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]

technically even slightly worse

If it's worse in all aspects, why would Nvidia stick with it? It's not like they're short on cash.

 

[adroc_thurston]

Ryzen moment happened mainly because Intel stayed on 14nm for so long, people forget that all the time

No it happened because 3950X.

 

[Meteor Late]

No it happened because 3950X.

Yeah 7nm vs 14nm. Also it's Intel, one of the less innovative company out there.

 

[exquisitechar]

Yeah, hopefully they do a transformer model. RDNA4 should support future upscaling models. I don't think they would switch to dedicated tensor cores after putting all the effort into beefing up the CUs

I think AMD said that FSR4's model already takes from both CNNs and transformers. There's room for improvement with future updates for sure, I hope they inch closer to DLSS4.

 

[adroc_thurston]

Yeah 7nm vs 14nm. Also it's Intel, one of the less innovative company out there.

Oh no that's not the issue.
They shipped twice the cores and that was the story.

 

[gdansk]

I'd like to see how RTX 5080 does when stuck with 20Gbps memory.
And maybe some memory OC on RX 9070 XT to meet in the middle.

 

[GTracing]

I'd like to see how RTX 5080 does when stuck with 20Gbps memory.
And maybe some memory OC on RX 9070 XT to meet in the middle.

Yeah, I've been hoping to see that as well. 5080 vs 9070 XT with the same power limit and memory bandwidth. Though the comparison would be slightly unfair to the 5080 since it was likely designed to use all that bandwidth.

 

[lixlax]

So sometimes it beats the XTX and even matches the 5080, sometimes it loses to the 7900XT.
GDDR7 and a bit more die area for shaders and it would've been a far better product at around $800.
They cheaped out a little bit too much to make it a clean value kill.

Given the lackluster memory BW they gave it I think RDNA 4 turned out to be faster than anticipated. 320bit with 20gbps bus would've helped a lot it seems and +4GB of vram would've been a nice selling point as well.

 

[adroc_thurston]

Given the lackluster memory BW they gave it I think RDNA 4 turned out to be faster than anticipated. 320bit with 20gbps bus would've helped a lot it seems and +4GB of vram would've been a nice selling point as well.

it just needs GDDR7.

 

[gdansk]

Though the comparison would be slightly unfair to the 5080 since it was likely designed to use all that bandwidth.

I'm interested as technical comparison, not a fair comparison or buying advice in anyway.
And unfortunately it's the closest we can get to equal setup as AMD probably doesn't have GDDR7 enabled memory controller or at least not validated.

 

[gdansk]

I have to agree they made a mistake on the 9070 price.
AMD thinks they know what they're doing but it won't work. You're not pushing them up to the XT, you're pushing them to the 5070. Not having a cheaper card is a mistake.

 

[GTracing]

I'm interested as technical comparison, not a fair comparison or buying advice in anyway.

Right, but even if you're just comparing them on an architectural basis it's not exactly fair. If all you're trying to do is see how bandwidth-efficient each architecture is, then it's fine. But if you try and draw conclusions about performance/area or performance/watt, then it's no longer apples to apples. Nvidia would have given it less cores if they were going to saddle it with 20Gbps memory. They likely would have decreased the cache as well. And less cores means less performance at a given power limit.

 

[adroc_thurston]

Nvidia would have given it less cores if they were going to saddle it with 20Gbps memory

They didn't give Blackwell any more shaders relative to Ada.
The perf gain is just GDDR7.

 

[PJVol]

thing is choking on memory mostly

Is it really?

To me it seems like some games might probably be not as friendly to the new rdna4 mem-cache optimizations.

 

[adroc_thurston]

To me it seems like some games are probably not as friendly to the new rdna4 mem-cache optimizations.

Which is where you need the stuff.

 

[GTracing]

They didn't give Blackwell any more shaders relative to Ada.
The perf gain is just GDDR7.

Even the 4080 super has 23Gbps memory though. And the 5080 does have 5% more cores than the 4080 super.

 

[branch_suggestion]

Is it really?
View attachment 118967
To me it seems like some games are probably not as friendly to the new rdna4 mem-cache optimizations.

That is a memoc too.
The raw perf difference between the XT and vanilla should be larger than it is, memory is the equaliser here.

 

[gdansk]

If all you're trying to do is see how bandwidth-efficient each architecture is, then it's fine. But if you try and draw conclusions about performance/area or performance/watt, then it's no longer apples to apples.

And if AMD knew they'd be making a GB203 competitor they'd have chosen different memory, etc.
Unfortunately, there is no apples to apples. But there are people out there with both cards who could try to make it close.

 

[adroc_thurston]

And the 5080 does have 5% more cores than the 4080.

Super added 2 SMs. It's not a real delta. Be serious.

 

[coercitiv]

Is it really?

You're missing one data point in that chart, one where the 9070 XT is only using mem oc.

 

[GTracing]

And if AMD knew they'd be making a GB203 competitor they'd have chosen different memory, etc.
Unfortunately, there is no apples to apples. But there are people out there with both cards who could try to make it close.

Yeah, I agree. That's basically my point. I was just saying that a comparison like that would favor the 9070 XT over the 5080.

 

[SolidQ]

They just need make GDRR7 version

 

[gdansk]

Yeah, I agree. That's basically my point. I was just saying that a comparison like that would favor the 9070 XT over the 5080.

I'm not convinced. Nvidia has a bigger die, meeting N48 halfway with slow memory is balanced approach.

 

[PJVol]

Which is where you need the stuff.

Maybe cost wouldn't justify the outcome?

The raw perf difference between the XT and vanilla should be larger than it is, memory is the equaliser here.

Not that it's a strong representative, but take note of XT -> XT-OC, the scaling still holds while the output is 4x more pixels.

 

[blckgrffn]

I have to agree they made a mistake on the 9070 price.
AMD thinks they know what they're doing but it won't work. You're not pushing them up to the XT, you're pushing them to the 5070. Not having a cheaper card is a mistake.

Doesn't matter if 5070 is unobtanium or only available at or above 9070XT prices.

We needed Nvidia to deliver to bring us better AMD prices. They let us down.

To quote Toms, of all places:

"And finally, what about the RX 9070 versus the RTX 5070, both nominally priced at $549? If you're at all good at math and were paying attention above, you'll already know that the 9070 comes out ahead. It's 12% faster at 4K, 8% faster at 1440p, and 4%/7% faster at 1080p. There are five games where the 5070 manages any lead at all, with Space Marine 2 being the biggest margin of victory and the only one where the 5070 leads at 4K. In general, though, the RX 9070 is clearly better for rasterization performance at native resolution."



There is zero reason for the 9070 to be cheaper when they will likely A) all sell out anyway and B) GPU prices will likely all rise in a month or three depending on shenanigans.

Would I have preferred a $450 9070? Yes. Would I have likely bought two of those for the coming fun times? Also yes. :/

9070xt or bust come tomorrow now.

And... if I can order a 9070 for my daughters PC and flip her 6700XT for a tidy sum while we all complain about the poor value of it but that means its on the shelves for more than a heartbeat, so much the better, IMO. 6700XT --> 9070 should be a huge jump and its right in that ~200W sweetspot.