News AMD FSR 4 Redstone path tracing software uses ML2CODE framework to run on any GPU

[marees]

Ray Regeneration is now live (for RDNA 4)

https://twitter.com/x/status/1989001142588641427

 

[coercitiv]

Here we go again, the puddle narrative:

Imagine spotting your opponent’s reflection rippling in a puddle beneath your feet as every ray of light is rendered with stunning precision.

Me: How did you die?
Soldier: Last thing I saw was the reflection of the enemy in a puddle. I was stunned by the precision of the light bounces, I just couldn't detect any noise. That's when I knew it was probably a time accumulation, I was looking half a second into the past. It was a trap. Then the photon stream stopped and everything went dark. Must have been a headshot, or a faulty 12V-2x6 connector.

That tweet smells of LLM salad with a bit of a touch-up from the marketing intern. Who am I kidding, it was probably a senior. Nothing of substance, just good enough to paste and tweet.

 

[Vikv1918]

Any predictions on this before it releases tomorrow? My prediction - 10% average perf uplift in path tracing, 30% uplift in some scenes with lots of foliage

 

[ToTTenTranz]

Any predictions on this before it releases tomorrow? My prediction - 10% average perf uplift in path tracing, 30% uplift in some scenes with lots of foliage

My prediction is everybody will be rooting for FSR4 on RDNA3 dGPUs and iGPUs, but instead we're getting the promise that future games with raytracing will be more performance competitive in RDNA4 GPUs (which still have very little marketshare compared to RDNA3).


My prediction is AMD will unfortunately show how they're failing to read the room.

 

[basix]

Any predictions on this before it releases tomorrow? My prediction - 10% average perf uplift in path tracing, 30% uplift in some scenes with lots of foliage

There are multiple things involved, I assume?

Potential new driver features:

• General performance enhancements for RDNA4
• Shader Model 6.9 support (preliminary, SM 6.9 release is targeted for Q1/2026). SM 6.9 brings SER, OMM and Cooperative Vectors
  • SW based support for SER (preview driver: https://www.amd.com/en/resources/su...-notes/RN-RAD-MS-AGILITY-SDK-25-10-07-01.html)
  • SW based support for OMM
  • Cooperative vector support (preview driver: https://www.amd.com/en/resources/su...-notes/RN-RAD-MS-AGILITY-SDK-25-10-07-01.html)
• Work graph improvements

FSR Redstone release:

• Updated and enhanced SR
  • Hopefully FSR4 for RDNA3 as well
• New and better FG, probably also MFG
• Introduction of RR
• NRC is delayed or not integrated in any available game
  • See SM 6.9 release (expected in Q1/2026)

 

[GodisanAtheist]

It's tomorrow and it looks like Redstone...

is kinda available in Darktide
Kinda available in COD
...
...
...

oh and don't worry it'll be in a lot of stuff *later*.

Well that was a damp squib.

 

[Vikv1918]

Anyone here with an RDNA4 card wanna test out the Radiance caching sample? I'm curious to see how it works https://github.com/GPUOpen-LibrariesAndSDKs/FidelityFX-SDK

 

[Vikv1918]

Sneak peek of Neural Radiance Caching on Redstone. Notice the "Path tracer" ms in bottom right.

Credit to SaltyKimchi on Optiscaler Discord.

 

[ToTTenTranz]

My prediction is everybody will be rooting for FSR4 on RDNA3 dGPUs and iGPUs, but instead we're getting the promise that future games with raytracing will be more performance competitive in RDNA4 GPUs (which still have very little marketshare compared to RDNA3).


My prediction is AMD will unfortunately show how they're failing to read the room.

Would you look at that, I guessed right.

 

[Ranulf]

is kinda available in Darktide

Hmm, with the new class dlc and some of its problems, I wonder how much it is redstone or just Fatshark being lazy/bad devs. I've never bothered with fsr in Darktide since dlss seems to work well enough.

 

[coercitiv]

Would you look at that, I guessed right.

AMD: If we don't ship FSR4 on RDNA3, they will buy more RDNA4.
The room: If they don't ship this for RDNA3, soon they won't ship that for RDNA4.

Kinda' makes sense though, RDNA3 is already 3 years old. Maintenance Stability mode is just around the corner. /s

 

[ToTTenTranz]

AMD: If we don't ship FSR4 on RDNA3, they will buy more RDNA4.

This logic wouldn't be all that terrible, had they not decided to gut all but 2 chips from the RDNA4 lineup.
There's no RDNA4 high-end or top-end dGPU, and no RDNA4 low-end for desktops, and no RDNA4 low-power laptops or handhelds. There's just not enough RDNA4 offer on the market now or in the near future to justify this planned obsolescence route for RDNA3.


If we count all the Phoenix + Hawk + Strix iGPUs, AMD is most probably selling today a lot more RDNA3/3.5 GPUs for gaming machines than RDNA4 ones. And the Steam Machine isn't even out yet. And the whole world knows they can run FSR4 on RDNA3 just fine.


Yet AMD decides to do a whole event where they Osborne all their current and near future APUs.
Just AMD's marketing decisions being as inept as usual. They just can't help themselves screwing their brand.

