Discussion RDNA4 + CDNA3 Architectures Thread

[DisEnchantment]

With the GFX940 patches in full swing since first week of March, it is looking like MI300 is not far in the distant future!
Usually AMD takes around 3Qs to get the support in LLVM and amdgpu. Lately, since RDNA2 the window they push to add support for new devices is much reduced to prevent leaks.
But looking at the flurry of code in LLVM, it is a lot of commits. Maybe because US Govt is starting to prepare the SW environment for El Capitan (Maybe to avoid slow bring up situation like Frontier for example)

See here for the GFX940 specific commits

History for llvm/lib/Target/AMDGPU - llvm/llvm-project

The LLVM Project is a collection of modular and reusable compiler and toolchain technologies. - History for llvm/lib/Target/AMDGPU - llvm/llvm-project

github.com

Or Phoronix

More AMD "GFX940" Enablement Work Landing In LLVM - Phoronix

www.phoronix.com

There is a lot more if you know whom to follow in LLVM review chains (before getting merged to github), but I am not going to link AMD employees.

I am starting to think MI300 will launch around the same time like Hopper probably only a couple of months later!
Although I believe Hopper had problems not having a host CPU capable of doing PCIe 5 in the very near future therefore it might have gotten pushed back a bit until SPR and Genoa arrives later in 2022.
If PVC slips again I believe MI300 could launch before it

This is nuts, MI100/200/300 cadence is impressive.



Previous thread on CDNA2 and RDNA3 here

Question - Speculation: RDNA3 + CDNA2 Architectures Thread

Man I have been dying to make this one for a while now. First rumours for RDNA3 are here so new thread time! Just going to start off with this one for now: kopite7kimi on Twitter: "@VideoCardz Ah, I mean a simple mcm design with 10240 cores is not enough. Because the lift from RDNA2 to RDNA3...

forums.anandtech.com

 

[Timorous]

Quick and dirty is that the 9070XT has about 50% more performance than the 7800XT with a very similar config.

Doing the same at 96CUs means 7900XTX + 50% with a very similar config.

So sure the numbers are super rough but even if error bars are +/-10% it gives you a ballpark.

 

[adroc_thurston]

Quick and dirty is that the 9070XT has about 50% more performance than the 7800XT with a very similar config.

Doing the same at 96CUs means 7900XTX + 50% with a very similar config.

So sure the numbers are super rough but even if error bars are +/-10% it gives you a ballpark.

The scaling doesn't quite work like that and N48 haez moar power.

 

[randomhero]

rant/
What a disappointment. Once again AMD snatches defeated from jaws of victory.
You can't win if you don't show up.

Great arch, poor execution . Where's halo AMD? You made strix halo, why not rdna halo?
/rant

P. S. I had to, sorry guys

 

[adroc_thurston]

Where's halo AMD? You made strix halo, why not rdna halo?

dead

 

[randomhero]

dead

Yeah, what a disappointment

 

[adroc_thurston]

Yeah, what a disappointment

hindsight is 20/20

 

[Saylick]

Quick and dirty is that the 9070XT has about 50% more performance than the 7800XT with a very similar config.

Doing the same at 96CUs means 7900XTX + 50% with a very similar config.

So sure the numbers are super rough but even if error bars are +/-10% it gives you a ballpark.

I came up with a similar conclusion.

RDNA 4's CUs have like 30+% more performance per clock than RDNA 3 since the 9070 XT has a 25% clock advantage but a 20% CU deficit against the 7900 GRE and yet is 40% faster. It kind of makes like AMD's Maxwell moment imo. Too bad they didn't scale this up any further.

 

[jpiniero]

I came up with a similar conclusion.

RDNA 4's CUs have like 30+% more performance per clock than RDNA 3 since the 9070 XT has a 25% clock advantage but a 20% CU deficit against the 7900 GRE and yet is 40% faster. It kind of makes like AMD's Maxwell moment imo. Too bad they didn't scale this up any further.

