Discussion RDNA4 + CDNA3 Architectures Thread

DisEnchantment · Mar 23, 2022

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

Mahboi · Apr 29, 2024

Saylick said:
"Please nuke your margins so that Nvidia is pressured to lower their prices, so that I can finally buy that high-end Nvidia GPU I've always wanted but could not afford"

FTFY.

The number of Plebbitors that went around echoing "AMD has to lower its prices so I can buy Nvidia for cheaper"...
Drove me crazy after awhile

Mopetar · Apr 29, 2024

I'm not even sure NVidia would lower their prices. Once upon a time they might have, but I think they've discovered that their customers are almost more akin to junkies looking for their next fix than regular consumers and that they don't need to cut prices.

That and AMD isn't making enough volume to really affect NVidia's sales.

SolidQ · Apr 29, 2024

Mopetar said:
Navi 50 basically looks like Navi 40

More WGP/Bus?

adroc_thurston · Apr 29, 2024

SolidQ said:
More WGP/Bus?

new uarch but yes more stuff.
More stuff is how you win.

SolidQ · Apr 29, 2024

adroc_thurston said:
new uarch

So RDNA3/4/5 all diffence uarch now?

adroc_thurston · Apr 29, 2024

SolidQ said:
So RDNA3/4/5 all diffence uarch now?

now?

Mahboi · Apr 29, 2024

Mopetar said:
I'm not even sure NVidia would lower their prices

Obviously it wouldn't since the brainlets are unable to compute basic economics.
"If they lower prices, others will buy and I will tell everyone to only buy Nvidia"

Has some real "to pump air into the wheel, I will suck air out of the wheel and blow it back in" vibes.

adroc_thurston said:
new uarch but yes more stuff.
More stuff is how you win.

Same interconnect?
I was under the impression that RDNA 4 was originally going for a silicon bridge or something, but that's a year old tweet I can't recall too well.
I thought the interconnect was the issue for RDNA 4.

marees · Apr 29, 2024

adroc_thurston said:
More stuff is how you win.

Assuming they can get that crossfire of the 3 GCDs to work in RDNA 5 (unlike RDNA 4)

marees · Apr 29, 2024

SolidQ said:
So RDNA3/4/5 all diffence uarch now?

RDNA 3 & 4 likely to have same raster architecture with some tweaks for RT (which might feature in PS5 pro)

I am guessing RDNA 5 has a major jump in architecture. It should be driven by xbox next requirements, I guess

adroc_thurston · Apr 29, 2024

Mahboi said:
I was under the impression that RDNA 4 was originally going for a silicon bridge or something, but that's a year old tweet I can't recall too well.

yeah there's a even a cross-section posted in the wild.

marees said:
RDNA 3 & 4 likely to have same raster architecture with some tweaks for RT (which might feature in PS5 pro)

Oh no. Not even close.
gfx11 and gfx12 are unrelated for a whole bunch of reasons.

marees said:
I am guessing RDNA 5 has a major jump in architecture. It should be driven by xbox next requirements, I guess

Consoles have quite literally zero input on AMD roadmap (besides the DirectX spec bumps usual but that's the same ISV stuff as it was 20 years ago).

marees said:
Assuming they can get that crossfire of the 3 GCDs to work in RDNA 5 (unlike RDNA 4)

3 AIDs.
It's not crossfire.
And it worked for RDNA4.

igor_kavinski · Apr 29, 2024

marees said:
Intel will do this to AMD with battlemage

And that should make AMD compete better. They are at their best when sufficiently challenged. Also, I don't see long time AMD users suddenly switching to Intel unless Intel starts giving away their cards. Battlemage may have the performance and even better features but drivers and control panel are still something Intel needs a few more years to refine. Nvidia is more at risk of their 3050/4060 customers opting for an Intel upgrade due to better price/performance since Nvidia is not going to give them more VRAM or more performance for the same price.

soresu · Apr 29, 2024

Mahboi said:
The number of Plebbitors that went around echoing "AMD has to lower its prices so I can buy Nvidia for cheaper"...
Drove me crazy after awhile

Not gonna lie, Plebbitors is the most well heeled tech industry insult I have ever heard 🤣

gaav87 · Apr 29, 2024

branch_suggestion said:
Slower than the 7900XTX so that might live for a bit longer, rest of the stack will go EOL very quickly though.

Well we know it will be slower in compute then 7900xtx but compute =/= gaming performance. I know its copium but if they fixed the dual issue problems and each simd can access other SIMD data then we could see 4080 non super raster performance with close RT performance.
6950xt has almost 2.5x lower theorethical performance but is only 30-50% slower then 7900xtx in gaming.

Glo. · Apr 29, 2024

H T C said:
Problem with AMD is that they aren't "as ruthless" as nVidia.

Pricing their cards SLIGHTLY cheaper than nVidia @ same or slightly better performance IS NOT ENOUGH: AMD needs to "aim for the jugular" and either have a similar price for SUBSTANTIALLY HIGHER PERFORMING cards (i mean like 40%+ higher), or MUCH CHEAPER price for similar performance cards (i mean like 40%+ cheaper), EVEN IF they take a monetary loss in the immediate future BECAUSE of it.

That's the only way i see AMD gaining ground on nVidia.

AMD couldn't afford to take this approach before ZEN, but THEY CAN now.

This is actually hilarious.

RTX 3050 vs RX 6600.

RX 6600 used the same amount of power, had equally bad RT as 3050, had the same amount of VRAM, and it was sold for 40$ less. Oh and in rasterization - 6600 was 30-35% faster.

nvidia GPU still outsold 4:1 AMD GPU.

