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

 

[insertcarehere]

Based upon NV's benchmarks the 9070 XT should be easily faster than the 5070 in raster. Close but not quite to the 5070 Ti.

Nothing realistic I've seen would have a 9070 XT being 'easily faster' than +30% vs a 4070..

 

[gdansk]

Based upon NV's benchmarks the 9070 XT should be easily faster than the 5070 in raster. Close but not quite to the 5070 Ti.

Wow. If that's true then it's going to be really sad for me (as a raster man) watching RDNA4 flop this generation.

 

[gdansk]

Nothing realistic I've seen would have a 9070 XT being 'easily faster' than +30% vs a 4070..

Is Far Cry 6 (RT) raster?
But yeah it should be about 1.2 - 1.25x perf of the 4070 based on some maths.

 

[jpiniero]

Nothing realistic I've seen would have a 9070 XT being 'easily faster' than +30% vs a 4070..

That's RT, not raster. RT got some arch improvement it looks.

From Compubase's review, 4070 Super by itself is 18% faster in RT than 4070 NS.

 

[marees]

5070 already has the sticker price. AMD can dump n48 out ~tomorrow.

Do we know the RT performance of 5070 ? I believe that should drive 9070xt price

 

[adroc_thurston]

Do we know the RT performance of 5070 ?

that's not relevant.

I believe that should drive 9070xt price

no, raster and upscaling results.

 

[SolidQ]

Nothing realistic I've seen would have a 9070 XT being 'easily faster' than +30% vs a 4070..

Ok, if AMD hint us 7900XT for real, than here

 

[insertcarehere]

Is Far Cry 6 (RT) raster?
But yeah it should be about 1.2 - 1.25x perf of the 4070 based on some maths.

That's RT, not raster. RT got some arch improvement it looks.

... Also based upon Nvidia's materials, which you implicitly trust are true but not when they list performance slides.

Expecting NV or whomever to cherry-pick is to be expected but I sure hope the same scrutiny is given to AMD when they show theirs. Afterall, these choice RTG marketing slides are still from when AMD launched N31, none of which AMD ever got close to at launch.

 

[Keller_TT]

Do we know the RT performance of 5070 ? I believe that should drive 9070xt price

From the RTX 5070 page on full specs detail, it is 40% more RT TFLOPs over 4070, and should be ~ 4070 TiSu with identical memory bandwidth.
So, the Q is how much pure raster gains have NV made in this gen? For it to be 4070 TiSu, 48 Blackwell SM units need to match 66 Ada SM -> 37.5% improvement.
So, for AMD to match a 7900 XT in raster, they need a 5700XT -> 6700XT type improvement from N32 to N48 and not another RDOA3 flop show.

 

[gdansk]

for AMD to match a 7900 XT in raster

Not happening. The Radeon team would not be banished backstage if they had delivered N48 at 3.5GHz.

 

[Win2012R2]

Seems no matrix is real problem why they cancel RDNA4 top.

They cancelled top to use that limited chiplet capacity to make MI3xx that sells for 10x the money

 

[adroc_thurston]

They cancelled top to use that limited chiplet capacity to make MI3xx that sells for 10x the money

no, the reason is far more dumb. trust me.
I'm seething like a Rot Fly seeing 5090.

 

[coercitiv]

The Radeon team would not be banished backstage if they had delivered N48 at 3.5GHz.

AMD is trying to convince the press that RDNA4 got pulled back because the time allocation they had for it at CES, being squeezed in 45 min with many other product offerings would not have done it justice. They claim to have pulled it back in favor of a bigger/better launch event.

I buy none of that. They had the press ready for briefings (including the pre-briefing), and then they went poof. In this day and age they could have launched the product on YouTube, all they had to do at CES was quote some rough performance numbers and offer a competitive price sticker. The press would have echoed all the details much better than any CES presentation could.

They cancelled top to use that limited chiplet capacity to make MI3xx that sells for 10x the money

Nope, at best they cancelled because they thought they would not sell enough of them, which is the opposite of fighting for limited capacity.

 

[Win2012R2]

Nope, at best they cancelled because they thought they would not sell enough of them, which is the opposite of fighting for limited capacity.

