OIST Deploys AMD EPYC Processors with Over 2 PFLOPs of Computing Power Dedicated to Scientific Research

[btarunr]

Today, AMD and Okinawa Institute of Science and Technology Graduate University (OIST), announced the deployment of AMD EPYC 7702 processors for use in a new, high performance computing system. The EPYC processor-based supercomputer will deliver the 2.36 petaflops of computing power OIST plans to use for scientific research at the University. The Scientific Computing & Data Analysis Section (SCDA) of OIST plans to implement the new supercomputer for supporting OIST computationally intensive research ranging from bioinformatics, computational neuroscience, and physics. SCDA adopted AMD EPYC after significant growth, including a 2X increase in users.

"2020 is a milestone year for OIST with new research units expanding the number of research areas. This growth is driving a significant increase in our computational needs," said Eddy Taillefer, Ph.D., Section Leader, Scientific Computing & Data Analysis Section. "Under the common resource model for which the computing system is shared by all OIST users we needed a significant increase in core-count capacity to both absorb these demands and cope with the significant growth of OIST. The latest AMD EPYC processor was the only technology that could match this core-count need in a cost-performance effective way."



Key factors of OIST's selection of the AMD EPYC processors included superior cost-performance, memory/PCIe bandwidth, and high core counts per server. OIST plans to also consider EPYC processors for other growing computational needs for University researchers in the future.

"AMD is proud to be working with leading global institutions to bring scientific research to the forefront through the power of high performance computing technology," said Ram Peddibhotla, corporate vice president, EPYC product management, AMD. "With high performance capabilities, ease of management and scalability, 2nd Gen AMD EPYC processors can assist OIST researchers with advancing technological innovations and supporting their research goals in bioinformatics, computational neuroscience, and physics."

Learn more about the AMD EPYC processor here.

View at TechPowerUp Main Site

 

[dragontamer5788]

Assuming 1.5 TFlops per 7702 EPYC (some number I randomly searched), that's well over 1000 CPUs, probably in a dual-socket x 500 node cluster.

 

[Vya Domus]

Assuming 1.5 TFlops per 7702 EPYC (some number I randomly searched)

Zen 2 ca do 32 single precision FLOPS/cycle, so 32*64*2 = 4 TFLOPS or 8 TFLOPS for FMA.

 

[AnarchoPrimitiv]

Epyc seems to be getting a lot of supercomputer/large installation wins lately... (El Capitan, Frontier, Indiana University, Purdue University and CERN as well as high-performance computing (HPC) cloud instances from Amazon Web Services, Google, and Oracle Cloud)... To bad private/corporate interests have so much inertia and just want to "stay with what they know".... In the show The Wire, Stringer Bell comments on the illegal drug business, "When the product is good we sell a lot, but when it's bad, we sell even more", and that's akin to Intel Xeon right now..... To make up for the lack of performance Xeon has, these companies just buy even more Xeon chips because they don't want to bother qualifying on Epyc even though it's superior in price and performance.... It's crazy how markets behave counter to what you'd expect

 

[prtskg]

Epyc seems to be getting a lot of supercomputer/large installation wins lately... (El Capitan, Frontier, Indiana University, Purdue University and CERN as well as high-performance computing (HPC) cloud instances from Amazon Web Services, Google, and Oracle Cloud)... To bad private/corporate interests have so much inertia and just want to "stay with what they know".... In the show The Wire, Stringer Bell comments on the illegal drug business, "When the product is good we sell a lot, but when it's bad, we sell even more", and that's akin to Intel Xeon right now..... To make up for the lack of performance Xeon has, these companies just buy even more Xeon chips because they don't want to bother qualifying on Epyc even though it's superior in price and performance.... It's crazy how markets behave counter to what you'd expect

AMD has actually also mentioned that they are focusing on bigger customers first as that brings more money in short timeas well as the reason you mentioned, smaller companies have more inertia along with some other reasons.

