AMD CDNA Discussion Thread

It's 1:1 with support for packed FP32, which delivers 2:1 when the math allows it (apparently most of the time)

 

https://ir.amd.com/news-events/pres...accelerated-data-center-premierevirtual-event

Well they're doing it before SC'21 it seems.

Yes.

 

Glad to see they're completely removing the "GPU" acronym from their Instinct series of products and announcements.

 

Because they've gutted out all the GPU parts besides the very lobotomized VCN instance.
They're genuine cray-cray SIMD machines now.

 

https://ir.amd.com/news-events/pres...accelerated-data-center-premierevirtual-event
CDNA2 will be very likely be announced in couple of weeks time.
Including some EPYC stuffs

 

Yes, but they had already done so with Mi100 and yet they still called that a GPU.
It's better that they don't call it a GPU IMO. It's just proper naming.

 

Yea.
Silly AMD and doing 11.5TF > 48TF DPFP in but one year.

 

It looks like Ethernet also accounts for most of Mellanox's revenue and they're pushing Ethernet into the higher speed tiers. No idea if that's relevant for flagship projects like El Capitan.

No rasterizers, texture units or ROPs in sight. It’s a big fat FP co-processor.

 

Yessssss.
Ethernet truly won.
Infiniband flavours not Mellanox were already killed by it, Cray proprietary stuff died in favour of an Ethernet superset thus only the last bastion remains now.

Those things have Cray as their prime contractor so interconnect options not Slingshot aren't available.

 

Official AMD tweet

 

sounds like a GPU of the future. ;p

 

For real.

I wonder how Nanite would fare on CDNA. Still a good amount of texture work and framebuffer blending involved I imagine.

 

Going by this report be the next platform, the lopsided race between AMD and Intel for the exaflops machine seems to be one for third place:
https://www.nextplatform.com/2021/10/26/china-has-already-reached-exascale-on-two-separate-systems/

 

For texture sampling, we could avoid emulating logic behind texture compression to lower computational cost but it would still be slow. Nanite has already solved the blending problem since it bans translucency altogether so the traditional restrictions behind the graphics pipeline following the API primitive order rule doesn't apply which makes Nanite's graphics pipeline by design more parallel ...

 

Aurora will be 2 Exaflops
https://www.nextplatform.com/2021/1...flops-by-2027-pushes-aurora-above-2-exaflops/

 

Not super relevant since it's now same timeline as El Cap.

 

Spock with M100.

what surprised me, porting of code was straightforward.

https://www.osti.gov/biblio/1817474

 

Yes but there's PITA after that

 

How many GPU do you see in the picture.

https://pbs.twimg.com/media/FDbmaNwWEAw0D7r?format=jpg&name=large

 

8, it's a standard OAM baseboard.