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

[adroc_thurston]

Just need to aim for 240p internal res with 10 fps target.

132p internal upscaled to 800p with FSR 3.1 to fit the other GEMM bits.

 

[itsmydamnation]

132p internal upscaled to 800p with FSR 3.1 to fit the other GEMM bits.

upgrade the screen to a crt will all be good.

 

[basix]

As I said, NRC and RR are probably too heavy for PS5 (and everything slower). But it should run decently enough on PS5 Pro.
Super Resolution could run well enough on base PS5 to make sense. The quality uplift from FSR2/3/TSR is easily worth it, even when lowering the rendering resolution a little to compensate for the performance hit of FSR4.

MegaLights and UE5 inbuilt denoising as pathtracing proxy are another thing. That was already shown on PS5. Faster systems like PS5 Pro will just look a bit more beautiful because of NRC and RR.

 

[Heartbreaker]

HWUB tests FSR4 Int8 (from the source code slip up):

It's not quite identical to the standard FSR4 FP8 mode, but it's VERY close. It is slower than FSR3, so you may have to drop to lower quality modes, but even it's performance mode is better than FSR3 or XESS quality.

Definitely use it, and AMD should definitely release it officially.

 

[marees]

HWUB tests FSR4 Int8 (from the source code slip up):

It's not quite identical to the standard FSR4 FP8 mode, but it's VERY close. It is slower than FSR3, so you may have to drop to lower quality modes, but even it's performance mode is better than FSR3 or XESS quality.

Definitely use it, and AMD should definitely release it officially.

FSR 4 int8 (for rdna 3 & 2) is better than xess is the conclusion

Minor pixel quality wins of xess loses to overall stability of FSR 4 int8

https://twitter.com/x/status/1973717348990521433

 

[regen1]

FSR 4 int8 (for rdna 3 & 2) is better than xess is the conclusion

Minor pixel quality wins of xess loses to overall stability of FSR 4 int8

https://twitter.com/x/status/1973717348990521433

FSR4 is good especially the FP8 version.
Since it was tested on Radeon cards those were compared to XESS DP4A version. Would like to see an IQ comparison with XESS XMX version as well.

 

[DAPUNISHER]

XESS XMX version as well.

I would appreciate that for science. No other reason to do it. Being as all 3 vendors now have upscalers that work best with their own proprietary hardware, the relevant comparison is transformer vs FP8 vs XMX. To date: they consensus rank in that order in almost every game tested.

Another consensus is AMD needs to officially support INT8. It will shut down a lot of criticism and reduce the sodium levels of some RDNA 2&3 owners. Mostly the ones with more powerful cards that can provide acceptable fps using the algorithm. I guess if you are on something like a 6650XT, a 10-20% bump in some games while avoiding TAA or older FSR is a win too?

 

[Heartbreaker]

Would like to see an IQ comparison with XESS XMX version as well.

Since the source was leaked can they run FSR4 INT8 on Intel GPUs?

 

[basix]

Yes, it does work in Intel Arc:

https://www.reddit.com/r/IntelArc/comments/1nj84n4/how_to_use_fsr4_in_gta_5_enhanced_story_mode

But not sure if you should better use Optiscaler for such stuff.

 

[regen1]

FSR 4 auf RDNA 2 & 3 im Test

Per „Hack“ ist FSR 4 auch auf Grafikkarten mit RDNA 2 (RX 6000) und RDNA 3 (RX 7000) lauffähig. Qualität und Performance im Test.

www.computerbase.de

 

[marees]

Digital Foundry video

My Takeaways:

I listened to it now. So let me describe the conclusion (from the last few minutes of the video). Let me know, if I got this right ?

• PS5 pro
  • This is the prime focus right now, to get FSR4 int8 work on PSSR on PS5 pro (a version of RDNA 3)
  • Sony & AMD have publicly committed to Q4 this year, but no firm date yet
• Base PS5
  • Doesn't support int 8, because it is RDNA 1 ?? What ???
• RDNA 3 cards
  • No issues in supporting FSR4 int8
  • Slight performance hit. Better than XeSS when normalized
  • Not an immediate focus as AMD might prefer you to "upgrade" to RDNA 4
  • Might be released along with support for RDNA 3.5 handhelds — such as RoG xbox Ally X (see below)
• RDNA 3.5 APUs
  • Quality is good but performance all across the map & varies by game
  • AMD expresses public interest to bring FSR4 to RDA 3.5 handhelds. Next gen zen 6 monolithic APus in 2026-2027 will also share same tech
  • Might release some time in H1 next year — after PSSR release this quarter
• RDNA 2 cards
  • Significant hit in performance even though quality is good
  • AMD likely will not publicly/officially release FSR4 for RDNA because of below reasons
    • Bad headlines due to performance hit
    • XeSS already available?
    • AMD would like RDNA 2 users to "upgrade" to RDNA 4
  • Small chance that this could come very late, if it also comes to current gen Xbox (see below)
• Xbox series X
  • Will benefit significantly as this is RDNA 2 unlike RDNA 1.5 based PS5
  • AMD (& Sony) have made no public statements of bringing FSR 4 to xbox (especially series X)
  • This will make PS5 look very bad in comparison — hence may never come
  • In case it comes to xbox then it might come to RDNA 2 cards also. Probably in 2027, if that ever happens
• Steam Deck
  • Probably limited scenarios where it is useful, hence not worth the trouble ??

