Fujitsu’s Fugaku and A64FX Take Arm to the Top with 415 PetaFLOPs
anandtech.com
anandtech.com
- AWS Graviton 2 is actually a viable option in the public cloud.
- Apple starts transitioning their workhorses to ARM.
- First place in the TOP500, apparently not due to custom magic or GPUs, but due to already existing/standardized ARM vector extensions.
ARM is becoming such a household name that Fujitsu appears to have traded getting a paid ARM license over the royalty free SPARC instruction set they used to work with. I hope POWER can survive this onslaught, as SPARC surely has no hope of survival after this.
[1] no x64 support, limited OpenGL/3D support, no Fax/Scan support... AFAIK.
It doesn't help that ARM themselves are not confidently offering any desktop worthy intellectual property, as the smartphone chipmakers are just using ARM's own core designs.
See, e.g., performance/watt of Altra vs Epyc: https://www.nextplatform.com/2020/03/18/stacking-up-arm-serv...
Apple is, though.
> performance/watt of Altra vs Epyc
Server CPUs are a different matter altogether, IMO.
In this era where all chips are getting power-state and speed-state shifts at millisecond speeds, eeking out even a 40% performance/watt benefit over comparable total-performance x86 SKUs in the laptop, desktop, and workstation categories would be a massive win for Apple and enough justification on top of unbridling themselves from Intel to warrant the shift.
40% benefit won't bring your average 60W desktop SKU down to a passively-coolable target. Or the 45W SKUs of a MacBook Pro class system (the passively cooled MacBooks were using 7W SKUs, and Apple axed that experiment after four years because even casual users found them too anemic to be marketable at Apple prices).
Even Raspberry Pi's are getting fans these days, and that CPU never draws more than 6W: https://www.jeffgeerling.com/blog/2019/raspberry-pi-4-needs-...
Lots of the ARM power wins in mobile rely on the light and infrequent use of the device. That's why Intel didn't bother with big.LITTLE style architectures until this year.
This is a good jumping off point if you want to dive more into ARM perf/watt, or see how performance/core is just starting to approach that of high end desktop x86 SKUs (with the A14 likely to approach parity): https://www.anandtech.com/show/14892/the-apple-iphone-11-pro...
That's all telling of apple likely dropping 4-16 core mobile and desktop SKUs that pull about 5W/core and are 30-40% more efficient than their x86 competition.
Yeah, a standardized ISA extension, but I'd bet Fujitsu spent a ton of time on the implementation details. They have a deep history of making high performance vector processors going back to the 80s, and high performance SPARC cores parallel to that. They may not be huge volume, but they know how to make a fast processor.
If it's not great marketing then it's a heck of a coincidence.
> The new Fugaku supercomputer is bigger than Summit in practically every way. It has 3.05x cores, it has 2.8x the score in the official LINPACK tests, and consumes 2.8x the power
So were the Nvidia accelerators what drove up Summit's efficiency?
Yes. And the next generation NVIDIA accelerators (A100) in the Selene computer (#7 on the list) are even more efficient, currently taking the #2 spot on the green500 list (which is essentially the top500 list ordered by performance/watt).
Edit: Disclaimer: I work for NVIDIA.
The year 2020 is the conclusion of a long push by ARM and now I am wondering can RISC-V will be taking a similar path.
https://www.hpcwire.com/2019/11/12/cray-fujitsu-both-bringin...
>The new HPE-Cray system, part of the Cray CS500 lineup, will employ the Fujitsu A64FX Arm-based processor with Arm Scalable Vector Extensions (SVE) and second-generation high-bandwidth memory (HBM). Named as customers in today’s release are Los Alamos National Laboratory, Oak Ridge National Laboratory, RIKEN Center for Computational Science, Stony Brook University, and University of Bristol. Cray and Fujitsu said they will be exploring engineering collaboration, co-development, and joint go-to-market strategies to meet customer demand as supercomputing extends into the exascale era.