 

[Vikv1918]

Yet AMD decides to do a whole event where they Osborne all their current and near future APUs.
Just AMD's marketing decisions being as inept as usual. They just can't help themselves screwing their brand.

Maybe Sony/MS has patents and/or corporate interests that prevent the free use of FSR4. After all they haven't opensourced the neural network model at all. Sony claims to have had a hand in the development of FSR4 and MS claims to have been involved in Redstone development.

I mean FSR4 is a pretty big deal, its basically a viable DLLS4 alternative. Making it truly open source means absouletly EVERYONE can use and can get a big leg up in the market. Valve can use it for their hardware and the entire industry hates Valve. Chinese can use it in their domestic GPUs. Big game companies like Epic or EA can make their own alternative and bypass nvidia and AMD entirely.

Im 100% sure theres some industry-level corporate shenanigans involved here and not just "AMD incompetent" which is a lazy and unsatisfying explanation. I mean we know for a fact from the former FSR 2/3 lead developer that AMD corporate deliberately stopped the development and release of major improvements to FSR2/3. Even prevented the release of major fixes that were ready to go. That goes way beyond "incompetence".

We need to understand that the market at that level essentially operates like a cartel where there's a web of mutual interests. Decisions aren't made in a vacuum. So these things don't make any sense if we look at it in a vacuum.

 

[jpiniero]

Im 100% sure theres some industry-level corporate shenanigans involved here and not just "AMD incompetent" which is a lazy and unsatisfying explanation. I mean we know for a fact from the former FSR 2/3 lead developer that AMD corporate deliberately stopped the development and release of major improvements to FSR2/3. Even prevented the release of major fixes that were ready to go. That goes way beyond "incompetence".

Does seem like patents in general could be a big reason.

 

[marees]

I believe eventually an RDNA 3 version will follow

But first focus would be on upgrading PSSR

 

[soresu]

I believe eventually an RDNA 3 version will follow

RDNA3 frame rate benefits are meager with FSR4 ML SR.

The problem will only be compounded by adding more ML features.

 

[itsmydamnation]

RDNA3 frame rate benefits are meager with FSR4 ML SR.

The problem will only be compounded by adding more ML features.

INT FSR as leaked doesnt do WMMA , so there is fair opportunity for performance improvement on RDNA3 , would make it not work on RDNA2.

 

[GodisanAtheist]

Looks like Claire Obscure picked up Redstone in its latest patch...

 

[ToTTenTranz]

INT FSR as leaked doesnt do WMMA , so there is fair opportunity for performance improvement on RDNA3 , would make it not work on RDNA2.

Are we sure about this?

 

[itsmydamnation]

Are we sure about this?

Yes

 

[marees]

AMD hints at officially open-sourcing FSR 4 upscaling and frame generation technology in the wake of accidental release — accidental release may have forced the company's hand​

News
By Bruno Ferreira Contributions from Jake Roach published 6 hours ago
You still have that salt shaker handy, right?

Hot on the heels of one of AMD's David McAfee suggesting that the company might dust off the blueprints for its Zen 3 CPUs for another production run, another exec let out a hint about AMD's future plans with FSR 4. In an interview at CES, AMD's president of GPU Technologies and Chief Software Officer Andrej Zdravković suggested that the FSR 4 Redstone technology might be open-sourced. The statements come in the wake of an accidental release of the FSR 4 code earlier this year.

The hint came during a Q&A that Tom's Hardware attended when Chips and Cheese's George Cozma asked Zdravković if an open-source release was in the cards. The software head responded that the accidental public release in August was unexpected, but that it intends to release the source for the FSR4 library while keeping the core technology closed, so as not to give Nvidia engineers an advantage.

He continued by stating that AMD intends to "work as openly as possible," and was then asked to be specific about FRS 4's open release. Zdravković then said "that's the long-term plan," seemingly corroborating an earlier remark that "open sourcing is in [AMD's heart and mind]".

https://www.tomshardware.com/pc-components/gpus/amd-hints-at-officially-open-sourcing-fsr-4-upscaling-and-frame-generation-technology-in-the-wake-of-accidental-release-accidental-release-may-have-forced-the-companys-hand

 

[Vikv1918]

AMD hints at officially open-sourcing FSR 4 upscaling and frame generation technology in the wake of accidental release — accidental release may have forced the company's hand

But does this mean fully opensourcing the model? That would be quite a move. That means Intel can straight up start using that.

 

[uzzi38]

INT FSR as leaked doesnt do WMMA , so there is fair opportunity for performance improvement on RDNA3 , would make it not work on RDNA2.

Tbh it's not like it would speed things up much even if it were to be implemented - peak throughput is still the same. I wouldn't expect more than low double digit percentage improvements in the best case scenario.

 

[itsmydamnation]

Tbh it's not like it would speed things up much even if it were to be implemented - peak throughput is still the same. I wouldn't expect more than low double digit percentage improvements in the best case scenario.

Yes no throughput increase, but it makes a big difference for register pressure and those registers still have to be shared with other shader code. So I would expect some perf improvement how much depends on what the bottleneck is.