mentally expect windell to be a slither of face behind 4 screens and a green background....
server fans go wrrrrrrrrrrrrr
thanks Wendell, very cool to see all this. this is super exciting to see
I've been WFH so long I forgot that noise. Don't miss it.
".. I'm sorry, the Linux kernel? Just not that complicated." Linus wishes to have a word with you... Good job of explaining the new tech as ever, with useful add-ins from industry insiders who are already familiar with the new tech. I won't be getting a Threadripper of any description any time soon but it remains at the top of my list of kit to buy if I ever get the cash/ (I need cash to afford cache!)
Funny thing is that Linus *wants* it this way :P He is known for bashing people over the head if he thinks a certain job could be done in a more straight forward way.
@@andersjjensen I get the impression that kernel merges are not a sunny day at the beach....
All the compute and all largely in a context outside of my reach...but I'm okay with that since I'm more of a software dude than hardware...still sick though. And I can definitely respect and appreciate the level of technology here.
_engagement_
I mean, I know that's not how this works.. But imagine running your OS within Cache and not DDR memory!
Definitely not windows though!! Windows just eats memory doing nothing
@@abdelfiala true :D was thinking more like ubuntu server etc which fits easily within 100-300MB after boot
@@ibonitog agreed
Uhm, that is exactly how it works... L3 cache eviction doesn't kick in until the cache is full. Then it runs an LRU (Least Recently Used) eviction policy. So, strictly speaking, if you only put 1.5GB of RAM in your box, and you have 1.5GB L3, then eventually you'll have all of it cached and nothing will ever get evicted :P
My spare laptop runs Slackware with XFCE on 2GB of RAM just fine...
For small applications and let's say 10-30GB of process maybe it works faster because of the quicker than ram transmissions in bursts. But the CPU when is topped with a process that is 1TB, as a whole it makes a smaller difference. When you need to modify 500GB, that cache won't suffice.
Depends on the access patterns. It's pretty unusual to see truly random access..CFD is regularly the size you describe and can go super linear if the job is sized right across the cluster.
AMD made the correct choice with their direction. Larger L1 and L2 cache are next.
I would love to see how this rips through the PostgreSQL workload in my company's stack.
it's so nice to be reminded that computers are being made for, you know, computing shit! Seriously. I wanna see those 3 car crash simulations! That's some shit you used to simulate by crashing 3 cars together!
Would love to see a standard test for showing graphics rendering (as in gaming not blender) on these CPUs (instead of GPUs) looking forward to when we potentially don't "need" locked-down graphics specific hardware. ex. can this CPU "emulate" a VooDoo2 or a GTX 560m maybe? Might bring in viewers because no one really does that?
TL:DR; I want to know what GPU these 128 (much more general compute) cores could equivocate to, because... CPU intrinsic good, GPU API calls too closed and abstracted.
(I know they can't be compared directly... but I want comparisons.)
I mean - you can run older games on CPU rendering only and it works fine. Just not very high demanding unoptimized games. Crysis 1 for example is notoriously badly optimized, but in Low settings on 64-core CPU you can achieve around 18-20 FPS. LinusTechTips did video with that. I don't watch a lot of them, but it was in my recommended so I gave it a try (normally they are too "mainstream" for me if you know what I mean). So you can achieve decent performance in many DirectX 9.0c games and older. And it's often not a matter of compute time, but combination of many factors of which raw CPU power is only one of. Memory is other, but there is also the fact that a lot of games have their software renderer so unoptimized (because no-one uses their CPU as their GPU for gaming). And also in the end you would still need graphics card to run an OS like Windows - so your monitor would still be connected to GPU - which is an additional hassle. I'm rather wondering if this technology could not give us APU's that are fast enough, though you'd either have to have obscene amount of L4 cache - which would be very area expensive or fast memory on the chip itself - like HBM. Which again is expensive.
I expect that APU's that will work with DDR5 might be able to do something more than current ones (I'm not counting Intel iGPU, because it's simply not powerful enough - yet), but that will then highly depend on what memory you're running.
Theoretically you could try to run games that require much less than 1GB of VRAM on such CPU (because CPU itself will still need L3 for itself), but again you run into the fact that not only CPU's are not built to handle such specific tasks, but that engines are utterly unoptimized to run on them. It's sort of vicious cycle, but again - no one is thinking of running games graphics on CPU for well over 2 decades now. Basically last engine that was somewhat optimized for this (and not superbly well, but at least it was) was Unreal Engine. The original one. And since even back then everyone had graphics card (but they didn't have 3D accelerators necessarily - which are the parts that run modern games on GPU, which now is combo of old Graphics card and 3D accelerator) this optimization was only partial. It still required card to have proper display.
I'd like to see Fujitsu's A64FX married with 3D-VCACHE technology (next to an update of their used HBM stack). That could make that already capable chip even more efficient. My guess would be that if used as a massive L3 cache, VCACHE might help to overcome some limitations in the memory system, as they lack the capacity of DDR memory, relying solely on their on-package HBM instead.
so much cache they might have to write a rap song about it.
Love your vídeos! Would it be a good video to test database (PostgreSQL or MySQL) on modern AMD CPU with super fast SSD (millions of IOPs) amd compare this previous hardware.... and maybe every 2 years or so to test it again on the lastest a d greatest harware available. Thanks
Woot!
I love it so much when Wendell says “BANANAS”!!
Your...bench network looks like mine at work. :D
Such a niche product, but it's so cool tho...very fitting for this channel.
very nice 😉
"We got one" to Quote Stefan from TheVerge :)
Would love to see what the performance uplift is for Apache Spark. I’d assume it depends on data size, amount of shuffling, and the job itself.
