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]

It's almost a week since the availability of PRO AI R9700 with SIs and OEMs. Still no 3rd party benchmarks?

You have my permission to be both the 2nd & 3rd party. As one of the few with an interest you'll even know exactly what benchmarks your 2nd party self wants your 3rd party self to run.

 

[mostwanted002]

Scorptec Computers

AMD Radeon AI PRO R9700 at Scorptec.com.au, the online Gaming PC, server computer and technology experts. Fast delivery to Australia.

www.scorptec.com.au

Radeon PRO AI R9700 Pre-orders listed for Australian vendor. roughly USD 1300

 

[jpiniero]

AMD quietly announces Radeon RX 9060 alongside new Adrenalin driver — new entry-level GPU is for 'select system integrators only'

A new RX 9000-series graphics card for prebuilts, not builders

www.tomshardware.com

Toms says the 9060 is coming but for "Select System Integrators" only. Ram speed down to 18 gbps and CUs cut to 28.

 

[SolidQ]

Ram speed down to 18 gbps and CUs cut to 28.

i think just as RTX 5060 level, but cheaper

 

[GodisanAtheist]

i think just as RTX 5060 level, but cheaper

- If AMD can pull out a 5060 for ~$230 they'd have a run-away hit on their hands.

Doubt they'd be able to make enough of them and the price would float up to $300 anyway.

 

[marees]

AMD quietly announces Radeon RX 9060 alongside new Adrenalin driver — new entry-level GPU is for 'select system integrators only'

A new RX 9000-series graphics card for prebuilts, not builders

www.tomshardware.com

Toms says the 9060 is coming but for "Select System Integrators" only. Ram speed down to 18 gbps and CUs cut to 28.

AMD page

https://www.amd.com/en/products/graphics/desktops/radeon/9000-series/amd-radeon-rx-9060.html

 

[DAPUNISHER]

AMD page

https://www.amd.com/en/products/graphics/desktops/radeon/9000-series/amd-radeon-rx-9060.html

Come on now; 100 fps in Resident Evil 4 1080 Ultra with RT on an 8GB card? I've never played, but it should blow pass the frame buffer at those settings correct?

 

[Dezii90]

RDNA4 is selling well

 

[SolidQ]

RDNA4 is selling well

Yeah, only Gabe need fix his steam survey, and show RX 9070 series, because 7800XT take 9070 data

 

[Keller_TT]

Nor sure if this will interest many, but I'm definitely excited. Modular, after announcing top-tier support for CDNA 3 with their Max platform and Mojo, have announced that they want to bring it to consumer GPUs too (RDNA 3 & 4) for the hobbyists at home. We're now this close to getting rid of the CUDA swamp.

The catch is that CDNA uses true Matrix cores with MFMA intrinsics, while RDNA does Matrix multiplication through WMMA instructions, which have been overhauled in RDNA4 and not backwards-compatible with RDNA 3 (also a main reason why FSR 4 isn't for RDNA 3). At the moment, they are relying on eager volunteers to help along.
I'm keeping an eye on it, and will offer what little I can do. I'm just genuinely chuffed that I can soon write GPU kernels on my RDNA4 that'll be fun, and not an experience in hell.

 

[igor_kavinski]

The catch is that CDNA uses true Matrix cores with MFMA intrinsics, while RDNA does Matrix multiplication through WMMA instructions, which have been overhauled in RDNA4 and not backwards-compatible with RDNA 3 (also a main reason why FSR 4 isn't for RDNA 3).

You just made me so glad for skipping RDNA3.

 

[mostwanted002]

You just made me so glad for skipping RDNA3.

+1

I particularly use hashcat, and the whole reason for me skipping RDNA3 was no significant improvements over my 6900 XT.

(From my limited analysis, that workload is more INT32/16 focused on the majority of algorithms)

 

[soresu]

Nu v3.0 of HIP RT finally incoming which should hopefully address RDNA4 performance in Cycles...

https://github.com/GPUOpen-LibrariesAndSDKs/HIPRT/commit/592a92b5afcc6de16c5371b1f41daec4ac59b77b

 

[ToTTenTranz]

The catch is that CDNA uses true Matrix cores with MFMA intrinsics, while RDNA does Matrix multiplication through WMMA instructions, which have been overhauled in RDNA4 and not backwards-compatible with RDNA 3 (also a main reason why FSR 4 isn't for RDNA 3). At the moment, they are relying on eager volunteers to help along.

My only problem with this big departure from RDNA3's ML capabilities is the fact that AMD is getting stuck with that architecture in their APUs for another 2 years.

As it stands, FSR4 is useless for all Phoenix, Strix and Medusa APUs, while Intel already has XMX cores in their Lunar Lake and Arrow Lake H SoCs that run Xess 2.

 

[marees]

My only problem with this big departure from RDNA3's ML capabilities is the fact that AMD is getting stuck with that architecture in their APUs for another 2 years.

As it stands, FSR4 is useless for all Phoenix, Strix and Medusa APUs, while Intel already has XMX cores in their Lunar Lake and Arrow Lake H SoCs that run Xess 2.

Microsoft partnership should net better APUs from 2027 onwards, maybe with MFMA too 🤔

 

[DAPUNISHER]

My only problem with this big departure from RDNA3's ML capabilities is the fact that AMD is getting stuck with that architecture in their APUs for another 2 years.

As it stands, FSR4 is useless for all Phoenix, Strix and Medusa APUs, while Intel already has XMX cores in their Lunar Lake and Arrow Lake H SoCs that run Xess 2.

Man, this is one of those situations where as a gamer, my criteria would have me choosing AMD. Here's why: The Intel software is a weak point. No integrated recording or streaming. Driver issues are more prevalent, and some engines just don't seem to play well with the hardware. The CPU overhead is another weak point that needs to be taken into consideration.

Also, if the APUs are in handhelds or thin and lights with small screens, the DP4a version of XeSS will do a great job on the AMD hardware. I find myself choosing XeSS on AMD cards in almost every game that has it. The performance is usually not significantly lower and the IQ is predominantly better. AMD in Linux is outperforming windows in quite a few games now, sometimes drastically so, with much better 1% lows too.

The Intel offering would have to be substantially faster in both windows and Linux at the same price to win me over.