by Nvidia's DLSS 5 neural rendering tech has been applied to hundreds of games (and plenty of non-game content too) using community tools in the week since it officially debuted in NBA 2K27, sparking reactions ranging from thoughtful discussion to dank memes.
The supporter question that came in on the latest edition of our DF Direct Q+A show is firmly in the former camp, with frequent contributor Skyrimfest asking whether rivals like AMD, Intel and ARM needed their own interpretation of neural rendering - or whether they ought to chart a different course, given the questions surrounding DLSS 5's impact on game artists, its performance impact and its general efficacy.
The short answer is that they do need it - and they're likely already working on it, but it needn't necessarily follow the path set by Nvidia with a screen-space solution. In fact, after we filmed the DF Direct segment above, AMD published their work into neural lighting, via GPUOpen - neural rendering at the core, but very different to DLSS 5. So, it's a forgone conclusion that other vendors like Intel and AMD are also at least exploring the technology, if not working on a proper feature for their GPUs. For example, we had discussions with Intel's Tom Petersen four years ago talking about a generative parse on game graphics, so the idea has been around for some time.
The driving force here isn't a desire to copy Nvidia, necessarily, but because each of these firms is responding to the same underlying conditions: hardware is getting more expensive, availability to cutting-edge semi-conductor processing nodes is getting harder and GPU vendors have hit the power and frequency walls. Brute-forcing extra performance is becoming less and less viable. That means that improving game frame-rates and delivering cutting edge rendering technologies will inevitably mean leaning on a machine learning approach
AI-based super sampling, frame generation and denoising have all followed the same pattern, with Nvidia typically blazing the trail with their heavy investment into research and development, while the likes of AMD, Intel and ARM have used those learnings and their own innovations to implement similar technologies - often specialised for specific hardware or made with slightly different philosophies (eg AMD's commitment to more open standards).
When it comes to DLSS 5, it's clear that the currently available iteration still leaves a lot to be desired - it can deliver an overall boost to image quality in ways that are less heavy-handed than initial demos suggested, but performance remains problematic for the visual return and integration requires significant investment to achieve good results. And there will always be limitations with the screen-space based approach.
Even deeper integration into game engines with more inputs would allow for better outcomes, but are currently off the table due to the sheer weight of the rendering task involved - and the difficulty of implementation for developers. It was a similar situation with SSAO back in the day, which had limitations due to working off screen-space information, but provided enough of a boost to be worthwhile - and, of course, smarter solutions fed with more data in the form of RTAO eventually came along to produce better results.
Right now, we're in that transitional phase, and it's clear that there's still quite some runway for Nvidia to improve its own version and other vendors to at least spec out their alternatives. The community interest - both fascination and disgust - has been intense, and it's likely that even Nvidia were unaware how quickly third-party tools would proliferate to run DLSS 5 on different games, on older GeForce cards and even on AMD GPUs.
That explosion of feedback prompted Nvidia to announce upcoming official support for 40-series RTX graphics cards, something they never hinted at before, and it's almost certainly causing discussion internally at rival GPU vendors too.
Where do you see DLSS 5 and neural rendering more broadly going in the future? Let me know in the comments below.