So there you go. First PS5 pro then RDNA 3 GPU & RDNA 3.5 handhelds. Maybe xbox & RDNA 2 GPUs. Very unlikely for steam deck & RDNA 2 handhelds

 

[ToTTenTranz]

Base PS5

• Doesn't support int 8, because it is RDNA 1 ?? What ???

IIRC the massive AMD github leak that had Orbis specs and ISA didn't list support for DP4a (4x INT8 throughput from a single FP32 ALU), leading people to believe the PS5's GPU is a RDNA1+2 hybrid.
Digital Foundry at the time picked this up and has since stuck to their theory that the PS5 lacks a bunch of RDNA2 features (mostly based on Microsoft told us so). Problem is they also assumed other stuff in the Orbis leak, like the PS5's GPU only clocking up to 1.8GHz making it a 8.3 TFLOPs that wouldn't be able to keep up with the Series X... and they were dead wrong.



What's even weirder is that DP4a is only absent in Navi 10 (5700XT). Navi 12 and 14 do support DP4a as referenced in the whitepaper. So, either:

- The PS5 does support DP4a but DF just kept believing their own FUD.

- The PS5 GPU has the clock speed optimizations and raytracing-on-TMUs of RDNA2, but is somehow lacking a feature from RDNA1 chips.

- Navi 12 and Navi 14 are actually more advanced than PS5's GPU, the hardware has raytracing and everything else on the PS5 plus DP4a, but AMD just decided to sandbag them (lol).

 

[coercitiv]

AMD would like RDNA 2 users to "upgrade" to RDNA 4

Makes no sense. If the performance hit is significant then users would be highly incentivized to upgrade, FSR4 on RDNA2 would be the perfect commercial for RDNA4+. Moreover, the press coverage would not reflect badly on AMD, it would actually praise AMD for releasing it and then tell consumers to buy RDNA4+ instead.

The only reason to not release it for RDNA2 is having the engineering teams spread too thin, which is likely the case considering AMD wants FSR improvements yesterday.

 

[ToTTenTranz]

Makes no sense. If the performance hit is significant then users would be highly incentivized to upgrade, FSR4 on RDNA2 would be the perfect commercial for RDNA4+. Moreover, the press coverage would not reflect badly on AMD, it would actually praise AMD for releasing it and then tell consumers to buy RDNA4+ instead.

The only reason to not release it for RDNA2 is having the engineering teams spread too thin, which is likely the case considering AMD wants FSR improvements yesterday.

Besides, if AMD really wanted RDNA2 users to upgrade to RDNA4 then they should have provided a viable upgrade path for all ranges.


After almost 20 years on Radeon GPUs on my main desktop, I was forced to go with a Nvidia GPU this time around because I sure as hell wasn't going to upgrade from a 6900XT to a 9070XT. Getting a ~30% performance upgrade after 4 years would've made no sense.

 

[adroc_thurston]

then they should have provided a viable upgrade path for all ranges.

those parts got KIA'd my friend

 

[ToTTenTranz]

those parts got KIA'd my friend

Therefore, no money from me. 🤷‍♂️

 

[SolidQ]

Fabio interviewing AMD's Adam Kozak

 

[marees]

FSR 4 on RDNA 2 & 3 in review: Performance and image quality analyses INT8 vs. FP8

Wolfgang has it working in 14 of 18 games tested on RDNA 2&3. https://www.computerbase.de/artikel/grafikkarten/fsr-4-rdna-2-rdna-3.94512/

His conclusion -

FSR 4 auf RDNA 2 & 3 im Test

Per „Hack“ ist FSR 4 auch auf Grafikkarten mit RDNA 2 (RX 6000) und RDNA 3 (RX 7000) lauffähig. Qualität und Performance im Test.

www-computerbase-de.translate.goog

 

[Vikv1918]

Is it true that there will be a preview driver of redstone this week? Heard a rumor on Reddit about this

 

[marees]

The way this is handled doesn't leave me hopeful of FSR 4 INT8 or Redstone coming to RDNA 2

https://twitter.com/x/status/1985093561533268169

But this is all self-inflicted damage by AMD
Nvidia has more software engineers than hardware engineers

It seems AMD not ready to make that commitment/leap to software. Maybe they are not convinced of the power/Magic of software ??

AMD is Becoming a Software Company. Here's the Plan

Just a few weeks ago, AMD invited us to Barcelona as part of a roundtable, to share their vision for the future of the company, and to get our feedback. On site, were prominent AMD leadership, including Phil Guido, Executive Vice President & Chief Commercial Officer and Jack Huynh, Senior VP &...

www.techpowerup.com

 

[GaiaHunter]

It seems AMD not ready to make that commitment/leap to software. Maybe they are not convinced of the power/Magic of software ??

