No drivers should not be a problem since this computing system is not going to operate on public roads or streets. It will likely sit indoors in at a fixed physical location and transport to that location will happen as parts that are assembled on site so that piece of the puzzle will likely depend on a shippong company that supplies its own vehicle operators.
ROFL... this is the only thing Nvidia fanboys can come with... drivers... I have zero issues with my 5700XT... same with Steve who tested dozens of them... ZERO ISSUES.
Funny, but maybe this quote will alleviate the concerns related to the state of the software environment: “As part of this procurement, the Department of Energy has provided additional funds beyond the purchase of the machine to fund non-recurring engineering efforts and one major piece of that is to work closely with AMD on enhancing the programming environment for their new CPU-GPU architecture.”
So true, what gets me is that AMD found a customer who will pay for their driver and software improvements on top of the equipment they are buying. We consumers will benefit greatly over the next 3-4 years. Maybe AMD will develop programing tools to accelerate x86 workloads for the Zen Arch. Along with gpu drivers or MCM for Instinct gpus before Nvidia! And I thought 2020 was going to be a very good year with Zen 3 and RDNA2. Man look at what we have coming in 2023! Intel can't even come close to Zen 3! What are they going to do by 2023? I bet they ask congress for a government bailout! OMG! Too Big to Fail all over again! What do you folks think? Should congress give a bailout to Intel in the near future? What about all those jobs world wide if Intel went bankrupt? I remember when they said the Banks where to big to fail. Look at how much that has cost US.
This is 2 supercomputer contracts, a few thousand parts. Remind me again how many CPUs Intel, or indeed AMD, shift in a year? You may also note that intel also produce network chips, FPGAs, memory, Mobile chips, graphics chips etc and also own their own FABs. AMD is the least of their worries, The ARM collective, on the other hand, is the real concern.
I think they concern is that they are fighting on several fronts - there‘s ARM with QC being one of the large players, nVidia, AMD and to a degree TSMC. Some of those are rather large, technically advanced in their segment and formidable competitors. This means that established strategies for dealing with competition may no longer work as they did years ago.
It's more about validating - the US is (in theory) working on a next-generation of warheads, but the biggest issue is really in ensuring that the stockpile, such as it is now, still works as it ages since we no longer test IRL.
So to make sure a nuke made ~30 years ago will still go critical if it's asked to, scientists take a lot of measurements of actual nukes, do non-nuclear explosive testing of some warheads, and run a ton of (very complicated) simulations. That's what this computer is for.
I have this suspicious for a long time that AMD will have two distinct GPU uArch, one for Gaming aka Navi and one for GPGPU which will be improved or derived from Vega. And GPGPU / ROCm wont be available on Navi.
Basically, Nvidia is dead in the water without its own CPUs to connect to. IBM fell behind when they sold their fabs and relied on GlobalFoundries to deliver 7nm (what a joke), leaving IBM without any future CPU for the time being. Even if they do deliver, its still POWER, not X86 or ARM.
Even Intel, which has no proof of being able to deliver two next generation products, managed to secure an amendment to the Aurora contract (for next next gen products to fill in), because they have their own x86 CPU to connect to.
Nvidia needs to get moving fast, or they will see their HPC ambitions die.
You're overstating it a little, but your point is valid. At the scale of hundreds of millions of dollars, everything will be customized and being able to provide both customized cpu and gpu outweighs the benefits of controlling the dominant parallel compute platform, ie cuda.
The CUDA thing doesn't mean much to these national lab supercomputers. Neither does providing both a CPU and a GPU. What matters is the price/performance the bidders are willing to offer.
However, providing a CPU & CPU that can more easily access shared memory space may be something they cannot ignore.
If IF3.0 means relatively seamless movement of data between main system DRAM and GPU and CPU, then AMD will be in a very strong position in the HPC market till Nvidia or Intel figure a counter.
As far as I know, there is no software development or system building issue with IBM's OpenCAPI or with the NVLINK connection between IBM's processors and NVIDIA's GPUs. From what I can tell, it's just uninformed, wild speculation that the processors coming from the same vendor mean anything. All of a sudden companies cannot cooperate on protocols? If anything, the IBM and NVIDIA implementation is far more mature.
Again, the reason for the decisions in the procurement, and this has been stated multiple times in interviews by those making the decisions, is the price/performance they were promised by the various bidders.
