AMD EPYC Processors Picked by Argonne National Laboratory to Prepare for Exascale Future

[btarunr]

AMD announced that the U.S. Department of Energy's (DOE) Argonne National Laboratory (Argonne) has chosen AMD EPYC processors to power a new supercomputer, called Polaris, which will prepare researchers for the forthcoming exascale supercomputer at Argonne called Aurora. Polaris is built by Hewlett Packard Enterprise (HPE), will use 2nd Gen EPYC processors and then upgrade to 3rd Gen AMD EPYC processors, and will allow scientists and developers to test and optimize software codes and applications to tackle a range of AI, engineering, and scientific projects.

"AMD EPYC server processors continue to be the leading choice for modern HPC research, delivering the performance and capabilities needed to help solve the complex problems that pre-exascale and exascale computing will address," said Forrest Norrod, senior vice president and general manager, Datacenter and Embedded Solutions Business Group, AMD. "We are extremely proud to support Argonne National Laboratory and their critical research into areas including low carbon technologies, medical research, astronomy, solar power and more as we draw closer to the exascale era."



Polaris will use the AMD EPYC 7532 and EPYC 7543 processors, and NVIDIA A100 Tensor Core GPUs, to deliver approximately 44 petaflops of peak double precision performance, which is 4x faster than Argonne's current supercomputers.

Initially, Polaris will be used by research teams participating in initiatives such as the DOE's Exascale Computing Project and the ALCF's Early Science Program. User communities within the DOE's Exascale Computing Project will also use Polaris for optimizing engineering tasks for Argonne's forthcoming exascale supercomputer, which includes scaling of combined CPU and GPU-enabled systems and the integration of workflows combining modeling, simulation, AI and other data-intensive components.

Polaris is scheduled to be delivered and installed in August 2021 and will go into use starting early 2022. The broader HPC community will access the system in spring of 2022 to prepare workloads for the next generation of DOE's HPC resources.

View at TechPowerUp Main Site

 

[ZoneDymo]

good but again, why would you possible pick anything else....unless you have no real ambition and get a massive discount from the competition....

 

[DeathtoGnomes]

will allow scientists and developers to test and optimize software codes and applications to tackle a range of AI,

Skynet will appreciate this.

 

[iO]

Wow, Intels Aurora has to be delayed by a lot to justify such an AMD/Nvidia stop gap solution...

 

[demu]

Better prepare for Polaris extensions.
Double the capacity twice a year, and Polaris will reach exascale before Intel Aurora is delivered...

 

[ARF]

good but again, why would you possible pick anything else....unless you have no real ambition and get a massive discount from the competition....

Because the something else is faster and more energy-efficient, maybe?
Look at the list of Top 500 Supercomputers in the world - ARM and IBM in the lead:

June 2021 | TOP500

 

[thesmokingman]

Because the something else is faster and more energy-efficient, maybe?
Look at the list of Top 500 Supercomputers in the world - ARM and IBM in the lead:

View attachment 214728
June 2021 | TOP500

You need to look at more than just the raw tflops to get the big picture. The AMD and even the newer Intel supercomputers are off the shelf parts not gigantor custom parts. And the efficiency of the stupid large systems is ridiculously poor and more a by product of their times. Fugaku uses up to 40MW, roflmao. It serves up nearly 7 times the tflops using 10 times the number of cores at 12 times the power. The AMD systems sip power comparatively. If one scaled the Epycs up... it'd theoretically assuming a 1:1 scale for the sake of this thought experiment to achieve the same raw tflops at just over half the power and a third fewer cores.

 

[dragontamer5788]

good but again, why would you possible pick anything else....unless you have no real ambition and get a massive discount from the competition....

The national lab computers need to not only be the best-in-class supercomputers, but also encourage best-in-class supercomputer development. This means tossing a bone to Intel / IBM / other players so that the NVidia continues to have someone to worry about.

Someone's gotta fund the competition, if the competition is to exist.

Because the something else is faster and more energy-efficient, maybe?
Look at the list of Top 500 Supercomputers in the world - ARM and IBM in the lead:

Japan's Fujitsu A64Fx processor is a very, very special case. It looks great, and I have to give props to Japan / Fujitsu for developing it. But its very much a special case, and I don't expect many A64fx computers to be deployed.

IBM's POWER9 doesn't have much compute power actually. Its just got a HELL of a lot of I/O, which allows IBM POWER9 to communicate very quickly with NVidia GPUs. That #2 Summit system has more GPU in it than CPU, with POWER9 only being used as communications (since its the only one with 300GBps CAPI support back when the computer was built). So POWER9 is the best computer for linking those GPUs together.

NVidia seems to have put more effort into NVSwitch / NVlink recently, so a big I/O heavy CPU like POWER9 is no longer necessary.

So the tl;dr is that NVidia dominates that list, with exception of the incredible Fujitsu A64Fx processor (which is more to do with Fujitsu's technology than ARM IMO). No other ARM CPU (be it Qualcomm, Apple, or Marvell) has anything close to Fujitsu's HBM2 enabled ARM chip.

 

[thesmokingman]

^^Don't forget to mention that Fugaku uses as much power a couple small towns.

 

[dragontamer5788]

^^Don't forget to mention that Fugaku uses as much power a couple small towns.

That may be true, but Fugaku was #1 when it was deployed on the Green500 list: meaning it was the most power-efficient supercomputer in the world (smaller, more efficient computers have been made since then however). Fugaku still holds #1 in not only Linpack (traditional measurement from the 1970s), but in HPCG as well (an alternative measurement on supercomputer strength, because LINPACK really only applies to dense matrix-multiplication problems).

Fugaku is one of Japan's strategic supercomputers. They're allowed to get #1 when they try for it, lol. And by all means: they got #1 on pretty much all counts that matter. Hopefully we in the USA reclaim #1 as Frontier is deployed. Maybe not in HPCG: that Fugaku HPCG benchmark is nuts.

 

[Minus Infinity]

But but Sapphire Rapids!

 

[dragontamer5788]

But but Sapphire Rapids!

The competitor to this supercomputer isn't Sapphire Rapids per se, but instead Intel Xe-HPC

 

[TheGuruStud]

The competitor to this supercomputer isn't Sapphire Rapids per se, but instead Intel Xe-HPC

If you believe intel's lies.