Microsoft convinced AMD and Nvidia to build a CPU with extraordinary features but it will never go on sale: 4th gen 9V64H has 88 cores and uses InfiniBand technology
Posted by Arthur_Morgan44469@reddit | hardware | View on Reddit | 83 comments
LordMohid@reddit
Quad socket??? Wtf would love to see the motherboard housing these 4P systems
titanking4@reddit
While it would be amazing to see, this part is going to use HBM and thus won’t have any DDr5 slots which use more space on the PCB these days than the sockets.
BuchMaister@reddit
It could have also (theoretically) have DDR 5 slots. See for example Xeon max that has both HBM on package and DDR 5 memory controllers.
Celaphais@reddit
I've seen server socs with both HBM and lpddr5
gumol@reddit
https://www.supermicro.com/en/products/motherboard/x10qrh+
however, this CPU doesn't need DDR slots
MDSExpro@reddit
4 sockets is nothing new or atypical in servers world.
b3081a@reddit
Same as MI300A. It's just using all Zen4 12*CCDs rather than 3*CCD + 6*XCDs.
Maldiavolo@reddit
This is such a terrible title. MS didn't convince anyone. It paid AMD to make a custom processor which AMD does as part of its custom division. Nvidia also had nothing to do with it because Infiniband is just networking. MS simply loaded the server with Nvidia/Mellonox Infiniband NICs and is using Mellonox Infiniband TOR (top of rack) switches. You can literally buy and use these in any server you want.
Tumleren@reddit
They convinced them with lots of money
proxgs@reddit
There was no convincing since AMD offers custom CPU for hyperscalers as a b2b product.
AbhishMuk@reddit
Nah, AMD was really reluctant. Until Microsoft said they were going to actually pay them money.
ExpletiveDeletedYou@reddit
Yeah, When I asked AMD to make me a b2b hyperscaler CPU there were initially interested until they learnt I had a budget of $600. Even though I offered to increase it by 50% they simply weren't interested.
DrWitchDoctorPhD@reddit
Christ I think my idea of they doing without any payment solely forthe exposure is starting to sound less likely
ExpletiveDeletedYou@reddit
i really should have offered to go 75% increase in hindsight
TenshiBR@reddit
Slow down man, let it cook
ParthProLegend@reddit
It's cooked already
Yebi@reddit
Yesterday I convinced McDonald's and Uber Eats to make me a sandwich. It's not for sale
dahauns@reddit
Thank you! The AMD/NVidia angle is especially dumb.
This would be like "Cartel Uncovered - AMD and Intel secretly working together?" because a lot of AMD mainboards have Intel NICs...
randomkidlol@reddit
more like microsoft paid AMD for a semicustom part, bought some other nvidia parts, and put together a special machine just for azure.
kakemone@reddit
Hardware engineering at its best. There are no limits when people/companies work together .
Miserable_Fault4973@reddit
You mean when companies have Billions to burn.
Creative_Purpose6138@reddit
Yes, the money grows hands and does it. No credit to engineers.
Miserable_Fault4973@reddit
Not sure what your point is. Obviously engineers do the work, but only a few companies can actually afford to pay that many engineers just for a custom design.
Creative_Purpose6138@reddit
Not sure what your point is. Obviously you need money to do engineering. Why'd you have to correct the OP which seems perfectly fine IMO.
Miserable_Fault4973@reddit
Because people aren't working together like it's some Christmas miracle; they're getting paid to do a job.
Soytaco@reddit
So you're telling me the folks at AMD and nVidia are getting paid? I gotta talk to my boss about this..
quildtide@reddit
I mean, if my company told me to engineer an insane CPU like this with basically no limits on per-unit expenses, I think I would consider that a Christmas miracle.
SaintsPelicans1@reddit
No one ever said anything like that lol
mernold@reddit
His point is that engineers would LOVE to tackle a project like that but the only firms that would do it have billions to gamble on said project being worth the cost
blenderbender44@reddit
It being really expensive just adds to the engineering marvel. There's plenty of things engineers currently cannot build even with infinite money.
Miserable_Fault4973@reddit
Yeah, but this ISN'T a marvel. It's a custom chip just mixing and matching existing tech, not something pushing through technology frontier.
blenderbender44@reddit
fair point. AMD thread rippers are upto what, 96 cores now?
danielv123@reddit
Milan is 196 cores per socket i think?
blenderbender44@reddit
Thread ripper might as well (or at least one in the works), can't be assed researching though
ProbablyPissed@reddit
Probably that your comment was obviously implied in the OP. Everyone knows about money, you aren't opening eyes here.
sixpointnineup@reddit
I hope this is MSFT trying to claw back/preserve market share from GC and AWS.
igby1@reddit
No limits? Faster-than-light travel confirmed!
Adromedae@reddit
LOL @ the wording of the title.
The CPU is most definitively not done for free and it is for sale, just not for consumer channels.
gumol@reddit
You can buy access to it (through Azure), but you cannot buy it outright.
MediocreAd8440@reddit
AMD has done similar stuff before, if you want a large enough quantity of something, they'll move mountains. So WE can't buy it, but some other large hyperscaler could easily request something like this
hhunaid@reddit
Does console chips fall under that? What about other consumer goods like Steam Deck?
PastaPandaSimon@reddit
It's a great example too. As the Steam Deck has gotten the far more polished handheld-friendly chip than the rebranded laptop chips like the Z1 that they are lazily releasing to that market themselves.
trenthowell@reddit
I suspect Valve's expertise working with hardware and their own Linux distro had no small part in that too. Bet it also made for a good experience for the AMD folks working with them on it too
drhappycat@reddit
Since it's not a general purpose cpu per se, you think there's zero chance of ES/QS versions showing up on ebay? Would a regular 4th gen server board recognize it?
munchkinatlaw@reddit
Quad socketing probably required a custom pinout, so probably not.
