Cray and AMD will build an exascale supercomputer for the Oak Ridge National Laboratory:
AMD today announced that it will partner with Cray to build Frontier, a supercomputer capable of "exascale" performance — one that can complete at least a quintillion floating point computations ("flops") per second, where a flop equals two 15-digit numbers multiplied together — for weather system simulation, subatomic particle modeling, and more. The two companies expect it will be the world's fastest supercomputer when it's delivered in 2021, with more than 1.5 exaflops of theoretical performance — roughly 50 times the speed of today's top supercomputers and faster than the top 160 combined. Frontier will be built at Oak Ridge National Laboratory in Oak Ridge, Tennessee.
[...] Driving Frontier's breakthrough compute is what AMD claims is the first "fully optimized" GPU and CPU design for supercomputing. It features a custom AMD Epyc processor packing a future Zen core architecture designed for high-performance computing (HPC) and AI workloads, along with a graphics processing unit (GPU) in AMD's Radeon Instinct product lineup of server accelerators. The GPUs feature HPC engines, "extensive" mixed precision operations, and high-bandwidth memory, and they're linked together — one Epyc processor to four Instinct graphics cards — by AMD's Infinity Fabric and Cray Slingshot high-bandwidth system interconnect architectures.
Also at AnandTech and The Verge.
See also: AMD's Supercomputer Deal Is a 'Landmark Win' for Chip Maker, Analyst Says
Related Stories
Hewlett Packard Enterprise to Acquire Cray for $1.3 Billion
This morning Hewlett Packard Enterprise and Cray are announcing that HPE will be buying out the supercomputer maker for roughly 1.3 billion dollars. Intending to use Cray's knowledge and technology to bolster their own supercomputing and high-performance computing technologies, when the deal closes, HPE will become the world leader for supercomputing technology.
Cray of course needs no introduction. The current leader in the supercomputing field and founder of supercomputing as we know it, Cray has been a part of the supercomputing landscape since the 1970s. Starting at the time with fully custom systems, in more recent years Cray has morphed into an integrator and scale-out specialist, combining processors from the likes of Intel, AMD, and NVIDIA into supercomputers, and applying their own software, I/O, and interconnect technologies.
The timing of the acquisition announcement closely follows other major news from Cray: the company just landed a $600 million US Department of Energy contract to supply the Frontier supercomputer to Oak Ridge National Laboratory in 2021. Frontier is one of two exascale supercomputers Cray is involved in – the other being a subcontractor for the 2021 Aurora system – and in fact Cray is involved in the only two exascale systems ordered by the US Government thus far. So in both a historical and modern context, Cray was and is one of the biggest players in the supercomputing market.
Related: Intel and Cray Will Build Aurora, U.S.'s First Exaflops Supercomputer, for $500 Million
Cray and AMD Will Build a 1.5 Exaflops Supercomputer by 2021
El Capitan Supercomputer Detailed: AMD CPUs & GPUs To Drive 2 Exaflops of Compute
This afternoon the DOE and HPE are announcing the architectural details of the [El Capitan] supercomputer, revealing that AMD will be providing both the CPUs and accelerators (GPUs), as well as revising the performance estimate for the supercomputer. Already expected to be the fastest of the US's exascale systems, El Capitan was originally commissioned as a 1.5 exaflop system seven months ago. However thanks to some late configuration changes, the DOE now expects the system to reach 2 exaflops once it's fully installed, which would cement its place at the top of the US's supercomputer inventory.
Overall, El Capitan is the second (and apparently final) system being built as part of the US DOE's CORAL-2 program for supercomputers. Like the similar Frontier system, El Capitan comes with a $600 million price tag and is intended to ensure the US's leadership in supercomputers in the exascale era. LLNL will be using the system to replace Sierra, their current IBM Power 9 + NVIDIA Volta supercomputer. All told, El Capitan will be 16 times more powerful than the system it replaces. LLNL will be using it primary for nuclear weapons modeling – substituting for actual weapon testing – while the system will also see secondary use as a research system in other fields, particularly those where machine learning can be applied.
[...] On the CPU side of matters, AMD will be supplying a standard version of their Zen 4-based "Genoa" EPYC processor. As it's still two generations out from AMD's current wares, the amount of information on Zen 4/Genoa is limited, but AMD is promising support for next-generation memory, Infinity Fabric 3, as well as broad promises of both single and multi-threaded performance leadership. Notably, this is a greater level of detail on the CPU than we currently have for Frontier, which is using an unspecified and customized next-generation EPYC CPU.
See also: AMD's CPU-to-GPU Infinity Fabric Detailed
Also at Wccftech.
Previously:
Cray and AMD Will Build a 1.5 Exaflops Supercomputer by 2021
June 2019 TOP500 List: All 500 Systems Above 1 Petaflops
(Score: 3, Informative) by takyon on Tuesday May 07 2019, @08:18PM (5 children)
Summit is currently ranked [top500.org] #1 at 143.5 petaflops Rmax, 200.8 petaflops Rpeak, using 10 MW.
The AMD system will apparently consume around 30 MW. 1.5 exaflops could be the Rpeak. So about 7.5 times the performance for triple the power consumption. 30 MW is probably near the limit of how much power we can tolerate a supercomputer using... maybe 50 MW.
