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

 

[gdansk]

Have these guys EVER worked at a company before?

Yes. I know it sounds delusional but they have actually worked with AMD's marketing team before which gives many more insights to their m.o. than mopping floors at Kroger.

AMD let one person run one benchmark. It's probably just incompetence but who knows.

 

[techjunkie123]

I think it will end up a bit faster than 5070 in raster, but AMD won't be able to charge more because of RT/DLSS/Reflex 2/etc

Better be at least one tier faster lol. 5070 probably won't be much faster than 4070 in raster, maybe ~7800XT-7900GRE.

Also if they did finally dedicate all the extra die space (~100mm^2 if 64 CUs) to tensor cores, then I hope the RT is good! Might not be though lol. And as you said the sad part is they're competing against RT + FGx3, not just pure RT.

 

[adroc_thurston]

And as you said the sad part is they're competing against RT + FGx3, not just pure RT.

They can fake more frames with FSR3 just fine.

 

[techjunkie123]

They can fake more frames with FSR3 just fine.

Yeah but gamers™ think FSR sux compared to DLSS. With some objectivity in there. Iso - feature is a loss if your features are worse .

 

[gaav87]

I always thought from my calc based on RDNA2 -> RDNA3 jump its ~320-353mm2 (rdna2->rdna3 transistor scaling is around 0.68x)
Maybe 128MB Infinity cache is back ? Or the 64CU is just wrong and its back to 6900xt levels that would explain the extra die size.

 

[iLLusiveMan]

I always thought from my calc based on RDNA2 -> RDNA3 jump its ~320-353mm2 (rdna2->rdna3 transistor scaling is around 0.68x)
Maybe 128MB Infinity cache is back ? Or the 64CU is just wrong and its back to 6900xt levels that would explain the extra die size.

Worst case scenario this will be around a 7900XT as their official slides suggested. Best case it will be around 7900XTX. I don't see how can a 9070XT be lower than the worst case.

 

[Abwx]

COD favors AMD heavily

12-13% over a 20 games average, so better than average but certainly not heavily.

AMD Radeon RX und Nvidia GeForce RTX im Benchmark

ComputerBase hat den Testparcours für Grafikkarten aktualisiert. AMD Radeon, Nvidia GeForce und Intel Arc sind im Benchmark mit dabei.

www.computerbase.de

 

[gaav87]

12-13% over a 20 games average, so better than average but certainly not heavily.

AMD Radeon RX und Nvidia GeForce RTX im Benchmark

ComputerBase hat den Testparcours für Grafikkarten aktualisiert. AMD Radeon, Nvidia GeForce und Intel Arc sind im Benchmark mit dabei.

www.computerbase.de

That would put 9070xt at 7% faster than 7900xtx in HUB benchmarks at 4K... Hope this sticks and not the 7900xt version xD

 

[adroc_thurston]

Yeah but gamers™ think FSR sux compared to DLSS. With some objectivity in there. Iso - feature is a loss if your features are worse .

Yes but if FSR4 is good enough at launch and Jack trucks a pile of Dolla to digital foundry, it's gonna be ok

 

[adroc_thurston]

I always thought from my calc based on RDNA2 -> RDNA3 jump its ~320-353mm2 (rdna2->rdna3 transistor scaling is around 0.68x)
Maybe 128MB Infinity cache is back ? Or the 64CU is just wrong and its back to 6900xt levels that would explain the extra die size.

They're just meaner CUs. They redid their caching setup. I dunno how exactly.

 

[Jan Olšan]

Guy says CoD 6 may have issue where it doesn't apply graphic setting changes until you restart the game - meaning, that 99FPS score may actually be deceiving and the benchmark could have silently used worse settings than advertised.

 

[adroc_thurston]

Guy says CoD 6 may have issue where it doesn't apply graphic setting changes until you restart the game - meaning, that 99FPS score may actually be deceiving and the benchmark could have silently used worse settings than advertised.

This implies they changed the settings and it did not default to Extreme.

 

[iLLusiveMan]

Guy says CoD 6 may have issue where it doesn't apply graphic setting changes until you restart the game - meaning, that 99FPS score may actually be deceiving and the benchmark could have silently used worse settings than advertised.

Says the guy who got 77 fps with a 7900XTX and 67 fps with a 7900XT. Something is seriously wrong with his PC...

 

[PJVol]

Did AMD said at CES that all future leaks are wrong either?

 

[Abwx]

That would put 9070xt at 7% faster than 7900xtx in HUB benchmarks at 4K... Hope this sticks and not the 7900xt version xD

Dunno for the final perf but you ll notice that at 4k Computerbase are quite below 99 FPS with all cards and despite FSR/DLSS enabled.

That being said i have no clue about AMD GPUs exact characteristics, according to Computerbase a single SM has two ALUs and can hence do 4 flops/cycle, at 4096 SP
and 2.8GHz this would yield about 45 Tflops, that s the number that matter for raster perf, so either they got dual issue working or they simplified the design by using two separate ALUs per SM.