I don't agree with the author's analysis. I think it's just a coincidence. Aurora was made because it was the result of a prior contract. Cray won the Frontier and El Capitan contracts, and they chose CPUs and GPUs from AMD, because of cost, not because of the CPUs and GPUs coming from a single vendor.
And big public supercomputers and a small part of HPC. The national labs and universities rely mostly on compiler directives to program accelerators, both because of the nature of their code bases and to avoid vendor lock in. However, commercial ventures, such as oil and gas companies and hospitals, are more apt to rely on native codes, such as CUDA. AMD's ecosystem is way off. In addition, enterprise data centers and hyperscalers similarly rely on CUDA code.
AMD's wins are high profile and they will result in revenue for further incentive to continue with GPGPU development, but both revenue and profit wise they are not that singificant. That means NVIDIA is not dead in the war and there is no reason to believe their HPC, or their larger data center business, are in any trouble.
And, of course, there is the mellanox acquisition and the fact they are free to develop ARM, OpenPower or RISC VI CPU stacks to glue their GPGPUs together into a full stack solution.
Nice article...;) I don't know what to make of the "Gosh, Ma, It doesn't come with drivers," comments...so I won't actually try to make anything of that...;) Yes, and you can't buy it off the shelf at Best Buy, either, either. Good for AMD. Likely LLNL has been able to use some early engineering samples of Zen4--which AMD said was already design-complete several months back--and some other sampled hardware. AMD is firmly on a solid execution path--Intel will have it's work cut out for it moving ahead. Frontier is due next year and El Capitan two years later, so I would imagine there would be differences in hardware--we shall see. Interesting stuff!
I don't expect Zen4 has any ES yet, I believe it currently exists only on paper and in simulations. But Zen3 should be available to test and AMD has a pretty good idea what to expect from Zen4.
Isn't it obvious? This will increase Intel's Gaming cred.
Obviously they didn't choose the superior clock speeds of Intel because these computers aren't meant for gaming. It will drive up the notion that Intel is Gaming King, and that AMD can only handle computers that need a billion cores. 8 Cores at 5.2GHZ is all we need for gaming. So this is the perfect thing for Intel!
After the scandalous waste and inefficiency revealed by DOE's performance in the Yucca Mountain nuclear repository fiasco, I've lost all faith in the organization. My impression now is that they spend billions producing nothing. This supercomputer will be more of the same.
Arbie, I have a hard time with your comment. Yes, our government has its inefficiencies, no one can argue that, but no country has better technology or better military because of innovations and risks taken by the US. (with the possible exception of Israel). Your statement basically is a back-hand to all the guys/gals who work for this project trying to do the right thing and are using their expertise.
I'm so glad I have the knowledge to post a rebuttal to your wild accusation.
The Department of Energy is split into two (2) entities at the top of the organization: the National Nuclear Security Administration and the Office of Science. Most of the national labs in the US fall underneath the Office of Science except for 3: Los Alamos (LANL), Lawrence Livermore (LLNL), and Sandia (SNL). Yucca Mountain falls underneath NNSA, not the Office of Science where most of the super computers are, and even then the supercomputers at NNSA labs do a lot more work than just simply Yucca Mountain.
As for the super computers themselves, the fastest computer in the world resides at Oak Ridge National Laboratory and is constantly busy. Scientists from across the world have to bid for time on the computer to run their simulations. Super computers like Frontier, Aurora, and El Capitan will also be just as successful if not more.
One of the applications being developed for Frontier, E3SM-MMF, will allow scientists to model climate changes up to 1 km in the atmosphere for all of North America for the next 100 years. How about the origin of chemical elements? ExaStar models stellar explosions to look at the source of gravity waves and what matter looks like at extreme densities.
