AMD CDNA Discussion Thread

Corpo can easily afford it.

https://wccftech.com/does-the-amd-instinct-mi300-accelerator-have-an-exascale-apu-mode/

 

"Setonix stage 2 will be delivered mid-2022."
This will include 750 next-gen AMD GPUs (MI200 presumably) and 200k AMD compute cores.
https://pawsey.org.au/pawsey-provides-the-first-look-at-setonix-wrapped-in-stars/

So, in all, a rather small system compared to the exascale projects.

 

https://ir.amd.com/news-events/pres...-ambitious-goal-to-increase-energy-efficiency

Ehehehe, they already are cooking something and they want to swoop in in 2025 like the white Knight

 

https://news.mynavi.jp/article/20210929-1985760/

Might as well post the one with all the actual slideware.

 

The measurement criteria is very convoluted though (expected) but all in all a very bold claim
Basically a product meant for launch in 2025 should have been done by 2024, so basically just two more product dev cycle after MI300 and Genoa

 

They're doing yearly updates on that goal anyway so MI200 and MI300 and whatnot all count.

 

https://www.hpcwire.com/2021/09/29/us-closes-in-on-exascale-frontier-installation-is-underway/
MI200 being deployed

 

is it with and Genoa?

 

Nope, custom 3rd Gen Epyc. Genoa is planned for "El Capitan"

 

Trento.

 

https://forums.anandtech.com/thread...scale-supercomputer-frontier-emerges.2597611/

 

It could be unspecified Epyc + 4x MI60 accelerators used in the presentation as 2020 energy efficiency point

 

More like some Generic shutterstock Architecture which some "tech media" claimed to be Frontier Architecture.

 

The thing on that diagram is very much a Frontier node.
One Trento, four MI200 with 4 Slingshot11 NICs attached to the slave MI200 dies.
You can even find the actual node diagram somewhere.

Because it is the Frontier node.

 

According to A. Schilling:

 

https://ascii.jp/elem/000/004/070/4070987/2/

 

https://www.nextplatform.com/2021/1...tier-exascaler-contrasted-to-argonnes-aurora/

 

https://www.hpcwire.com/2021/10/21/...p-mi200-gpu-readies-for-a-major-global-debut/

 

https://wccftech.com/amd-instinct-m...a-4x-increase-in-fp16-performance-over-mi100/

 

See?
Bazillions of DPFP per 500W all on some good old N7.

Time to kneel o-onegai.