As a data engineer I’d enjoy seeing other tools covered like Kafka, Presto, Cassandra etc. on the channel.
Yeah, the key really seems to feed this thing data in just the right way that you can fully take advantage of the cache, but with over 2x gains to be had it seems very much worth the time to spend optimising
This is INSANE!
The wailing noise is also "epic".
Milan X gonna give it to ya. Hearing damage that is.
More cache per core than my first hard drive. nice.
It's great they brought Mulan to computing.
is there anything from team blue that can match milan? even if you ignore the utterly dumb pricing team blue has.
You should hit up Linus regarding that $1 Million dollar unboxing... a Petabyte worth of Flash storage with tons of custom hardware.
Whats this push a 6900xtu to on timespy PMSL
I think Wendell needs to go have a lie down in a dark room :D
My backup-server does not have a 1.5GB cache but a whopping 2GB DDR (400MHz).
what does more work in a simulation a processor with half the cores (32/64) and the new cache or double the cores(64/128) with the old cache? Thank you!
Could you run Linux as an os in just the lvl 3 cache maybe as a hand down to the consumer market high l3 with consumer counts and on die graphics to truly have a system on a chip
If your data fits into the cache and your jobs require you to make lot of to and fro from your data then this makes a difference otherwise no. Eg: A database which makes use of cache for storing index. This will become fast.
Maybe not related to this topic, but can Microsoft Directstorage accelerate databases? Because if lower API cost for multiple file reads, it may not only be useful for games.
So if these big datacenters are going to drop their OG Milan... how's that secondary market looking?
Sorry, your Epyc has been fucked by AMD™ PSB®
ps: pls buy our new shit so we can afford giving our wonderful CEO Mrs Su her 6-figure paycheck and give more homelabbers and enthusiasts like Buildzoid the middle finger
Hopefully cheap and in great quantity! :D
Though the vendor lock "feature" might cause some havok
Not that impressive. Good luck getting motherboards lmao
imagine how big your clip studio brushes can get!
768mb is 12 times the RAM of my first PC on 1999
What does "BANANAS" means? Frequently hearing it in Level1 videos
Totally and completely bonkers.
nucking futs
n>>>>>>> f
Crazy / Amazing / Incredible / Hopefully that helps define what "bananas" in this context means.
"absolutely mind bogging", "totally far out", "stupendously wild", "out of the stratosphere", etc.
However, if a person "went bananas" it can mean "became violent", "started yelling angrily" or... "became so overwhelmed with joy they started acting irrationally". And only context can give you a clue to which one it is.
So the days were "bananas" was just plural of a bland-ish fruit with a short shelf life are certainly over. But at least we have more L3 cache now.
keeping momory close to the work is the goal
This is why there are no other 3D Desktop CPUs from AMD. ;)
Sounds like it has a drone impersonating side job. But hey, if it gets the job done before schedule, why not?
How much ram can you have as a max, is there a limit?
hehe, yesssssssss
Wow Daytona systems are old. Bet these would crush even more on a retail system.
Wtf has Intel been doing for the past 10 years xD
Trying to eat their cake and have it.... Bob Swan was a bean counter who forgot that ultimately a silicon manufacturer lives and dies with their ability to manufacture.... silicon...
the ability to say 0,76 gb of cache oh no 1.5 gb awesome
Thanks for the informative video. I wonder how this will affect desktop with the R7 5800X3D vs R7 5800X and Zen 4 probably having 3D V-Cache on its higher end SKUs. Intel went with big.LITTLE and is doubling E cores between Alder Lake and Raptor Lake. I wonder how the two different design decisions from Intel and AMD will compare performance wise.
I think Zen 4 will have the single threaded performance crown by a sizeable margin. Whether or not the take the perf/watt crown in multi threaded is another matter..
can it run crysis?
It's fitting that servers sound like very busy bees.
Server singing the song of its people.
What is the reasoning behind L2 cache? I get L1 is per core and L3 is shared, but why is there an L2? (More pedantic this would be why are there 3 levels, with L2 then being the shared cache?)
How does this translate to gaming; will the additional 64MB of L3 be sufficiently utilised to noticeably increase performance (frame rates and timing)...?!??
...or should I stick to my 5800x and not waste money on the 5800x3D ?!?? ...and such; is it more worthwhile to go for the 5950x for long term viability of the AM4 socket ?!??
I think you can feasibly install Arch on just the cache of these things
How
It's too big
768mb? So Milan X has a 8800gtx on it? 😁
I can't tell you that 768MB of L3 is good for, but I can tell you that 512MB of RAM was good for windows xp running gta 3 and firefox in the background.
That's where the magic lies😂
In future we can run GTA V in L3
i loved my 42MB hard drive
Way over my head, but you presented the information so well that I was able to get the implications and time/power/$$$ savings for companies who drop 1 million plus on their "desktops".
Scream like you mean it
@Wendel, you should rename the channel to Level3Techs, the time has come!
I'm mad at 12:58 he said gazillion lol
3:10, 5:17, 7:05 Superlineality.
what if we 10x the L3 cache? what would happen?
please lets get to the end of this journey instead of being mirco transacted into it with each update product and service. new product: maybe AMD EPIC THREADRIPPER?
But how many FPS in CS:GO?
You know, at 7900$, i hope there's a way to push a 7763 over 280w TDP knowing how much a 8-16 cores could pull...for 500-900$
...but can it run Crysis?
Also, need a Collab with Patrick from STH
AMD 2223: "CPU now with integrated banana"
When your cache is multiple times bigger than your entire OS...
Im here from 2050, we have 1 TB of L3 Cache