Edit : That s their presets :

Extrem-Preset, Tessellation Alle, DLSS/FSR Native
Full HD: DLSS/FSR Quality

 

[GodisanAtheist]

COD favors AMD heavily so AMD beating a 4090 isn't all that magical as it sounds. Will be very different in other games especially NV favoring ones with heavy RT.

If the die size is indeed around 350mm2 or bigger (vs the leaked 270mm2 ish) and the huge ass coolers + 3 pin options pointing to high power use = high clocks, that performance doesn't seem that unrealistic. I'm thinking 3.2ghz clocks or more.

There is one way: bad performance per area. And the forecast is looking likely.

- To channel my best Adroc PPA MAX IS ALL THAT MATTERS

 

[iLLusiveMan]

Did AMD said at CES that all future leaks are wrong either?

They said that all leaks are inaccurate because nobody has the full performance driver besides them.

Journalist: Are you guys cool with partner leaks? Everyone’s going to know the specs just by seeing the cards.




McAfee: I think you’ll see static demos of cards. Everybody loves a good wall of boards. And I think you’ll see that from all of our partners in their spaces. I don’t think you’ll see any live demos, or you better not see any demos from partners — I’ll put it that way. But, you know, power connectors, things like that, TDPs, I think you’ll see a lot of that stuff out there.


Azor: By the way, if you do see demos out there, just know that they don’t have the production [software] driver.




McAfee: All these performance leaks, well, it is accurate for the way the driver performs on the card right now. It is nowhere near where the card will actually perform once we release the full performance driver.


Journalist: Did that also factor into your decision?




Azor: It’s not a readiness issue.


McAfee: We have in house the full performance driver. We intentionally chose to enable partners with a driver which exercises all of the, let’s call it, thermal-mechanical aspects of the card, without really running that risk of leaking performance on critical aspects of the product. That’s pretty standard practice.

Souce: PcWorld

 

[PJVol]

VideoCardz mentions the CoD test showed artifacts, so there is possibility of FPS being lower once the rendering is fixed with proper drivers.

She (IGN reporter) said there was a bug similar of missing gun in benchmark. What artefacts? )

They said that all leaks are inaccurate because nobody has the full performance driver besides them

So they let her run a game benchmark which is a priori inaccurate?)

 

[Jan Olšan]

She (IGN reporter) said there was a bug similar of missing gun in benchmark. What artefacts? )

The missing gun bug is apparently common (not a failure of these particular drivers and GPU), the dude replicating the benchmark in a video few posts back was getting it too.

 

[iLLusiveMan]

So they let her run a game benchmark which is a priori inaccurate?)

Either way I don't care. I want the official specs and third party reviews.At least the official specs should have been presented in the keynote. They screwed this up big time.
I'm tired of all the leaks and speculations.

 

[tajoh111]

If the die is indeed 380mm2, that would make more sense with 64 CU's with upgraded ray tracing. It did not make sense at 250mm2. However, I can see why AMD wants to pull out of the discrete market to some extent with that die size.

If this comes out as a 500 dollar products AMD cannot be particularly happy. That means selling a product comparable in cost to manufacture as a 6900xt for 500 dollars. This product also costs more to make than a 7800xt while selling for the same amount. That makes it difficult to recover R and D.

It would make sense why Lisa wants to remove resources from the division because RTG financial trajectories are not promising.

 

[adroc_thurston]

that would make more sense with 64 CU's with upgraded ray tracing

not even what eats the area there.

 

[Elfear]

If the die is indeed 380mm2, that would make more sense with 64 CU's with upgraded ray tracing. It did not make sense at 250mm2. However, I can see why AMD wants to pull out of the discrete market to some extent with that die size.

If this comes out as a 500 dollar products AMD cannot be particularly happy. That means selling a product comparable in cost to manufacture as a 6900xt for 500 dollars. This product also costs more to make than a 7800xt while selling for the same amount. That makes it difficult to recover R and D.

It would make sense why Lisa wants to remove resources from the division because RTG financial trajectories are not promising.

If true, much to everyone's detriment. 😕

 

[Heartbreaker]

If the die is indeed 380mm2, that would make more sense with 64 CU's with upgraded ray tracing. It did not make sense at 250mm2. However, I can see why AMD wants to pull out of the discrete market to some extent with that die size.

AMD has often only competed at mid to low end, that isn't pulling out of the market.

It would make sense why Lisa wants to remove resources from the division because RTG financial trajectories are not promising.

Where have we seen they are pulling resources from the GPU business?

 

[adroc_thurston]

AMD has often only competed at mid to low end, that isn't pulling out of the market.

they did not.
Stuff like Polaris is an outlier.

Where have we seen they are pulling resources from the GPU business?

you didn't.