News [Toms] Beyond Rome: AMD's EPYC and Radeon to Power World's Fastest Exascale Supercomputer

Hitman928

Diamond Member
Apr 15, 2012
5,182
7,632
136
https://www.tomshardware.com/news/amd-epyc-radeon-frontier-exascale-supercomputer,39275.html

AMD announced today that it had been selected to power Frontier, which is set to be the world's fastest exascale-class supercomputer when it comes online in 2021. . . The new Frontier supercomputer is expected to deliver a leading 1.5 exaflops of performance powered by next-generation variants of AMD's EPYC processors and Radeon Instinct GPUs.

AMD didn't reveal which specific generation of its GPUs and CPUs will power the system, although CEO Lisa Su did announce that both components are customized for the deployment.
AMD's EPYC processor has been optimized with support for new instructions that provide optimal performance in AI and supercomputing workloads, with Su saying, "it is a future version of our Zen architecture. So think of it as beyond..what we put into Rome."
 

exquisitechar

Senior member
Apr 18, 2017
655
862
136
AMD went from nothing to the very top in such a short amount of time. Well done, Lisa Su and co.

It is a particularly pleasant surprise that Frontier is powered by Radeon GPUs. I think my prediction that Intel and AMD will provide serious competition to Nvidia in this sector will come true. The Epyc CPUs will likely be Genoa, I presume?
 

DrMrLordX

Lifer
Apr 27, 2000
21,582
10,785
136
Could be an early run of Genoa chips. I wouldn't expect full availability of Genoa until pretty late in 2021.

I'm more curious about the Instinct GPUs that will be in there. That would be the "future architecture" we've seen on roadmaps? Or will it be a custom GCN hackjob?
 

exquisitechar

Senior member
Apr 18, 2017
655
862
136
Could be an early run of Genoa chips. I wouldn't expect full availability of Genoa until pretty late in 2021.

I'm more curious about the Instinct GPUs that will be in there. That would be the "future architecture" we've seen on roadmaps? Or will it be a custom GCN hackjob?
https://www.nextplatform.com/2019/05/07/cray-amd-tag-team-on-1-5-exaflops-frontier-supercomputer
The CPU is a unique, custom device that is not based on the impending “Rome” second generation Epyc processor and it is not based on the future “Milan” follow-on, either, but is rather a custom CPU.
Yeah, it isn't Genoa based after all, I guess.
 
  • Like
Reactions: prtskg

moinmoin

Diamond Member
Jun 1, 2017
4,934
7,619
136
"All told, the supercomputer is projected to be the most expensive machine ever built and be faster than the top 160 supercomputers in the world, combined."

Pure insanity. Congratulations to AMD (semi-custom business?) for wining this contract.
 

amrnuke

Golden Member
Apr 24, 2019
1,181
1,772
136
This is the type of contract that allows AMD to develop even more experience. By building custom CPUs not based on current and future consumer/enterprise CPUs, there could be a bit of a "halo effect" of taking the lessons learned in development and applying them to future publicly available CPUs.

Needless to say, I'd like to hear more about the GPUs and whether they will be custom as well. Specifically, for such a large project I have to imagine there is some impetus to increase power savings, and it'll be interesting to see how that might be applied to AMD's future consumer GPUs.
 

DisEnchantment

Golden Member
Mar 3, 2017
1,590
5,722
136
The US DoE will award 100M USD as development contract, so probably a good chunk of that will go to AMD.
More funds for R&D.
And interesting for Open Source folks is that ROCm will be used which means we can expect lots of features upstreamed.
 
Last edited:
  • Like
Reactions: lightmanek

bsp2020

Member
Dec 29, 2015
103
114
116
AMD did this in Opteron times. Good to see it again.

This one is more significant because both CPU and GPU are from AMD. During the Opteron days, AMD did not have GPU.

It seems providing both CPU and GPU is a significant advantage. When Intel Aurora was announced earlier this year, I was really surprised that Intel's GPU was part of the deal. But it seems DOE wants to award contracts to as diverse companies and I'm sure the next award will go to another IBM+NVidia design.
 

Mopetar

Diamond Member
Jan 31, 2011
7,797
5,899
136
The CPU is obvious given the current situation, but the GPU isn’t necessarily a surprise. We know that AMD GPUs have some good compute potential and compete well if you just look at raw FLOPs. Some of their tech like HBCC that didn’t have much practical consumer use might have also factored in to the decision as well.
 

maddie

Diamond Member
Jul 18, 2010
4,723
4,628
136
This one is more significant because both CPU and GPU are from AMD. During the Opteron days, AMD did not have GPU.

It seems providing both CPU and GPU is a significant advantage. When Intel Aurora was announced earlier this year, I was really surprised that Intel's GPU was part of the deal. But it seems DOE wants to award contracts to as diverse companies and I'm sure the next award will go to another IBM+NVidia design.
R&D is certainly spread around thus keeping the home knowledge base intact. Many industries benefit from this.

The custom CPU is quite interesting. What a change from the normally associated role of the custom division (gaming and lower end) and what else might we see in the next few years.

