What if someone other than Nvidia, AMD or Intel delivered a suite of AI upscaling, frame generation and ray reconstruction technologies? And what if they're comparable and perhaps, maybe even better in some regards? Apex Compute has just released a beta version of its Apex neural rendering suite and while it's early beta code focusing just on frame generation, there is promise here - and it poses questions on the nature of integrations into future games. In a world where it seems difficult to get FSR 4 and XeSS support into new games, what if third party developers offer compelling tech? Surely we shouldn't need Optiscaler to make it work in new games?
There's something else to consider too. Apex Compute has its own ambitions in producing its own GPUs, but ANR is part of its suite of open source technologies - and not only that, it's developed by just one person within the organisation. To produce anything comparable with DLSS is surely quite the feat.
Right now, the technology is delivered via Optiscaler, which reroutes DLSS inputs into its own DLL. The DLL itself is actually a drop-in replacement for Nvidia's own frame gen DLL, but the developer recommends using Optiscaler for now. The Github provides easy-to-use installers for games that already have Optiscaler installed and for those that do not. A second, supplementary package is also on offer - nr-record - which is fascinating in its own right in delivering frame accurate comparisons between super resolution and frame gen processors.
Let's get straight to the testing though, based on a Ryzen 7 98000X3D system with 6000MT/s DDR5 combined with an RTX 5070 Ti. Watch the video up top and you'll see that first I tested the intro scene to Marvel's Spider-Man. Apex averages at 130.5fps up against DLSS at 124.9 so yes, it is faster on this title at least. Apex has a 7.2fps advantage at the 1% lows, rising to 9.9fps at 0.1% low - but I honestly think 0.1% metrics only really make sense with a very very long benchmark sequence. The important thing to note is that frame-times are broadly consistent between the technologies - I understand from speaking to the developer that Apex ties into Nvidia reflex here, so it's to be expected. PC Latency comes in at 41.1ms with DLSS, 43.1ms with Apex. No problems there then.
Let's move from 2x to 3x frame gen as we look at RT overdrive path tracing in Cyberpunk 2077 with this 9800X3D/RTX 5070 Ti combo. DLSS is around 10 percent faster in this run - a 0.85ms average frame time advantage to DLSS here with a 13% deficit at one percent lows. Because of the performance difference, DLSS has a 3.9ms latency advantage too. I do wonder the extent to which the Optiscaler wrapper may be disadvantageous to Apex but it's certainly the case that based on these conditions, ANR is slower.
One more benchmark and I moved from 3x MFG to 4x MFG in Crimson Desert. 2160p, optimised settings and a 203fps average on DLSS up against the optiscaler-fuelled 195.7fps from Apex - a 3.6 percent deficit but an irrelevant average frame-time difference of just 0.18ms. Remember we are looking at DLSS with machine learning acceleration vs a compute shader here. PC latency - that's impressive. 50.4ms DLSS vs 51.1ms on Apex.
Image quality is interesting and difficult to compare, especially on multi frame generation where the artefacts present very, very differently. Frame generation is most challenging on fast-moving screen-filling elements, where Spider-Man leaping out of his apartment window is a pretty good test. It's difficult to choose a winner here. Artefacts around the webslinger are present with both generation techniques, but arguably less noticeable in motion on the Apex side.
However, ghosting on the weblines is more of an issue with Apex compared to DLSS. Running past these railings, DLSS has less ghosting and remains more consistent across the run. Overall, both techniques have strengths and weaknesses here. I notice that in this title, DLSS interpolates 2D bitmap elements with some failures. Apex looks cleaner, but only because it runs those elements at a lower frame-rate which is not ideal.
Next up, 4x MFG in Crimson Desert at optimised settings with 4K DLSS performance mode with our notoriously heavy benchmark scene. To begin with, initial tests reveal quite similar image quality overall, with similar ringing artefacts around large objects moving quickly in front of the camera. However, once subtitles appear on-screen, Apex now presents without 2D element mitigation, presenting almost like interlacing in motion, while DLSS just looks absolutely fine.
Both generative techniques can have problems with camera cuts, merging imagery from multiple frames rather than producing a clean cut - but Apex seems less capable of spotting those cuts and can take longer to resolve. Overall quality looks fine though.
I concluded with The Last of Us Part 2, 4x MFG again, ultra settings, 4K DLSS performance mode. The presentation is pretty good on both. General motion flow looks fine on both but I notice that on fast character movements close to the camera, artefacts can present quite differently. They're not ideal on both but Apex does seem to look a touch less egregious. This game is one that the developer highlighted for ANR's advantages over DLSS.
I also tried ANR very quickly on non-Nvidia GPUs. It uses a compute shader pipeline so it should run just fine on any modern GPU. The author recommends Nvidia GTX 16-series upwards. To begin with, I tried, Cyberpunk 2077 at 1440p on high settings on an Intel Arc B580 in XeSS Quality mode with ANR 3x MFG. I saw a 186fps average displayed across the run with an average PC latency of 67ms, which is a bit high. I understand that Optiscaler is routing Reflex inputs through XeLL, but regardless, frame health looks consistent.
I also tried a Radeon RX 6700 which would otherwise be relying on FSR 3 frame generation. a 149fps displayed framerate average, latency even higher at 82ms average. The sawtooth frame-times there do suggest integration back into the AMD pipeline where frame-pacing isn't fixed on old tech. These are just some initial observations but enough to suggest that ANR deserves proper attention once it is out of beta, or when its super resolution component arrives. In the meantime, the beta is available to all - do check it out.





Comments 1
Thanks Rich! I love a good article over the full fat video for these ones. Reminds me of the eurogamer write-ups of old.
Show Comments
Leave A Comment
Hold on there, you need to login to post a comment...