AMD CDNA Discussion Thread

[Deleted member 90741]

For revenue or for bring-up? I'm asking because, you know, Intels 10nm products have been "shipping to customers since 2017" (maybe they used a sailboat for that and had strong headwind). So, there clearly is a difference between shipping (samples for qualifications and bring-up) and shipping (for actual market introduction).

Maybe AMD just replaced SPOCK systems at ORNL with some real ones for now.

Isn't that what happened with Milan as well. Perlmutter was active before Milan was launched.
Similarly, Frontier will be active by the time CDNA2 launches if not before.
If HPE gets CDNA2 in end of 21 then they can only deliver Frontier by Q2 2022 at best if not later
There is a thread on freedesktop during discussion on HMM that HPE, AMD and partners are working with the real HW for testing HMM. So the bringup is already done by Q1, and now we have HMM already in 5.14 and ROCm 4.3 already launched with HMM.

 

[del42sa]

https://videocardz.com/newz/amd-ins...aldebaran-gpu-might-feature-110-compute-units

 

[Leoneazzurro5]

It is obviously wrong (their CU estimation) as it was laready leaked that Aldebaran would have 110CU per die, 220 total (2 Die MCM).

 

[Bondrewd]

as it was laready leaked that Aldebaran would have 110CU per die, 220 total (2 Die MCM).

?
112/224.

 

[Leoneazzurro5]

?
112/224.

Yep, I checked the leak and It was indeed 112, even if there are more CU per die (yield reasons, I suppose). I was confused by the name.

 

[Kaotik]

Pretty sure it's actually 128/256, with some disabled for that 112/224, similar to how Arcturus was 120 active on 128
edit: even though disabling that many sounds a lot, 112 just sounds wrong after Arcturus.

 

[no-X]

https://videocardz.com/newz/amd-ins...aldebaran-gpu-might-feature-110-compute-units

So VideoCardz is saying, that AMD (compared to MI100) tripled bandwidth, quadrupled memory capacity and changed form factor to allow twice as high TDP… all that to support less CUs?

 

[Kaotik]

So VideoCardz is saying, that AMD (compared to MI100) tripled bandwidth, quadrupled memory capacity and changed form factor to allow twice as high TDP… all that to support less CUs?

They're suggesting 110 per chiplet quite surely, not 110 total.

 

[no-X]

Their table states 110 per Aldebaran.

 

[Bondrewd]

So VideoCardz is saying, that AMD (compared to MI100) tripled bandwidth, quadrupled memory capacity and changed form factor to allow twice as high TDP… all that to support less CUs?

Uh that's per tile and they're full rate DPFP now.

 

[Granath]

https://videocardz.com/newz/amd-instinct-mi300-gpu-to-feature-4-graphics-chiplets

 

[Bondrewd]

https://twitter.com/i/web/status/1438564239933853698

Ouch, a beaniespill.

MI300 is yes, a BIG APU.

 

[Leoneazzurro5]

https://twitter.com/i/web/status/1438564239933853698

Ouch, a beaniespill.

MI300 is yes, a BIG APU.

Epyc + CDNA on the same socket?

 

[Bondrewd]

Epyc + CDNA on the same socket?

Yyyyes but actually no.
There's a nicer, sleeker and saner way to build big APUs present in AMD's toolkit.

 

[Leoneazzurro5]

Yep, I meant obviously with stacking and enhanced interconnects.

 

[Bondrewd]

Yep, I meant obviously with stacking and enhanced interconnects.

Yes.
They don't just put CPU and GPU side by side a-la Grace Hopper naaah.

It's 4 GCDs each having a few CCDs directly attached to them.

 

[Jawed]

That sounds epic!

 

[Bondrewd]

That sounds epic!

Epic(ally expensive), yes.

 

[Deleted member 90741]

That sounds epic EPYC!

Indeed

 

[Lightman]

Epic(ally expensive), yes.

Price is relative, technological progress is what's important.