AMD appears to have fully embraced a Lego approach to complex SOC designs. Contrast to Intel integrating designs with a locked monolithic philosophy.
 

maddie

Diamond Member
Jul 18, 2010
4,723
4,628
136
The CPU is obvious given the current situation, but the GPU isn’t necessarily a surprise. We know that AMD GPUs have some good compute potential and compete well if you just look at raw FLOPs. Some of their tech like HBCC that didn’t have much practical consumer use might have also factored in to the decision as well.
HBCC was never really done for gaming. Using the HBM as a cache really only works for huge datasets. as saving a few GB on a card is minimally cost effective. Two paths going forward.
 

Mopetar

Diamond Member
Jan 31, 2011
7,797
5,899
136
HBCC was never really done for gaming. Using the HBM as a cache really only works for huge datasets. as saving a few GB on a card is minimally cost effective. Two paths going forward.

It likely wasn’t, but they still brought it up and showed it off in a gaming scenario. I think a lot of people have written AMD GPUs off because they’re viewed as power hungry with gimmick features that don’t do a lot.

I won’t say this is some kind of master plan on the part of AMD, but rather that people might be sleeping on their GPUs more than they should be. What I’m saying is that it shouldn’t be as much of a surprise as some make it out to be.
 
  • Like
Reactions: OTG

maddie

Diamond Member
Jul 18, 2010
4,723
4,628
136
It likely wasn’t, but they still brought it up and showed it off in a gaming scenario. I think a lot of people have written AMD GPUs off because they’re viewed as power hungry with gimmick features that don’t do a lot.

I won’t say this is some kind of master plan on the part of AMD, but rather that people might be sleeping on their GPUs more than they should be. What I’m saying is that it shouldn’t be as much of a surprise as some make it out to be.
Think Raja. He couldn't stop talking.

Tahiti was both, Vega started to emphasize highend compute, Navi, if more gaming focused, will confirm a dual portfolio going forward.
 

zinfamous

No Lifer
Jul 12, 2006
110,512
29,099
146
This is the type of contract that allows AMD to develop even more experience. By building custom CPUs not based on current and future consumer/enterprise CPUs, there could be a bit of a "halo effect" of taking the lessons learned in development and applying them to future publicly available CPUs.

Hopefully their marketing team is on top of this very soon

"AMD: Building the world's fastest computers, period."
 
  • Like
Reactions: DarthKyrie

jpiniero

Lifer
Oct 1, 2010
14,510
5,159
136
By mandate this purchase could have only been an AMD system.

No, they could have used nVidia. It just couldn't have been Intel.

Edit: Should add that since it's still unknown what node Xe is using, there's still a very good chance Intel won't be able to deliver on Aurora. Using AMD helps out still being able to use the software developed if Intel can't make it work.
 
  • Like
Reactions: prtskg

Phynaz

Lifer
Mar 13, 2006
10,140
819
126
Edit: Should add that since it's still unknown what node Xe is using, there's still a very good chance Intel won't be able to deliver on Aurora. Using AMD helps out still being able to use the software developed if Intel can't make it work.

Lol, sure. There’s a much larger chance AMD won’t be able to deliver. As a matter of fact I say it’s 100x more likely AMD will fail than intel.

There, I made a statement just as valid as yours, based upon the same amount of information.

Anyone can make stuff up.
 

prtskg

Senior member
Oct 26, 2015
261
94
101
Wow! Major win. Good to see AMD win The deal. It will help us consumers too as this means more competition. Good quality product and competitive prices. :D
 

DrMrLordX

Lifer
Apr 27, 2000
21,582
10,785
136
Lol, sure. There’s a much larger chance AMD won’t be able to deliver. As a matter of fact I say it’s 100x more likely AMD will fail than intel.

There, I made a statement just as valid as yours, based upon the same amount of information.

Anyone can make stuff up.

Why would AMD have a larger chance of being unable to deliver? Their track record since 2017 has been nearly spotless, at least on the CPU side. On the GPU side, things are a bigger unknown, though they came through for PS5/Xbox 2(?) so I would not dismiss their ability to get hardware ready for a specific client. If they have neglected anyone, it's the consumer PC market. Now that AMD is using TSMC for nearly everything, they'll be ready on the fab side.

Intel has absolutely no recent record on delivering on the dGPU front. They did have to cancel an entire uarch upgrade (Cannonlake) on the CPU side, along with Gen10 graphics, so that certainly isn't good. Intel could improve their prospects by using an outside fab for Xe. We still haven't seen them deliver a working dGPU product, though.

Does the mandate say anything about x86? Even still, NVIDIA was an option has been stated.

NV was always an option. POWER/NV wound up in Summit, after all.
 

NostaSeronx

Diamond Member
Sep 18, 2011
3,683
1,218
136
Why would AMD have a larger chance of being unable to deliver? ... Now that AMD is using TSMC for nearly everything, they'll be ready on the fab side.
AMD can lose priority at TSMC, which can lead to them being unable to deliver. Which was why AMD went to GlobalFoundries exclusively. No fellow customers means they would be the only customer.

Semi-custom however has extra weight in TSMC's eyes.
 
  • Like
Reactions: Olikan