Clock speed difference might be a lot more than 25%.

 

[adroc_thurston]

Clock speed difference might be a lot more than 25%.

Na, it ain't.

 

[jpiniero]

Na, it ain't.

7900GRE is basically crappiest tier grade quality, no? If you were comparing it to the 7900 XT, that might be 25%.

 

[APU_Fusion]

A lot of assumption it could scale up it seems. Also, given Nvidia’s power blow up at the high end I wonder if amd chose to not ride the high power 600 watt lightning knowing even if they matched 5090 their yields would suck and no one would buy them anyway.

 

[adroc_thurston]

7900GRE is basically crappiest tier grade quality, no? If you were comparing it to the 7900 XT, that might be 25%.

RDNA4 clocks vary a lot more depending on how hard you hammer the memory.
Remember, this thing is not like the others.

 

[Saylick]

Clock speed difference might be a lot more than 25%.

25% is within the ballpark. It really depends which SKUs you compare.

I really hope the next gen arch aims to double 9070XT performance across the board. That would get us to 5090 performance but they need to reach up into a higher performance bracket imo. With 3nm and 50% larger die (~525mm2), I think it's possible.

 

[adroc_thurston]

I really hope the next gen arch aims to double 9070XT performance across the board. That would get us to 5090 performance but they need to reach up into a higher performance bracket imo. With 3nm and 50% larger die (~525mm2), I think it's possible.

boring and niche, if you're spending this R&D you do a 500 on 500, N2 on N4C and you win.

 

[SolidQ]

I really hope the next gen arch aims to double 9070XT performance across the board. That would get us to 5090 performance but they need to reach up into a higher performance bracket imo. With 3nm and 50% larger die (~525mm2), I think it's possible.

We still don't know how many CU have RDNA5, and per CU perf
Question how is NV Vera Rubin same arch, just fattier, or new arch

 

[adroc_thurston]

We still don't know how many CU have RDNA5, and per CU perf

Not many and a lot more.

 

[SolidQ]

Not many

Per cu?

 

[adroc_thurston]

Per cu?

As in not many CUs.
They're more interested in funnier things.

 

[GTracing]

I came up with a similar conclusion.

RDNA 4's CUs have like 30+% more performance per clock than RDNA 3 since the 9070 XT has a 25% clock advantage but a 20% CU deficit against the 7900 GRE and yet is 40% faster. It kind of makes like AMD's Maxwell moment imo. Too bad they didn't scale this up any further.

Wouldn't the 7800 XT be a better comparison? The 7900 GRE is a lopsided design. It has 33% more cores than the 7800 XT, but is only 10% faster in games.

 

[adroc_thurston]

The 7900 GRE is a lopsided design. It has 33% more cores than the 7800 XT, but is only 10% faster in games.

N31 chopchop clocks lot lower than the full N32.
Raw compute delta isn't anything too high iirc.

 

[Saylick]

boring and niche, if you're spending this R&D you do a 500 on 500, N2 on N4C and you win.

Well, yeah that would be very interesting if they can FINALLY get GPU gaming chiplet to work, but it's a big gamble if they can't. Not sure if Lisa/Jack has the risk appetite for it when they can fallback on another monolithic design.

 

[GTracing]

N31 chopchop clocks lot lower than the full N32.
Raw compute delta isn't anything too high iirc.

The 7900 GRE only clocks ~8% lower than 7800 XT. Even taking that into account, the 7800 XT achieves ~25% more frames per CU.

 

[adroc_thurston]

Well, yeah that would be very interesting if they can FINALLY get GPU gaming chiplet to work, but it's a big gamble if they can't

Oh I'm not advocating for N4c sequel.
Something far far simpler instead.

 

[Timorous]

Oh I'm not advocating for N4c sequel.
Something far far simpler instead.

E core spam, GPU edition.

 

[adroc_thurston]

E core spam, GPU edition.

AMD is doing outright the opposite though.