The interesting part is that these have unified memory across the CPU and GPU! It will mean that apus with this tecnology aka Zen4 will be really interesting! Chiplet GPUs and GPUs in same socket most likely. That is something that amd has been trying out to do for Many years.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
53 Comments
Back to Article
DominionSeraph - Wednesday, March 4, 2020 - link
Half a billion dollars for a supercomputer with no drivers...extide - Wednesday, March 4, 2020 - link
Oh my sweet summer child...PeachNCream - Wednesday, March 4, 2020 - link
No drivers should not be a problem since this computing system is not going to operate on public roads or streets. It will likely sit indoors in at a fixed physical location and transport to that location will happen as parts that are assembled on site so that piece of the puzzle will likely depend on a shippong company that supplies its own vehicle operators.Irata - Thursday, March 5, 2020 - link
Hehe...great replyeva02langley - Wednesday, March 4, 2020 - link
ROFL... this is the only thing Nvidia fanboys can come with... drivers... I have zero issues with my 5700XT... same with Steve who tested dozens of them... ZERO ISSUES.https://www.youtube.com/watch?v=1uynVO4ZXl0
eva02langley - Wednesday, March 4, 2020 - link
By the way, these are not gaming GPUs on windows, they are used for creating massive simulations and rendering in server environment.69369369 - Wednesday, March 4, 2020 - link
Gamer(TM) detected.TeXWiller - Wednesday, March 4, 2020 - link
Funny, but maybe this quote will alleviate the concerns related to the state of the software environment: “As part of this procurement, the Department of Energy has provided additional funds beyond the purchase of the machine to fund non-recurring engineering efforts and one major piece of that is to work closely with AMD on enhancing the programming environment for their new CPU-GPU architecture.”It will run nuclear crisis, I tell you!
jerry_watson14 - Wednesday, March 4, 2020 - link
So true, what gets me is that AMD found a customer who will pay for their driver and software improvements on top of the equipment they are buying. We consumers will benefit greatly over the next 3-4 years. Maybe AMD will develop programing tools to accelerate x86 workloads for the Zen Arch. Along with gpu drivers or MCM for Instinct gpus before Nvidia! And I thought 2020 was going to be a very good year with Zen 3 and RDNA2. Man look at what we have coming in 2023! Intel can't even come close to Zen 3! What are they going to do by 2023? I bet they ask congress for a government bailout! OMG! Too Big to Fail all over again! What do you folks think? Should congress give a bailout to Intel in the near future? What about all those jobs world wide if Intel went bankrupt? I remember when they said the Banks where to big to fail. Look at how much that has cost US.Makaveli - Wednesday, March 4, 2020 - link
Government Bail out for intel by 2023 ?Did you look at how much money they have lol.
Whiteknight2020 - Thursday, March 5, 2020 - link
This is 2 supercomputer contracts, a few thousand parts. Remind me again how many CPUs Intel, or indeed AMD, shift in a year? You may also note that intel also produce network chips, FPGAs, memory, Mobile chips, graphics chips etc and also own their own FABs. AMD is the least of their worries, The ARM collective, on the other hand, is the real concern.Irata - Thursday, March 5, 2020 - link
I think they concern is that they are fighting on several fronts - there‘s ARM with QC being one of the large players, nVidia, AMD and to a degree TSMC. Some of those are rather large, technically advanced in their segment and formidable competitors. This means that established strategies for dealing with competition may no longer work as they did years ago.Makste - Sunday, December 20, 2020 - link
This reminds so much about the history of the Roman EmpireAdhesiveTeflon - Thursday, March 5, 2020 - link
If you base the largest semiconductor company in the world on just desktops, you're going to have a bad argument.dan_ger - Wednesday, March 4, 2020 - link
Drivers? Intel doesn't even a discreet GPU and they get Aurora. Drivers? LOL!ianmills - Wednesday, March 4, 2020 - link
Because of security issues I betHStewart - Thursday, March 5, 2020 - link
Worst things is AMD has the history sending technology to China.Qasar - Thursday, March 5, 2020 - link
oh and intel is any better with all its security issues ??Vitor - Wednesday, March 4, 2020 - link
Damn, AMD is on a roll,quite beautiful to witness.eva02langley - Wednesday, March 4, 2020 - link
Impressive stuffToTTenTranz - Wednesday, March 4, 2020 - link
So.. GPU is Arcturus / MI100 ?extide - Wednesday, March 4, 2020 - link
Probably not -- given it's time frame I would assume it is a successor of Arcturus.mrvco - Wednesday, March 4, 2020 - link
Hard to read that and not keep thinking that they're running OSX 10.11.... somehow.p1esk - Wednesday, March 4, 2020 - link
Nuclear weapons simulations? What are they trying to develop?sing_electric - Wednesday, March 4, 2020 - link
It's more about validating - the US is (in theory) working on a next-generation of warheads, but the biggest issue is really in ensuring that the stockpile, such as it is now, still works as it ages since we no longer test IRL.So to make sure a nuke made ~30 years ago will still go critical if it's asked to, scientists take a lot of measurements of actual nukes, do non-nuclear explosive testing of some warheads, and run a ton of (very complicated) simulations. That's what this computer is for.
