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

 

[SolidQ]

Guys like this is something...

 

[Keller_TT]

Same as RDNA3 at a similar junction?

Perhaps we shouldn't read too much into this. Maybe AMD were just hesitant to go first and were licking their chops hoping for Jensen to peg the 5070 at min $600.
As for the 9070 XT, if they did an average case improvement from RDNA3 (7800 XT/7900 GRE) w.r.t to CU/Clocks/PPW, they will be in 4070 TiSu territory, and 7900 XT isn't a stretch if RDNA 4 is more sorted. For RT, my guess is that they'll be marginally better than 4070 Ti.
If they do manage this, $500 max is a good price for MBA card.

 

[Win2012R2]

Maybe AMD were just hesitant to go first

For sure, but why did they even plan to say anything about RDNA4 at CES when they had lots of other stuff? Should have just delayed it, seems like obvious move - let Jensen show his cards first

 

[jpiniero]

For sure, but why did they even plan to say anything about RDNA4 at CES when they had lots of other stuff? Should have just delayed it, seems like obvious move - let Jensen show his cards first

AIBs wanted to show their cards at CES.

 

[Heartbreaker]

For sure, but why did they even plan to say anything about RDNA4 at CES when they had lots of other stuff? Should have just delayed it, seems like obvious move - let Jensen show his cards first

Reading between the lines of that AMD Q&A, RDNA4 really isn't quite ready yet (sounds like HW is so that leaves software).

NVidia is doing a January release, it sounds like AMD will be releasing later in the quarter.

 

[Win2012R2]

AIBs wanted to show their cards at CES.

And?

It's terrible timing, AMD missed Xmas - they should have released before Nvidia.

Reading between the lines of that AMD Q&A, RDNA4 really isn't quite ready yet

Another good reason not to show it then.

 

[SolidQ]

View this content on Instagram

View this content on Instagram

24.1.2025?

 

[marees]

NVidia is doing a January release

5070 is early February

N48 will probably be launched after (comparing) that, I think

Seems that 5070 has raster of 4070 super & RT of 4070 ti / super

That is what N48 needs to beat

 

[In2Photos]

AIBs wanted to show their cards at CES.

Gigabyte had their cards at the show.

https://twitter.com/x/status/1876442039321251860

 

[Keller_TT]

Reading between the lines of that AMD Q&A, RDNA4 really isn't quite ready yet (sounds like HW is so that leaves software).

NVidia is doing a January release, it sounds like AMD will be releasing later in the quarter.

My guess too. I think FSR 4 isn't yet ready for showdown with well honed DLSS 3.x stack, let alone the *transformed* neural numbness to come. Remember how FSR 3 went after hyped up *reveal* and it launched after an eternity and yet did nothing for upscaling or denoising. AFMF 2 did more for the peasants with potatoes than 3rd FSR and is one good tick for Team Red.

 

[Heartbreaker]

My guess too. I think FSR 4 isn't yet ready for showdown with well honed DLSS 3.x stack, let alone the *transformed* neural numbness to come. Remember how FSR 3 went after hyped up *reveal* and it launched after an eternity and yet did nothing for upscaling or denoising. AFMF 2 did more for the peasants with potatoes than 3rd FSR and is one good tick for Team Red.

I think now they will want/have to cram in MFF - Multiple Fake Frames, if they weren't already doing that.

 

[insertcarehere]

It does not look like RT improvements are meaningful at all as otherwise CP2070 would have gotten a lot quicker even without DLSS

The original point being that:
-As per NV's figures 5070 is roughly 30% faster than 4070 in FC6 RT.
-From TPU testing, RT performance declines are remarkably even across the board from Ada to RDNA2, despite how well suited (or not) each architecture handles RT.
-Therefore it stands to reason that unless Blackwell breaks this trend of 2 OEMs and 4 architectures 5070's raster perf vs 4070 should be in line with RT perf in FC6 (I. E roughly 30% advantage).
- With the above assumptions, it is hard to see reasonable perf projection where a 9070XT (I. E not 4080/7900XTX) would 'easily beat' a 5070 (4070+30%) in Raster performance.

 

[eek2121]

I think this release is going to shock people, in a good way. AMD is going to nail the execution. They just didn’t have room at CES and didn’t want to launch a card when NVIDIA is stealing the show.

The pricing and performance is going to be better than people think.

 

[exquisitechar]

I think this release is going to shock people, in a good way. AMD is going to nail the execution.

AMD hasn’t been nailing the execution of anything after Zen 3 and RDNA2. 9070 XT is going to be another typical “slightly undercut NV for the same performance” kind of GPU. At least, maybe they catch up a bit to NV in terms of PPA after Ada trounced the flop that was RDNA3.

