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

 

[Keller_TT]

I think I know exactly what to expect after reality has dawned. There is enough proper CES reporting info and new leaks ballpark.

A 250-260W 9070 and 300W 9070 XT MBA cards. (AMD could do a deceptive 295W naming)
Performance: 9070 a 7900GRE Pulse in raster, catching up to 4070S in RT, <5% better in some cases.
9070 XT MBA: 4070 TiS ± 5% across the board. AIBs can crank up perf by 10% for even inferior efficiency.

A few odd titles favoring AMD like CoD will be shown as uplift ballpark and it'll be a Zen 5% like launch affair after all the hype.

Does Lisa have the will to shake off Cheap & Inefficient tagline of AMD GPUs?

 

[adroc_thurston]

A 250-260W 9070 and 300W 9070 XT MBA cards. (AMD could do a deceptive 295W naming)
Performance: 9070 a 7900GRE Pulse in raster, catching up to 4070S in RT, <5% better in some cases.
9070 XT MBA: 4070 TiS ± 5% across the board. AIBs can crank up perf by 10% for even inferior efficiency

All of this is wrong.
The full die is 7900XTX perf, the chop is 7900XT. That's it. Both MBA, idk the wattage.
They're boring but good parts. Pricing will decide their fate more or less.

Does Lisa have the will to shake off Cheap & Inefficient tagline of AMD GPUs

No she personally killed Navi40/41/42.

 

[Win2012R2]

Just good GFX IP, down to the very basics.

Will they change everything in UDNA or build on that?

 

[adroc_thurston]

Will they change everything in UDNA or build on that?

RDNA5 is gfx13 aka more good more better.
UDNA doesn't exist. Never existed. Will never exist.
DC and client parts live on very very separate tracks and design targets.

 

[Win2012R2]

RDNA5 is gfx13 aka more good more better.

N3-ish I hope, and they'd better have bigger chip, ideally next year too

 

[SolidQ]

The full die is 7900XTX perf, the chop is 7900XT.

12GB n48 is 7800XT, but for like 330$?

RDNA5 is gfx13 aka more good more better.

This time N52 or N51 max? But no N50

9070 XT MBA: 4070 TiS ± 5% across the board.

You saw already worst case scenarion for AMD in Wukong near 4080s, in Cyber RT 4070S. Normal scenario would be better

 

[adroc_thurston]

N3-ish I hope, and they'd better have bigger chip, ideally next year too

N3p probably.

12GB n48 is 7800XT, but for like 330$?

I've no idea.

This time N52 or N51 max? But no N50

numbers do not matter, all that matters is +1 die targeting kilobuck range.

 

[SolidQ]

all that matters is +1 die targeting kilobuck range

Based on Nvidia 5080 price, seems it's xx80 class GPU from AMD

 

[adroc_thurston]

Based on Nvidia 5080 price, seems it's xx80 class GPU from AMD

duh

 

[Win2012R2]

N3p probably.

So maybe +10% clocks and +20% density?

At least memory will be GDDR7 and by that point 35-40 gbits should be around, AMD needs 3D cache under GPU, hope RDNA5 will get it sorted

 

[SolidQ]

The full die is 7900XTX perf, the chop is 7900XT

if full die real XTX, that mean RDNA5 like xx60 clas = XTX, xx70 class = 4090, xx80 class above, that usually happens with every tier on next cards

 

[adroc_thurston]

So maybe +10% clocks and +20% density?

N4 2-2 to N3 2-1 should be that or a bit more I guess? Don't remember TSM stats for N3e + N3p cumulatively.

AMD needs 3D cache under GPU

ain't happening.

hope RDNA5 will get it sorted

dead. Colossals are Very Dead. Only the simplest monodies ever on the roadmap now.

if full die real XTX, that mean RDNA5 like xx60 clas = XTX, xx70 class = 4090, xx80 class above, that usually happens with every tier on next cards

It doesn't mean anything since cost per xtor stopped declining.