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 2) by Gaaark on Tuesday May 07 2019, @08:26PM
Cray...the final Frontier. These are the voyages of the AMD -- Exaflop.
Or.... something.
Yeah.
--- Please remind me if I haven't been civil to you: I'm channeling MDC. ---Gaaark 2.0 ---
(Score: 1, Touché) by Anonymous Coward on Tuesday May 07 2019, @09:49PM (3 children)
I say 60 MW.
(Score: 3, Interesting) by bob_super on Tuesday May 07 2019, @10:55PM (2 children)
Rather than a 30k people town in Tenessee, they should build those near a million-people town in WI, MN, WA, or northern IL, and use the massive output to provide heating to the town, or for some industrial/chemical application.
There is not reason to limit a computer to 60 or 100 MW, as long as you use the exhaust for something useful. Top-secret bits don't leak.
(Score: 2) by takyon on Tuesday May 07 2019, @11:10PM (1 child)
If you make it too big, you might hold onto it for too long and be stuck with an inefficient and expensive supercomputer. There are some technologies on the horizon that will likely slash power consumption by at least an order of magnitude. We'll eventually see a 1 exaflops system with 1 MW or lower power consumption.
The smaller systems near the top of the Green500 list [top500.org] should be the ideal, as long as you are getting the amount of performance you require. If you want that full 1 exaflops in the early 2020s, fine, but you probably shouldn't shoot for 4 exaflops if the system will end up consuming nearly 100 MW.
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 2, Funny) by Anonymous Coward on Wednesday May 08 2019, @01:45AM
That's why I never start anything. If I start it later, I'll be older and wiser and will finish it quicker.
(Score: 2) by richtopia on Tuesday May 07 2019, @08:36PM (1 child)
Earlier on SN we saw the Aurora Supercomputer which will use Intel discrete GPUs: https://soylentnews.org/article.pl?sid=15/04/10/0216205 [soylentnews.org]
It will be interesting to see a comparison between two systems so closely integrated with a single hardware vendor. Cray is manufacturing both systems.
(Score: 2) by takyon on Tuesday May 07 2019, @08:49PM
That obviously didn't happen. In 2017 they changed the Aurora plans to target exaflops [soylentnews.org].
Xeon Phi is a discontinued product line and Intel will be making actual discrete GPUs starting next year. It's not clear which chips Aurora will use beyond Intel Xeon Scalable and Optane memory. Maybe the Xe-branded discrete GPUs will make an appearance.
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 2) by sshelton76 on Tuesday May 07 2019, @09:07PM (2 children)
The jokes on this write themselves. But good on them. Always nice to see this field continue.
I was shocked to see Cray is still a going concern, I thought they flopped.
/me ducks
(Score: 3, Funny) by ikanreed on Tuesday May 07 2019, @09:09PM (1 child)
The weird thing is you never see Clark Computer and Super Computer photographed together.
(Score: 0) by Anonymous Coward on Tuesday May 07 2019, @11:28PM
I guess one could say that this computer is expected to be a giant flop?
(Score: 0) by Anonymous Coward on Tuesday May 07 2019, @09:09PM (6 children)
Massively parallel computers are so last century ... simply Piling Higher and Deeper, aka PhD.
Here is an area where theoretical computer science could shine, discover a general method to turn a sequential process problem into a parallel processing problem.
(Score: 2) by sshelton76 on Tuesday May 07 2019, @09:18PM (2 children)
That's a nice thought isn't it? But it would only be possible if P=NP and right now that doesn't appear to be the case.
https://en.wikipedia.org/wiki/P_versus_NP_problem [wikipedia.org]
So that's the first step. Prove that P=NP
(Score: 1, Funny) by Anonymous Coward on Tuesday May 07 2019, @09:39PM (1 child)
It does, for N=1. Please send Fields Medal.
(Score: 0) by Anonymous Coward on Wednesday May 08 2019, @05:32AM
Thats nothing I can do a proof in C for any random P
int P = rand();
int NP;
if(P = NP){
println("Proof Correct!");
}else{
println("Proof Failed!");
}
(Score: 2) by takyon on Tuesday May 07 2019, @09:32PM (2 children)
They're just as important as ever, and they are being used for the new hotness, machine learning.
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 2) by DannyB on Tuesday May 07 2019, @09:41PM (1 child)
The new hotness is AI's creating photorealistic images of human beings.
The lower I set my standards the more accomplishments I have.
(Score: 1, Funny) by Anonymous Coward on Wednesday May 08 2019, @01:47AM
15 quinitillion 8008135 per second!!!!
(Score: 2) by stormwyrm on Wednesday May 08 2019, @02:48AM
High floating point operations per second are pointless if you can't get the data to process quickly enough. A lot of the important supercomputing applications, such as Lattice QCD (simulations of the strong interaction in particle physics), computational fluid dynamics, and so on, are limited far more by memory and I/O bandwidth than by computational power. The I/O interconnects (Infinity Fabric/Slingshot) and high-speed memory are thus the things to really watch for.
Numquam ponenda est pluralitas sine necessitate.