AbhishMuk@reddit
Honestly that’s nicer than what I thought (exclusively build for MS’s servers as their only client)
noiserr@reddit
You can buy mi300c which is basically the same thing.
gumol@reddit
Where can you buy MI300C? When was it released?
I can only find rumors about MI300C, nothing concrete.
noiserr@reddit
https://www.tomshardware.com/news/amds-special-mi300c-ai-accelerator-emerges
It was announced last year. There are 3 variants of the mi300 platform.
mi300x (pure GPU accelerator)
mi300a (APU, with unified memory)
mi300c (CPU variant)
c version didn't get much attention because it is not for AI. So it's probably why it flew a bit under the radar. I also remember Lisa talking about it on stage briefly at one of the big events (forgot which).
gumol@reddit
Your link is just rumors.
UsernameAvaylable@reddit
Who would think its for free. "Never go on sale" means its proprietary...
AreYouAWiiizard@reddit
I'm kind of surprised anyone would even think that... Personally when I first read the title, it made me assume it never made it to production :/
tecedu@reddit
Unless they have made special NICs for these. not sure why Nvidia is mentioned here?Should be standard connect x7 as far I know
karatekid430@reddit
I have always found Infiniband interesting and thought that it is a good replacement for Ethernet, especially in that you can tunnel IP over it, but also rDMA for performant applications.
danielv123@reddit
The part i don't understand is why? What is the advantage? Rdma is a thing on Ethernet too
tecedu@reddit
RDMA over Ethernet has been a mess for a while until the AI revolution forced advancements, now its Nvme/roce for a most things which makes life easier for the networking people. But for all of those years, Infiniband was the only way to do MPI and RDMA. And Mellanox was just plug and play.
tecedu@reddit
I'd love to use it but unfotunately my azure region is two generations behind so never getting this there.
sascharobi@reddit
Every article written these days is just clickbait.
bob69joe@reddit
This article claims that zen4c and 5c have SMT disabled. Which isn’t true.
Evilbred@reddit
For these chips they probably are.
They run better optimized workloads so there is less need for pipeline scheduling.
bob69joe@reddit
The article said that “this chip doesn’t have SMT like zen 4c-5c”. But normal zen4c-5c does have SMT.
UsernameAvaylable@reddit
Reading comprehension?
The quote says that 4c-5c have SMT and this chip does not? What you claim would be "This chip has SMT unlike zen 4c-5c"?
VenditatioDelendaEst@reddit
Writing comprehension. The real quote is
The article does, in fact, incorrectly claim that Zen 4c and 5c have SMT disabled.
Geddagod@reddit
The only way this might make sense is if Microsoft's Bergamo and Turin Dense also have SMT disabled, don't know how true that is though.
VenditatioDelendaEst@reddit
Hmm...
I'm not finding any Bergamo or Turin Dense options in Azure's catalog, but they do apparently have configurations on the whole spectrum of SMT settings.
bob69joe@reddit
Don’t feel like re-reading the article but i know what i read.
Evilbred@reddit
It's not a problem bud, the phrasing is a bit weird and could be interpreted both ways.
VenditatioDelendaEst@reddit
What you read is not what you wrote, and in this case the distinction is very important. Funny thing is... you were actually right the first time.
System0verlord@reddit
I’m not so sure about that one.
battler624@reddit
Microsoft also confirmed that the chip is SMT-disabled (just like the Zen 4c and Zen 5c parts)
Maybe you (and everyone else that isnt bob) are the ones with reading comprehension
Evilbred@reddit
They meant it like "this chip doesn't have SMT, like Zen 4c and 5c does."
Just worded vaguely.
anival024@reddit
You're parsing that incorrectly.
this chip doesn’t have SMT like zen 4c-5c
a dog doesn't meow like a cat
RandomGenericDude@reddit
Microsoft's versions do apparently.
XyaThir@reddit
« Clearly, single socket is where the market is going, except when there’s a clear case - and a hyperscale client with near unlimited funds - to go for quads. » Could not be more wrong. All high end solutions are with 2/4/8 sockets. SXM or Grace Hopper modules.
hardolaf@reddit
Even my workstation at work is dual socket.
E-werd@reddit
Lets consider a world where we could buy CPUs with HBM integrated. This is already what Apple does with their silicon. We've seen the crazy advantages to on-die memory with Apple's M-series SoCs and with AMD's Radeon Vega/VII line.
It complicates the market a bit and would require some new hardware, but the performance would be incredible.
ScepticMatt@reddit
It actually has 96 cores each but 88 are available to the VM:
https://www.servethehome.com/this-is-the-microsoft-azure-hbv5-and-amd-mi300c-nvidia/
Brisngr368@reddit
I'm more interested as to whether this will actually make Microsoft a viable competitor for large scale HPC in university's for example. Or are they just gonna charge more for this special chip just to suck more money out of people who already use it
bobbie434343@reddit
HUB gaming benchmarks with depressed Steve or it does not exist ! /s
6950@reddit
Intel already has HBM Sapphire Rappids( Xeon Max) it is not some marvel created for the first time it is more about paying someone to be able to make them for the money/engineering resources that is required for it as for Nvidia just lend their networking how they have spined the PR is ridiculous take
ElementII5@reddit
Yes, and they are absolutely terrible. Intel couldn't figure out how to connect the HBM without them failing after a while.
So yeah it is some kind of marvel.
6950@reddit
It was not that they couldn't figure out how to it's just that SPR is the worst product in general due to so many delays and stuff if they didn't figure out how we even got xeon max it was more like SPR in general was really problematic Node + design delays and we got disaster
theophys@reddit
Woowie, I'd love to have me one of those.