 

[SolidQ]

dead. Colossals are Very Dead.

Lisa after looking RTX 5090 results, "Revive Colossal for RDNA6!"

 

[poke01]

https://twitter.com/x/status/1878700875285410062

its from chiphell, real or bs?

 

[ToTTenTranz]

It doesn't mean anything since cost per xtor stopped declining.

It may (or should) not be this way forever. The only real reasons for this are TSMC having few top-of-the-line fabs doing the highest-end processes, Intel and Samsung not being able to compete at those processes, and AI chips with massive markups taking up such a huge portion of the high-end fab capacity.


It's not like all these factors are bound to stagnate forever. TSMC has the "we need to stay relevant to the world or China invades" factor but still they're building fabs for high-end processes outside Taiwan. Intel and Samsung aren't sitting idle, and AI hardware sounds increasingly like a bubble.


It's not like TSMC has enormously increased costs for their newer processes, e.g. ASML isn't charging them 20x more for their 2nm scanners than they did for the 28nm ones and the costs for their scanners are also paid off after a couple months into full production.
This is all supply and demand caused by AI demand and lack of competition over TSMC's high-end fabs.

https://twitter.com/x/status/1878700875285410062

its from chiphell, real or bs?

It's both real and.. not new? AMD confirmed they were also waiting for Nvidia to play their hand about a week ago.

 

[SolidQ]

its from chiphell, real or bs?

Maybe they waiting 5070/5070ti raster perf, if bad, than up price?

 

[Keller_TT]

You saw already worst case scenarion for AMD in Wukong near 4080s, in Cyber RT 4070S. Normal scenario would be better

Wukong is software RT that rewards raster. Besides, that was AIB OC card by the look of it (330W)
I really don't care about Cyberpunk. Not a game in our house and it's sponsored by and optimized for Nvidia to show off DLSS, RT. All lighting gimmicks are meh as the setting is all artificial and can just as well be like "realism" of using another set of lights and time of day. It looks more than punk enough on PS5.

I'll judge performance by Avatar FoP, Plague's Tale Requiem, RDR2, Indiana Jones, Rachet & Clank, F1 24, sim racing titles in VR without glitches.

 

[marees]

Maybe they waiting 5070/5070ti raster perf, if bad, than up price?

That will be known only 1st week of feb

AMD may not set RDNA 4 price until then

 

[soresu]

So they preserved CUDA and added the DLSS lockin

CUDA is their entire compute platform that keeps the industry locked into their hardware, tensor cores just being the newest domain specific addition to that.

The compute silicon they call "CUDA cores" is the ALUs used for shading and other assorted GPU compute in games.

If nVidia had gone down AMD's path and bifurcated their GPU development with a CDNA equivalent lineage then they probably would have kept DLSS on CUDA cores if they ever made it.

 

[SolidQ]

That will be known only 1st week of feb

Unless spy in Nvidia sending them results

Wukong is software RT that rewards raster.

a lot games working and coming on UE5 engine.

 

[soresu]

AMD may not set RDNA 4 price until then

It wouldn't be the first time they set a price and then revised it afterward.

The only thing they can't do after is revise it upward.

I mean, they can - it would just be a terrible idea.

 

[poke01]

I'll judge performance by Avatar FoP, Plague's Tale Requiem, RDR2, Indiana Jones, Rachet & Clank, F1 24, sim racing titles in VR without glitches.

These are good starting titles to test out RT for RDNA4. I'll add them to the RT benchmarking thread later when results comes out.

 

[adroc_thurston]

If nVidia had gone down AMD's path and bifurcated their GPU development with a CDNA equivalent lineage

That's funny. Because they did, with H100.

 

[soresu]

Maybe they waiting 5070/5070ti raster perf, if bad, than up price?

Given at least one game result for raster on 5090 was shown during CES they should be able to infer 5070 perf from that.

 

[SolidQ]

on 5090 was shown during CES they should be able to infer 5070 perf from that.

What games that was?
What if RTX 5070 is real 4070S perf or +5%