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

 

[coercitiv]

9070 pricing is mainly a gift for NVidia.

It's the other way around: Nvidia gifted AMD the higher 9070 pricing. Remember the 5070 FE is an imaginary GPU.

 

[SolidQ]

They still a step behind in RT and AI scaling.

DLSS Transformer is better at texture quality but FSR4 is better with particle details, that to be where they are trading blows.

 

[adroc_thurston]

DLSS Transformer is better at texture quality but FSR4 is better with particle details, that to be where they are trading blows.

I just wanna see execution traces on both.

 

[GaiaHunter]

In the UK, with prices still to be confirmed and only able to order them tomorrow, overclockers https://www.overclockers.co.uk/pc-c...cs-cards/amd-radeon-rx-9070-xt-graphics-cards has the following prices (it seems they are updating models and prices as I write).

Sapphire Pulse Radeon RX 9070 XT £569.99
PowerColor Radeon RX 9070 XT Reaper £569.99
Asrock Radeon RX 9070 XT Steel Legend £569.99
PowerColor Radeon RX 9070 XT Hellhound £659.99
Sapphire Pure Radeon RX 9070 XT Gaming OC £689.99
Sapphire Nitro + Radeon RX 9070 XT Gaming OC £699.99
PowerColor Radeon RX 9070 XT Red Devil £709.99
Gigabyte Aorus Radeon RX 9070 XT Elite £749.99
So the basic models seem decently priced but the better models better have some 10%+ performance increases.
The 9070 seem to start at £529.99 for the reaper and pulse models.
The mid tiers like the Hellhound and Pure are £569.99 and the premiums like the Red Devil and Nitro+ at £599.99

 

[branch_suggestion]

Why are there quite a number of instances where the 9070 XT performs relatively better against other GPUs at 4K than it does at 1440p?

Could be a hint, for reference L2 BW and capacity both massively exceed N31, and MALL BW is similar with 2/3 the bus width.

 

[marees]

Well, isn't price the most important factor though? Next to availability that we will soon see.

Launch Availability is not a issue

However once sold out fresh stocks will be at a new higher MSRP

 

[SolidQ]

Path Tracing seems problem for AMD seems need RNDA5 for that, but in AMD sponsored game PT will be better

 

[GTracing]

DLSS Transformer is better at texture quality but FSR4 is better with particle details, that to be where they are trading blows.

That's a much more optimistic description than Alex gave in the video. He pointed out the aliasing on the confetti as one edge case where FSR4 does better than DLSS transformer. I guess that technically qualifies as trading blows, but overall DLSS transformer is noticeably better than FSR4.

 

[SolidQ]

but overall DLSS transformer is noticeably better than FSR4.

I'd assuming AMD is gonna make own transformer, question for RDNA4 or for RDNA5 with matrix cores

 

[branch_suggestion]

Path Tracing seems problem for AMD seems need RNDA5 for that, but in AMD sponsored game PT will be better

Outside of BM:W and Indy PT it is alright.
Roughly matches 5070 in even some notorious first party PT mods.

 

[gdansk]

The impression really varies by review.
Some even show power consumption on par and others quite a bit higher. Telemetry vs PCAT? No, even some PCAT reviews show it in line with NVidia. I guess some AIB XT models are really cranked. But HWUB has lower performance on average than most other reviews and still has high power consumption numbers.

In any case, I think the price is low enough it should work out provided AMD can have a reasonable quantity of MSRP cards available.

 

[GTracing]

I'd assuming AMD is gonna make own transformer, question for RDNA4 or for RDNA5 with matrix cores

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

 

[GaiaHunter]

I would wager that the low power 9070 XTs wont be around for long and will be replaced by the OC 9070s.

So in a few months you would have a few standard 9070s, a bunch of 9070 OC and the 9070 XT will be all the high power models.

The OCed 9070 XT seems to match/surpass the 5070 TI.

 

[adroc_thurston]

The OCed 9070 XT seems to match/surpass the 5070 TI.

It runs out of bandwidth all the same.

 

[Meteor Late]

I expected a bit more out of 9070 XT to be honest.
AMD lied about 9070 XT performance relative to 7900 GRE, it is much lower than the performance increase they gave. Then in RT, looks good in some games but terrible in others when using transformative settings. While it is much better than before overall, it's just a lot behind in some games with PT and very high RT settings. Also, power consumption a bit higher than expected.
I think the only expected "good" result is FSR4, definitely looks very good, slightly better than CNN, although still behind DLSS Transformer. Want to see HUB FSR4 vs DLSS Transformer comparison, however.

 

[branch_suggestion]

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.

 

[linkgoron]

Not nearly. Price is really the only win. They still a step behind in RT and AI scaling.

Ryzen didn't beat Intel across the board until much later than Ryzen.

Really RDNA1 was much more of a Ryzen moment. RDNA2 was excellent, it's just that RDNA3 was a dud. RDNA4 is pretty good, just mostly suffering because AMD made the wrong decision of not doing a bigger die.

 

[Panino Manino]

Random thought: I think the power efficiency on the cards tells a lot as to why AMD gave up on the high end. If one 9070 XT is barely able to get close to a 4080, I doubt that chiplets stacked together into 120 CUs would have done great.

Imagine the chips and bigfefs chips running at a lower clock.

 

[branch_suggestion]

Then in RT, looks good in some games but terrible in others when using transformative settings. While it is much better than before overall, it's just a lot behind in some games with PT and very high RT settings.

Games sponsored, developed for and sometimes developed by NV.
Also VulkanRT for Indy which AMD does not properly support, drivers might still help these edge cases but the devs really need the cards to be optimised for.

 

[adroc_thurston]

I expected a bit more out of 9070 XT to be honest.
AMD lied about 9070 XT performance relative to 7900 GRE, it is much lower than the performance increase they gave

They sure as hell didn't.
It's a good product, shame no G7 SKUs to equalize the membw baseline.

 

[MrTeal]

Somehow great and still disappointing. It really would have been better had AMD not pushed out the +40% over GRE claims.

At $200 or more cheaper than a 5070Ti if MSRP stock is good it's a great product though

 

[gdansk]

Also VulkanRT for Indy which AMD does not properly support

Well, that's still AMD's problem.

 

[branch_suggestion]

Random thought: I think the power efficiency on the cards tells a lot as to why AMD gave up on the high end. If one 9070 XT is barely able to get close to a 4080, I doubt that chiplets stacked together into 120 CUs would have done great.

Perf barely goes down with the power dialled back, thing is choking on memory mostly.
Lots of clock headroom without the memory to let it shine.

 

[Mahboi]

Perf barely goes down with the power dialled back, thing is choking on memory mostly.
Lots of clock headroom without the memory to let it shine.

Well then that's just silly.

 

[Meteor Late]

Ryzen didn't beat Intel across the board until much later than Ryzen.

Really RDNA1 was much more of a Ryzen moment. RDNA2 was excellent, it's just that RDNA3 was a dud. RDNA4 is pretty good, just mostly suffering because AMD made the wrong decision of not doing a bigger die.

Ryzen moment happened mainly because Intel stayed on 14nm for so long, people forget that all the time. Let's see Nvidia stay at 4nm (well, technically even slightly worse with 4N custom node) until 2029 and see AMD with "Radeon moment" 2 years from now.