I am sure they were also super concerned that they won't sell enough and maybe that was good enough reason on it's own, but when you have far overriding reason (we'll make more money from same silicon like 10x more) then product becomes dead in the water, either way (even if it went ahead) it would have been made in tiny quantities with capacity used for MI300, still they should have gone for it.

N3E GCD and they would have nailed it.

 

[SolidQ]

the reason is far more dumb.

can't be even worse. Raja style?

 

[adroc_thurston]

Nope, at best they cancelled because they thought they would not sell enough of them, which is the opposite of fighting for limited capacity.

it's actually funnier. they killed it because they thought GB202 would make it uncompetitive.

N3E GCD and they would have nailed it.

THEY KILLED THE CHUNGUS BECAUSE THEY THOUGHT THEY COULD NOT COMPETE
aaaaaaaaaaaaaaaa

can't be even worse. Raja style?

worse. they offed it because they thought gb202 gonna be a massive massive massive uplift, being the first NV 512b part since GT200.

 

[Win2012R2]

THEY KILLED THE CHUNGUS BECAUSE THEY THOUGHT THEY COULD NOT COMPETE
aaaaaaaaaaaaaaaa

 

[SolidQ]

THEY KILLED THE CHUNGUS BECAUSE THEY THOUGHT THEY COULD NOT COMPETE

interesting why always in amd someone work like a spy?
They need a strong leader who will say, "We're doing it this way," while banging their fist on the table.
only question why they even miss chance make at least 4090+10%, that not far from XTX.

 

[adroc_thurston]

yeah.

interesting why always in amd someone work like a spy?
They need a strong leader who will say, "We're doing it this way," while banging their fist on the table.
only question why they even miss chance make at least 4090+10%, that not far from XTX.

well it would be 384b (3 base times 128b) versus the 512b chungus with new uarch and chungus die so they just kinda wrote their chances off. too bad!

 

[Win2012R2]

It does not seem like 512 bit on 5090 is helping that much in raw perf, perhaps with 384 it would have been back to 4090 perf

 

[reaperrr3]

Not happening. The Radeon team would not be banished backstage if they had delivered N48 at 3.5GHz.

Why would they need ~60% more oomph than a 78XT to match a 79XT?

Even if the CU-IPC-improvement was zero, just the 4 additional CUs and 33% more SE+ROPs would improve per-clock perf enough that the only reason to not hit 79XT perf would be bandwidth, not GPU clock.
3.0GHz would be enough to catch up in raster, as long as the bandwidth is sufficient.
Which it might not be, but that's a different matter.

it's actually funnier. they killed it because they thought GB202 would make it uncompetitive.

THEY KILLED THE CHUNGUS BECAUSE THEY THOUGHT THEY COULD NOT COMPETE
aaaaaaaaaaaaaaaa

worse. they offed it because they thought gb202 gonna be a massive massive massive uplift, being the first NV 512b part since GT200.

You mean Nvidia's propaganda successfully manipulated AMD leadership into self-deterring themselves from showing up to a fight they actually would've won surprisingly easily if they did?

If true: Damn.

 

[SolidQ]

and chungus die so they just kinda wrote their chances off.

Seems some analytic so bad, need some determined guys, a new guys.
Like: Gamer maybe not going buy much? We can make RocM and sell like cupcakes as AI card. Overall they not thinking in right way, of they would beat RTX 5090, with N4C gamers gonna buy it

 

[adroc_thurston]

You mean Nvidia's propaganda successfully manipulated AMD leadership into self-deterring themselves from showing up to a fight they actually would've won surprisingly easily if they did?

No, they just modelled a huge uplift based off 50% more SM, new Uarch, GDDR7 and 33% moar memory width.

Which it might not be, but that's a different matter.

RDNA4 new caches are neat, actually.

Seems some analytic so bad, need some determined guys, a new guys.

NV is good at shipping big die. GB202 was the chunguests, meanest thing they've shipped since arguably GT200 (their last 512b offering).
Plus, remember, they got stomped by Ada being 144SM on the top-end.

 

[gdansk]

Why would they need ~60% more oomph than a 78XT to match a 79XT?

It isn't. It's 44% more oomph to deliver 30% more performance. Many such cases.

 

[SolidQ]

33% moar memory width

that was original 33%? Because currently seeing 1tb vs 1.8tb