 

[dragontamer5788]

Epyc seems to be getting a lot of supercomputer/large installation wins lately... (El Capitan, Frontier, Indiana University, Purdue University and CERN as well as high-performance computing (HPC) cloud instances from Amazon Web Services, Google, and Oracle Cloud)... To bad private/corporate interests have so much inertia and just want to "stay with what they know".... In the show The Wire, Stringer Bell comments on the illegal drug business, "When the product is good we sell a lot, but when it's bad, we sell even more", and that's akin to Intel Xeon right now..... To make up for the lack of performance Xeon has, these companies just buy even more Xeon chips because they don't want to bother qualifying on Epyc even though it's superior in price and performance.... It's crazy how markets behave counter to what you'd expect

There's numerous soft advantages to Intel, even if they're falling behind EPYC in raw cache and/or core counts. Performance discussions have been discussed to death: a brief summary would be AVX512, unified L3 cache (vs EPYC's "split L3 cache"), lower latencies. AMD's system is overall better, but there's enough differences that your code may have to be re-tuned to reach optimal performance on EPYC.

The soft advantages, which are particularly important to HPC, would be Intel's far superior tooling. Including hardware performance counters, Intel MKL (Math Kernel Library), ICC, VTune, and the like. AMD doesn't offer any real competition to Intel's software suite of tools, which is hugely important for optimization. AMD does offer uProf, GCC / CLang is pretty good... but they are definitely steps behind Intel's set of tools.

Even if AMD's CPUs are faster, optimizing code on AMD's CPUs will be a slightly harder job than using Intel's VTune. Especially if your developers are already familiar with VTune, why make them switch to AMD uProf?

 

[mahirzukic2]

There's numerous soft advantages to Intel, even if they're falling behind EPYC in raw cache and/or core counts. Performance discussions have been discussed to death: a brief summary would be AVX512, unified L3 cache (vs EPYC's "split L3 cache"), lower latencies. AMD's system is overall better, but there's enough differences that your code may have to be re-tuned to reach optimal performance on EPYC.

The soft advantages, which are particularly important to HPC, would be Intel's far superior tooling. Including hardware performance counters, Intel MKL (Math Kernel Library), ICC, VTune, and the like. AMD doesn't offer any real competition to Intel's software suite of tools, which is hugely important for optimization. AMD does offer uProf, GCC / CLang is pretty good... but they are definitely steps behind Intel's set of tools.

Even if AMD's CPUs are faster, optimizing code on AMD's CPUs will be a slightly harder job than using Intel's VTune. Especially if your developers are already familiar with VTune, why make them switch to AMD uProf?

To save money, duh. Or get better performance for the same money, duh.
Hell, you said it yourself:

AMD's system is overall better.

 

[dragontamer5788]

To save money, duh. Or get better performance for the same money, duh.
Hell, you said it yourself:

Code optimization isn't that simple. If code was highly tuned to run well on a 2x28-core Skylake Xeon Gold using AVX512, it would take significant modifications to have it run as well on a 2x64-core EPYC 7702 with only AVX256. (Mind you: 2x64 core EPYC would be 8-node NUMA, while 2x28 core Xeon Gold is just 2-node NUMA, maybe 2x2-node with sub-numa clustering).

You'd have to retune the code to "unlock" the performance of the EPYC. For many people, I'm sure it will be easier, and cheaper, to remain on Xeon Gold (even if the overall performance of the system is lower).

EDIT: And by "tuning", that might include rewriting portions of the Intel MKL library (which doesn't perform as well on EPYC servers).

 

[FinneousPJ]

Very true. Never underestimate corporate laziness (or change resistance).

 

[quadibloc]

I suspect there's a mistake in that article. The supercomputer may have two petaflops of floating-point muscle, but the majority of that likely comes not from the EPYC CPUs, as powerful as they may be, but from GPU accelerator cards.