News Valve Enables Experimental Nvidia DLSS Support For DirectX 12 in Proton

I see absolutely no connection between Nvidia's proprietary DLSS and the full Amd setup of the SteamDeck, but sure, it is nice in the way that it helps Linux users get access to Dlss.

On the other hand, it could be a sign Valve might release an Nvidia-based Steamdeck in the near, perhaps?
 
  • Like
Reactions: joesaiditstrue
Assuming they are somehow translating the calls properly, dlss could work in proton, hardware regardless, but it would be much slower than current versions on rtx hardware with tensor cores, and probably slower than the earlier cuda core dlss also due to having to handle the calls properly. Dlss running on tensor cores is completed in just 1-2 ms due to tensor cores sacrificing color accuracy for speed. There's no hardware specific tricks, just low latency and a rtx card lock. Dlss can run on cuda cores/shader cores, and Nvidia's first release of 2.0 ran on cuda cores, but subsequent releases went to tensor cores, which meant instead of 6-10ms latency, you only had 2ms added dlss upscale latency at worst. In a handheld like the steam deck, I don't think dlss would be worth the added latency of translating the calls properly for the amd soc.

More likely, this is something for third party hardware developers who want to make steam decks with Nvidia discrete graphics, like a low wattage 3050/ti mobile gpu.
 
Last edited:
How would DLSS benefit Deck users? DLSS is Nvidia only. Unless it translates DLSS into FSR?
It doesn't translate DLSS to FSR. That's not even a thing, the entire concept of "translating" DLSS to FSR is nonsensical. DLSS and FSR are upscaling technologies and they work completely differently, there's nothing that could be translated. DLSS uses AI and FSR uses set algorithms with no AI, and I mean there would be nothing to translate. Once the image is upscaled then it would just be ready to be painted on the screen, what would there be to "translate?"

Not to mention the fact that an AMD GPU trying to run DLSS would have to run it on the stream processors (cuda core equivalents) which would incur a massive performance penalty, and that would be completely pointless especially considering the fact that, you know, FSR can be used for any game run in Wine/Proton. So why would you translate DLSS into FSR instead of just using FSR, even if it were possible?

But yes, whoever wrote the article seems to have had a massive derp moment when they said this would benefit Steam Deck users. I'm honestly kind of shocked that this got all the way through the writing process without anyone realizing how ridiculous a statement that was.
 
  • Like
Reactions: joesaiditstrue
How does a Toms Hardware editor not know that DLSS doesn't and will never work on the Steam Deck? like how do you write this stuff and one of your colleagues doesn't pull you aside and say, "Uhhh look buddy, DLSS is proprietary Nvidia tech. Steam Deck is running AMD hardware, you need to edit this article." 😬😒
 
It doesn't translate DLSS to FSR. That's not even a thing, the entire concept of "translating" DLSS to FSR is nonsensical. DLSS and FSR are upscaling technologies and they work completely differently, there's nothing that could be translated. DLSS uses AI and FSR uses set algorithms with no AI, and I mean there would be nothing to translate. Once the image is upscaled then it would just be ready to be painted on the screen, what would there be to "translate?"
It's true that translating DLSS to FSR is impossible and nonsensical.

What is possible, however, is translating calls to DLSS into calls to FSR. In other words, instead of sending the rendered frame to a DLSS API that doesn't exist, it would instead be forwarded to FSR for upscaling.
 
  • Like
Reactions: hushnecampus
It's true that translating DLSS to FSR is impossible and nonsensical.

What is possible, however, is translating calls to DLSS into calls to FSR. In other words, instead of sending the rendered frame to a DLSS API that doesn't exist, it would instead be forwarded to FSR for upscaling.

It's not impossible at all. It's been shown that only versions after 2.0 use tensor cores, an Nvidia specific feature. 2.0 and earlier versions are running on cuda cores (which have been successfully emulated on amd shader/compute units before, and there's various existing libraries to run cuda code on amd) However the cost of running it on amd, especially in a low wattage chip like the steam deck, doesn't seem worth it. Any latency lost via dlss would be gained back and potentially even increased. While you could just swap out dlss for fsr, I doubt that's why this is being added. Numerous third party OEMs have shown interest in making their own steam deck hardware. Supporting dedicated rtx 3050/3050 ti feature sets makes sense as with a bit more weight/bulk and a lot less battery, you could put a dedicated low wattage rtx entry mobile gpu in this device.