T1beriu - Wednesday, March 4, 2020 - link
>For the first time AMD is naming their Infinity Fabric 3.0It's actually called "3rd Gen Infinity Fabric".
ksec - Wednesday, March 4, 2020 - link
I have this suspicious for a long time that AMD will have two distinct GPU uArch, one for Gaming aka Navi and one for GPGPU which will be improved or derived from Vega. And GPGPU / ROCm wont be available on Navi.jeremyshaw - Wednesday, March 4, 2020 - link
Basically, Nvidia is dead in the water without its own CPUs to connect to. IBM fell behind when they sold their fabs and relied on GlobalFoundries to deliver 7nm (what a joke), leaving IBM without any future CPU for the time being. Even if they do deliver, its still POWER, not X86 or ARM.Even Intel, which has no proof of being able to deliver two next generation products, managed to secure an amendment to the Aurora contract (for next next gen products to fill in), because they have their own x86 CPU to connect to.
Nvidia needs to get moving fast, or they will see their HPC ambitions die.
quorm - Wednesday, March 4, 2020 - link
You're overstating it a little, but your point is valid. At the scale of hundreds of millions of dollars, everything will be customized and being able to provide both customized cpu and gpu outweighs the benefits of controlling the dominant parallel compute platform, ie cuda.Yojimbo - Wednesday, March 4, 2020 - link
The CUDA thing doesn't mean much to these national lab supercomputers. Neither does providing both a CPU and a GPU. What matters is the price/performance the bidders are willing to offer.Atari2600 - Thursday, March 5, 2020 - link
However, providing a CPU & CPU that can more easily access shared memory space may be something they cannot ignore.If IF3.0 means relatively seamless movement of data between main system DRAM and GPU and CPU, then AMD will be in a very strong position in the HPC market till Nvidia or Intel figure a counter.
Yojimbo - Thursday, March 5, 2020 - link
As far as I know, there is no software development or system building issue with IBM's OpenCAPI or with the NVLINK connection between IBM's processors and NVIDIA's GPUs. From what I can tell, it's just uninformed, wild speculation that the processors coming from the same vendor mean anything. All of a sudden companies cannot cooperate on protocols? If anything, the IBM and NVIDIA implementation is far more mature.Again, the reason for the decisions in the procurement, and this has been stated multiple times in interviews by those making the decisions, is the price/performance they were promised by the various bidders.
Yojimbo - Wednesday, March 4, 2020 - link
I don't agree with the author's analysis. I think it's just a coincidence. Aurora was made because it was the result of a prior contract. Cray won the Frontier and El Capitan contracts, and they chose CPUs and GPUs from AMD, because of cost, not because of the CPUs and GPUs coming from a single vendor.And big public supercomputers and a small part of HPC. The national labs and universities rely mostly on compiler directives to program accelerators, both because of the nature of their code bases and to avoid vendor lock in. However, commercial ventures, such as oil and gas companies and hospitals, are more apt to rely on native codes, such as CUDA. AMD's ecosystem is way off. In addition, enterprise data centers and hyperscalers similarly rely on CUDA code.
AMD's wins are high profile and they will result in revenue for further incentive to continue with GPGPU development, but both revenue and profit wise they are not that singificant. That means NVIDIA is not dead in the war and there is no reason to believe their HPC, or their larger data center business, are in any trouble.
