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

 

[soresu]

The paper also shows Neural Ambient Occlusion (NAO)

Doesn't necessarily mean Redstone has anything to do with it.

It's a research paper so it may well just be using previous studies in the NR gfx field to demonstrate the efficacy of GATE.

Demonstrating a single NR technique like NRC would be fine if it was purely for PR purposes, but in a research paper using only a single point of reference to demonstrate gains over a previous method is bad science, and probably wouldn't have passed peer review.

 

[soresu]

My kingdom for a FSR Redstone leak

Waiting for Redstone.....

 

[soresu]

Probably the gddr7 version of 9070xt that MLID has been wishing for

More likely stacked or 2x density (16 -> 32 gbit) GDDR6.

 

[coercitiv]

HUB tests driver overhead on RDNA4 vs. Blackwell:

Something to consider for people considering a GPU upgrade while keeping their old CPU... such as... myself?! Oh noes, I have an old CPU.

 

[SolidQ]

Oh noes, I have an old CPU

and 7600 not old CPU

 

[JustViewing]

HUB tests driver overhead on RDNA4 vs. Blackwell:

Something to consider for people considering a GPU upgrade while keeping their old CPU... such as... myself?! Oh noes, I have an old CPU.

I would assume most will run max settings as possible for the given GPU, so the overhead might not be an issue. But seeing half the power usage for 5080 did surprise me

 

[coercitiv]

I would assume most will run max settings as possible for the given GPU, so the overhead might not be an issue.

The overhead only becomes an issue for people on older systems who upgrade the GPU. The CPU would have to be even older than the R5 7600, as in this case it takes a high refresh gaming scenario to show a performance delta. IMHO it's not a big deal for most people, just something to keep in mind.

But seeing half the power usage for 5080 did surprise me

I keep telling people they can achieve wonders by limiting FPS to a limit they find acceptable. Even with high refresh monitors and single player games there's often a tradeoff that is well worth in terms of IQ - FPS - Power balance.

 

[Win2012R2]

I keep telling people they can achieve wonders by limiting FPS to a limit they find acceptable.

Why buy stupidly overpriced 5080 then?

 

[coercitiv]

Why buy stupidly overpriced 5080 then?

It has nothing to do with the 5080, you can also do this with most high TDP cards.

 

[DAPUNISHER]

I would assume most will run max settings as possible for the given GPU, so the overhead might not be an issue.

As Steve talks about in the video, max settings ae often the most CPU bound now. Crowds, traffic, and ray tracing, are good examples of the higher the setting, the more CPU intensive it is.

 

[Win2012R2]

It has nothing to do with the 5080, you can also do this with most high TDP cards.

Like 5090?

 

[coercitiv]

Like 5090?

Like my 6800 XT.

 

[soresu]

Another update for the upcoming v3 of HIP RT, hopefully a release sooooon:

Update to 3.0.9ba63f3 · GPUOpen-LibrariesAndSDKs/HIPRT@13a4fd2

Contribute to GPUOpen-LibrariesAndSDKs/HIPRT development by creating an account on GitHub.

github.com

 

[marees]

That was using Lilac board, which is the Rembrandt socket. Valve was likely testing how the Lenovo Legion Go S performed with Proton.

Does the Deckard need a Fremont like device to connect ?


Valve speculated to launch all these 4 devices together

1. Fremont — TV console
   1. a TV focused PC box/console running SteamOS.
   2. Has dedicated HDMI port
   3. Has a semi-custom CPU based on Hawk Point 2
   4. But with a removed iGPU. And dedicated RX 7600 GPU (No shared RAM)
   5. Has custom Valve mobo/case
2. Ibex — gamepad for above
3. Deckard — arm powered XR headset
4. Roy — controller for Deckard

https://twitter.com/x/status/1958025002105598107

 

[PJVol]

Is the 9070XT even bandwidth limited?

I skipped RDNA3, but compared to RDNA2 I see a lot more scenarios where the bandwidth limitation becomes apparent.

 

[soresu]

YUSSSSSSSSSSSSSSSSSS!

I was right on the money, HIP RT v3 just released, hopefully it makes it into Blender v5 🎉

Release 3.0.9ba63f3 · GPUOpen-LibrariesAndSDKs/HIPRT

RDNA 4 RT features Compressed BVH8 and triangle packets Intersectable instances Oriented bounding boxes (for HQ build) BVH import API changed Generic BVH2 input format Internally transformed to...

github.com

 

[Saylick]

Looks like AMD is going to go full OSS with FSR4:

https://videocardz.com/newz/amd-releases-fidelityfx-sdk-2-0-fsr-4-goes-full-open-source

Unfortunately, quickly after they announced it, they pulled the GitHub files for FSR4 because they accidentally put files on there that they didn't original intend.

https://twitter.com/x/status/1958209847507480834

 

[igor_kavinski]

Re-uploaded 36 minutes ago: https://github.com/GPUOpen-LibrariesAndSDKs/FidelityFX-SDK/releases

 

[soresu]

Looks like AMD is going to go full OSS with FSR4:

I always assumed that as a given.

Not that it will do much good for those who lack the hw to make use of it performantly.

Will be interesting to see if ARM churn out a mobile variant like with ASR, which is basically FSR2 with an extra backend for OGL ES.

 

[basix]

There is already an ARM variant of DNN based "neural super sampling", presented 1 week ago: NSS which builds on top of ASR

Arm Neural Technology Delivers Smarter, Sharper, More Efficient Mobile Graphics for Developers

Explore Arm Neural Technology, the first to integrate neural accelerators into GPUs, enhancing mobile graphics with AI.

newsroom.arm.com

And there was a Siggraph 2024 presentation, which used FSR2 as basis and extended the algorithm with a DNN based upscaling parameter predictor / estimator:

https://community.arm.com/cfs-file/__key/communityserver-blogs-components-weblogfiles/00-00-00-20-66/siggraph_5F00_mmg_5F00_2024_5F00_mobile_5F00_nss_2D00_LiamONeil_2D00_v7_2D00_speakers_5F00_notes.pdf