 

[jpiniero]

The original point being that:
-As per NV's figures 5070 is roughly 30% faster than 4070 in FC6 RT.
-From TPU testing, RT performance declines are remarkably even across the board from Ada to RDNA2, despite how well suited (or not) each architecture handles RT.
-Therefore it stands to reason that unless Blackwell breaks this trend of 2 OEMs and 4 architectures 5070's raster perf vs 4070 should be in line with RT perf in FC6 (I. E roughly 30% advantage).
- With the above assumptions, it is hard to see reasonable perf projection where a 9070XT (I. E not 4080/7900XTX) would 'easily beat' a 5070 (4070+30%) in Raster performance.

The 5070 might even be slower than the 4070 NS in raster.

 

[gdansk]

The 5070 might even be slower than the 4070 NS in raster.

NonSuper? I don't see how. It has more tflops and more memory bandwidth.

 

[poke01]

The 5070 might even be slower than the 4070 NS in raster

How?? I don’t see that.

 

[insertcarehere]

The 5070 might even be slower than the 4070 NS in raster.

Based on what? Nvidia's provided figures (c. +30% in FC6 RT), FC6 doesn't actually discriminate between architectures when converting Raster - > RT performance penalties so there's no evidence that this gap wouldn't mostly hold.

Based on what's provided, with a c. 15% perf penalty for activating RT w/ 4070 in FC6. The only way one squares a 5070 being 30% faster in RT while being slower than 4070 in raster is if the 5070 is faster with RT on than no RT - I. E not possible.

 

[jpiniero]

Based on what?

Suspecting there's an game clock regression. I'd say it's more likely to be about the same in raster. So the 5070 would be about the same to slightly slower than the 4070 NS in raster and about 20-30% faster in RT.

 

[insertcarehere]

Suspecting there's an game clock regression. I'd say it's more likely to be about the same in raster. So the 5070 would be about the same to slightly slower than the 4070 NS in raster and about 20-30% faster in RT.

As I said, It's basically not possible for a 5070 to do such a thing, at least not in Farcry 6, which is the basis here:


A 4070 only has a 15% penalty between RT and no RT, for 5070 to get flat raster and +20%-30% RT would literally mean higher FPS under RT. If Blackwell perchance does achieve that (no penalty RT vs Raster) then AMD has a much bigger problem on their hands than relative raster perfromance..

 

[Heartbreaker]

Suspecting there's an game clock regression. I'd say it's more likely to be about the same in raster. So the 5070 would be about the same to slightly slower than the 4070 NS in raster and about 20-30% faster in RT.

I wouldn't expect that. We see lower clocks for the 5090, because the die is so much bigger (+30% CUDAs), to keep power/heat in check.

But down on the 5070, they only have 4% more CUDA, so they will be boosting clocks to compensate. Power is also up 25%/50watts, despite nearly same CUDA count so the must really be pushing clocks on this one, plus it has a lot more memory BW to play with.

I have no doubt raster of 5070 will exceed 4070, but I think it won't get much beyond 4070 Super, if at all.

 

[SolidQ]

9070 non XT is 16GB, Seems 9600XT is N48 12gb version

 

[JustViewing]

I'm setting the bar to this low in order to be pleasantly surprised by AMD beating it, instead of dissapointed *again*:

Yes, that is the correct way to go with AMD product launches. Being pessimistic has its benefits, like I am not at all dissapointed at Zen 5 core performance.

 

[Panino Manino]

I think this release is going to shock people, in a good way. AMD is going to nail the execution. They just didn’t have room at CES and didn’t want to launch a card when NVIDIA is stealing the show.

The pricing and performance is going to be better than people think.

Nail the execution?
So what? AMD went from trying to compete with 3090, to competing with the 4080 to now competing against the 5070.

So what if "price and performance" is "good"? Just lower the price of the current generation, or just redirect people to Nvidia store as AMD see to be trying to offer the same tricks Nvidia perfected and of course will fail to deliver.


... aah, before AMD stops pretending and gives up for real from this market I have to give the example first.
I don't care anymore, only thing left around Radeon is frustration and this is something I don't need. I'll stop seeing any news and comments.

Sorry to bother you good people.

 

[Gideon]

9070 non XT is 16GB, Seems 9600XT is N48 12gb version

I really hope this ends up true, but are you certain though this is RX 9070 and not a 9700 XT where the "XT" part is poorly obscured by the I/O bracket?

Most of the sites that feature (no doubt copy-pasted material from ASUS) seem to mention it's a 9700 XT (1, 2, 3, ....)