Whiteknight2020 - Thursday, March 5, 2020 - link
And, of course, there is the mellanox acquisition and the fact they are free to develop ARM, OpenPower or RISC VI CPU stacks to glue their GPGPUs together into a full stack solution.WaltC - Wednesday, March 4, 2020 - link
Nice article...;) I don't know what to make of the "Gosh, Ma, It doesn't come with drivers," comments...so I won't actually try to make anything of that...;) Yes, and you can't buy it off the shelf at Best Buy, either, either. Good for AMD. Likely LLNL has been able to use some early engineering samples of Zen4--which AMD said was already design-complete several months back--and some other sampled hardware. AMD is firmly on a solid execution path--Intel will have it's work cut out for it moving ahead. Frontier is due next year and El Capitan two years later, so I would imagine there would be differences in hardware--we shall see. Interesting stuff!redtail3 - Wednesday, March 4, 2020 - link
They said Z3 is design complete. Haven't said anything about Z4 yetFreckledTrout - Wednesday, March 4, 2020 - link
Nope they haven't but then Zen4 is a new design, new sockets, likely DDR5 etc. I wouldn't let out exactly what I was doing either.Zizy - Thursday, March 5, 2020 - link
I don't expect Zen4 has any ES yet, I believe it currently exists only on paper and in simulations. But Zen3 should be available to test and AMD has a pretty good idea what to expect from Zen4.WaltC - Thursday, March 5, 2020 - link
Thanks for the correction guys!--memory is a tricky thing--I thought I had read that awhile back--guess it was Zen 3...Lord of the Bored - Thursday, March 5, 2020 - link
Wait, where are the paid shills in the comments? I really wanted to see someone explaining how this was a boon for Intel.Xyler94 - Thursday, March 5, 2020 - link
Isn't it obvious? This will increase Intel's Gaming cred.Obviously they didn't choose the superior clock speeds of Intel because these computers aren't meant for gaming. It will drive up the notion that Intel is Gaming King, and that AMD can only handle computers that need a billion cores. 8 Cores at 5.2GHZ is all we need for gaming. So this is the perfect thing for Intel!
Qasar - Thursday, March 5, 2020 - link
intel kept saying we only need 4 cores, and stuck mainstream at 4 cores for how many years ?Xyler94 - Thursday, March 5, 2020 - link
I was being sarcastic of course. But to answer your question, from at least 2011 to 2017.Qasar - Thursday, March 5, 2020 - link
:-)Arbie - Thursday, March 5, 2020 - link
After the scandalous waste and inefficiency revealed by DOE's performance in the Yucca Mountain nuclear repository fiasco, I've lost all faith in the organization. My impression now is that they spend billions producing nothing. This supercomputer will be more of the same.A giant self-licking ice-cream cone.
HSO4 - Thursday, March 5, 2020 - link
Arbie, I have a hard time with your comment. Yes, our government has its inefficiencies, no one can argue that, but no country has better technology or better military because of innovations and risks taken by the US. (with the possible exception of Israel). Your statement basically is a back-hand to all the guys/gals who work for this project trying to do the right thing and are using their expertise.Ninhalem - Thursday, March 5, 2020 - link
I'm so glad I have the knowledge to post a rebuttal to your wild accusation.The Department of Energy is split into two (2) entities at the top of the organization: the National Nuclear Security Administration and the Office of Science. Most of the national labs in the US fall underneath the Office of Science except for 3: Los Alamos (LANL), Lawrence Livermore (LLNL), and Sandia (SNL). Yucca Mountain falls underneath NNSA, not the Office of Science where most of the super computers are, and even then the supercomputers at NNSA labs do a lot more work than just simply Yucca Mountain.
As for the super computers themselves, the fastest computer in the world resides at Oak Ridge National Laboratory and is constantly busy. Scientists from across the world have to bid for time on the computer to run their simulations. Super computers like Frontier, Aurora, and El Capitan will also be just as successful if not more.
One of the applications being developed for Frontier, E3SM-MMF, will allow scientists to model climate changes up to 1 km in the atmosphere for all of North America for the next 100 years. How about the origin of chemical elements? ExaStar models stellar explosions to look at the source of gravity waves and what matter looks like at extreme densities.
Ninhalem - Thursday, March 5, 2020 - link
Forgot to add that you can find all of this information and more at the Exascale Computing Project's home page: exascaleproject.org.AdhesiveTeflon - Thursday, March 5, 2020 - link
But will it run crysis at all high?haukionkannel - Thursday, March 5, 2020 - link
The interesting part is that these have unified memory across the CPU and GPU!It will mean that apus with this tecnology aka Zen4 will be really interesting! Chiplet GPUs and GPUs in same socket most likely. That is something that amd has been trying out to do for Many years.
SanX - Friday, March 6, 2020 - link
Yes, you will get 23547434 fps. Your classmates will be jealous next 14500 years.Lolimaster - Saturday, March 7, 2020 - link
Intel system power consumption not available LOL.MenhirMike - Friday, March 13, 2020 - link
Pfft, El Capitan is so 2015. Catalina is where it's at!