That is the reflection of how consumers make decisions on GPUs. So no. AMD had clearly better product than Nvidia, on every front. Perf/dollar, in sheer performance, and in the rest was tied.

Didn't matter. Lets put this lie to the grave, where it belongs. AMD will not gain traction, untill the market will change.

adroc_thurston · Apr 29, 2024

gaav87 said:
I know its copium but if they fixed the dual issue problems and each simd can access other SIMD data then we could see 4080 non super raster performance with close RT performance

Dual issue isn't the point.
It's not really dual issue either, closer to math packing.
Only MI300 has real dual issue and real wave slot count for that.

gaav87 · Apr 29, 2024

adroc_thurston said:
Dual issue isn't the point.
It's not really dual issue either, closer to math packing.
Only MI300 has real dual issue and real wave slot count for that.

Well it is a problem when 7900xtx has 61TFLOPs of FP32 (30.5 no dual issue) and 6950xt only 24.
If they changed wave slot amount or simd count per wgp fixed dual issue or got rid of it ?
Then the leaked 7900xtx(61 at 355w)>>7900xt(51 at 315w)>N48 (50 at 215W)>7900gre(46 at 260w) Does not look so bad does it ?
50FP32 is 4080-4080super of real power territory.
Worst case scenario if they did not change, a thing 7900xt 215W with rtx4000 rt performance levels
Best case scenario compute slower then 7900xtx but as fast as it in gaming at 215w ?

adroc_thurston · Apr 29, 2024

gaav87 said:
Well it is a problem when 7900xtx has 61TFLOPs of FP32

It doesn't.

gaav87 said:
If they changed wave slot amount or simd count per wgp fixed dual issue or got rid of it ?

That's not the point.
It's just a very opportunistic throughput hack (for w64, VOPD w32 is memey).

gaav87 said:
Then the leaked 7900xtx(61 at 355w)>>7900xt(51 at 315w)>N48 (50 at 215W)>7900gre(46 at 260w) Does not look so bad does it ?

FLOPs don't matter.
Just do clock by shader core count, designs are very convergent anyway.

gaav87 said:
Best case scenario compute slower then 7900xtx but as fast as it in gaming at 215w ?

Not happening.
It's not that good.
Still a cool part.

DisEnchantment · Apr 29, 2024

New patches incoming , nothing interesting yet I spotted (few HW bugs as usual)

Unified MES (precursor to one MES per SED?)

AMD X.Org drivers - Patchwork

patchwork.freedesktop.org

GFX12 IP

AMD X.Org drivers - Patchwork

patchwork.freedesktop.org

DVGPR seems not related to increasing the VGPR file, but pushing the content to LDS and re-allocating more VGPR to certain wavefronts whichever are getting executed and popping back the VGPR from LDS when the corresponding wavefront has finished waiting on the barrier. Instead of the usual partitioning of the VGPRs across whatever number of concurrent wavefronts RDNA can handle (10 I think)

adroc_thurston · Apr 29, 2024

DisEnchantment said:
across whatever number of concurrent wavefronts RDNA can handle (10 I think)

8 per SIMD, for 2-onwards.

H T C · Apr 29, 2024

Glo. said:
This is actually hilarious.

RTX 3050 vs RX 6600.

RX 6600 used the same amount of power, had equally bad RT as 3050, had the same amount of VRAM, and it was sold for 40$ less. Oh and in rasterization - 6600 was 30-35% faster.

nvidia GPU still outsold 4:1 AMD GPU.

And was the card launched @ that performance / price ratio, or did it become like that AFTER some time?

Because i was talking about LAUNCH DAY performance / price ratio.

If it was, then i concede and won't talk about it further.

Glo. · Apr 29, 2024

H T C said:
And was the card launched @ that performance / price ratio, or did it become like that AFTER some time?

Because i was talking about LAUNCH DAY performance / price ratio.

If it was, then i concede and won't talk about it further.

Launch day perf/dollar ratio for 3050 vs 6600 was +12% more expensive for 6600 for 35-40% more performance.

Absolute destruction.

The funniest part.

RX 6600 at launch was 2% faster than RTX 3060. Yet it cost 40$ less.

RX 6600 was in a tier above its price. Nobody cared. Because it was AMD product. Not Nvidia.

Timorous · Apr 29, 2024

adroc_thurston said:
3 AIDs.
It's not crossfire.
And it worked for RDNA4.

RDNA5 has aids? Is that why it will cost so much cash?

blckgrffn · Apr 29, 2024

Glo. said:
Launch day perf/dollar ratio for 3050 vs 6600 was +12% more expensive for 6600 for 35-40% more performance.

Absolute destruction.

The funniest part.

RX 6600 at launch was 2% faster than RTX 3060. Yet it cost 40$ less.

RX 6600 was in a tier above its price. Nobody cared. Because it was AMD product. Not Nvidia.

Continues to be a great pick now in the sub $200 space.

adroc_thurston · Apr 29, 2024

Timorous said:
RDNA5 has aids?

yea.

Timorous said:
Is that why it will cost so much cash?

the halos yeah.

Glo. · Apr 29, 2024

blckgrffn said:
Continues to be a great pick now in the sub $200 space.

Best perf/dollar GPU of current offerings.

Discussion RDNA4 + CDNA3 Architectures Thread

Golden Member

Senior member

Diamond Member

Senior member

Platinum Member

Senior member

Platinum Member

Senior member

Member

Member

Platinum Member

Lifer

Platinum Member

Junior Member

Diamond Member

Platinum Member

Junior Member

Platinum Member

Golden Member

Platinum Member

Senior member

Diamond Member

Golden Member

Diamond Member

Platinum Member

Diamond Member