"AMD finewine" is likely a combination of AMD being much less aggressive with game devolopment, which leads to them being able to squeeze a bit more performance with drivers than Nvidia (which has almost max performance in most new games) and more hardware (mostly vram) than the comparable NVIDIA cards, which allows them to brute force future games.

 

[soresu]

It seems AMD not ready to make that commitment/leap to software. Maybe they are not convinced of the power/Magic of software ??

Ignore the short term PR facepalm fests that AMD are (really) so good at.

Concentrate on what they have been doing overall.

It seems like they have really been working hard on ROCm/HIP, and little by little it has begun to pay off for them in just how much faster that they can support new model releases - to the tune of literally same day in some cases, if not at least the same week.

HIP finally scored a commercial renderer with V-Ray with supposedly significant support from AMD directly.

(it's not clear whether this is just regular HIP or HIP RT btw)

Rome was not built in a day as they say.

You can't take prioritisation of newer vs older gfx µArch driver code just for games as an overall sign of their software strategy when it represents only a limited subset of their overall software efforts.

Efforts which are not limited to driver code or software frameworks.

It also includes their research side, both on their own and working with academic institutions as published on the GPUOpen site.

It also includes their Amethyst joint venture with Sony, and perhaps also some work with MS if they are feeling left out with AMD/Sony's increasing closeness, and the significant gap of gaming performance on Windows vs Linux even when using DXVK, which really isn't making MS look too great at the moment - case in point Linux just broke 3% overall in the Steam survey for the first time, just 13 months after hitting 2% in October last year.

It should also be noted that AMD shuttered their own AMDVLK driver efforts earlier this year in favor of concentrating on RADV, which is clearly killing it in raster gaming, if not yet so well in RT/PT.

That will cut unnecessarily duplicated work.

 

[Vikv1918]

Day 170 waiting for any news about this. I really wish this company would stop announcing things so far in advance

 

[marees]

AMD Driver Confusion: We Still Have Unanswered Questions About RDNA 2's Future​

Will FSR 4 and FSR Redstone support RDNA 2? We still don't know.

We also asked about whether RDNA 2 owners can expect future support in other aspects. We asked about upscaling technologies like FSR 4, an upscaling system whose early, leaked version already functions on RDNA 2 GPUs. Additionally, and arguably more crucial for the average PC gamer, is the question of whether this RDNA 2 driver branch will support DirectX shader model updates, which developers rely on to implement new coding functions or translate emerging lower-level functionality that has fueled console versions of their games.

AMD reps did not answer these questions. And AMD's lack of commitment to both issues is worrying, if not telling. Having tested FSR 4 on RDNA 2 GPUs, we can confirm it may be a matter of taste in both fidelity and performance on older GPUs, but it works, so lack of official support for it would be disappointing. The lack of a commitment to FSR Redstone support for RDNA 2 is also testing.

 

[soresu]

3 new papers just dropped on GPUOpen:

Geometric Integration for Neural Control Variates

Abstract:

Control variates are a variance-reduction technique for Monte Carlo integration. The principle involves approximating the integrand by a function that can be analytically integrated, and integrating using the Monte Carlo method only the residual difference between the integrand and the approximation, to obtain an unbiased estimate. Neural networks are universal approximators that could potentially be used as a control variate. However, the challenge lies in the analytic integration, which is not possible in general. In this manuscript, we study one of the simplest neural network models, the multilayered perceptron (MLP) with continuous piecewise linear activation functions, and its possible analytic integration. We propose an integration method based on integration domain subdivision, employing techniques from computational geometry to solve this problem in 2D. We demonstrate that an MLP can be used as a control variate in combination with our integration method, showing applications in the light transport simulation.

Cardioid Caustics Generation with Conditional Diffusion Models

Abstract:

Despite the latest advances in generative neural techniques for producing photorealistic images, they lack generation of multi-bounce, high-frequency lighting effect like caustics. In this work, we tackle the problem of generating cardioid-shaped reflective caustics using diffusion-based generative models. We approach this problem as conditional image generation using a diffusion-based model conditioned with multiple images of geometric, material and illumination information as well as light property. We introduce a framework to fine-tune a pre-trained diffusion model and present results with visually plausible caustics.

Joint Denoising and Upscaling via Multi-branch and Multi-scale Feature Network

Abstract

Deep learning-based denoising and upscaling techniques have emerged to enhance framerates for real-time rendering. A single neural network for joint denoising and upscaling offers the advantage of sharing parameters in the feature space, enabling efficient prediction of filter weights for both. However, it is still ongoing research to devise an efficient feature extraction neural network that uses different characteristics in inputs for the two combined problems. We propose a multi-branch, multi-scale feature extraction network for joint neural denoising and upscaling. The proposed multi-branch U-Net architecture is lightweight and effectively accounts for different characteristics in noisy color and noise-free aliased auxiliary buffers. Our technique produces superior quality denoising in a target resolution (4K), given noisy 1spp Monte Carlo renderings and auxiliary buffers in a low resolution (1080p), compared to the state-of-the-art methods.