Full path tracing might not be such a GPU hog in future, as Nvidia claims over 2x performance improvements in new research

Ray tracing is one of those technologies that's mighty impressive but also seems to take a mighty amount of time to get right. We saw the first proper go at it with the RTX 20-series, and we're now three generations on and still relying on workarounds to achieve global illumination. One technique to help make path traced lighting more efficient is ReSTIR, and now Nvidia researchers claim to have improved its performance by over 2x.

The new research lays claim to an "enhanced" version of the technique(s), called ReSTIR PT Enhanced, and explains how these improvements are made.

But first, the original ReSTIR itself. Here is the explanation as the researchers put it:

"Reservoir-based spatiotemporal importance resampling (ReSTIR) applies chained GRIS passes across frames to progressively improve the sample distribution, greatly improving real-time image quality. Each pixel maintains a reservoir, which is essentially a tuple (????,????????, ????) carrying a sample, its unbiased contribution weight, and a confidence weight (originally called an “effective sample count”) used to weigh domains/techniques when computing resampling MIS. In a GRIS pass, the current pixel’s reservoir provides the canonical sample, and other pixels contribute neighboring samples. The reservoir is then updated to the selected output sample ????, whose confidence weight accumulates those of all input samples."

Err… yeah, I'll stick to the simple version. Which is, as far as I can tell, that for each lit pixel, you check neighbouring ones and previous frames to figure out which light samples are worth using, and then reuse these samples whenever relevant. In a game scene with tons of light sources and potential light paths, cutting down the number of these you actually trace is a good way to keep things efficient and the performance overhead low.

This technique has been most useful for full path tracing, which hits powerful GPUs very hard even while using ReSTIR. Case in point: a game like Alan Wake 2. That's one of the reasons proper path tracing is so rare—often we get workarounds like Lumen to achieve global illumination rather than full-fat path tracing.

The improvements to the tech these Nvidia researchers are proposing should hopefully go some way towards changing that. They include the following:

• "Halving shift mapping costs in spatial reuse by reciprocal neighbor selection
• New ray footprint thresholds that adapt to scene and materials
• Reducing correlation artifacts by sample duplication maps
• Improving quality and cost by unifying ReSTIR for direct and indirect light
• Other optimizations that boost performance and improve robustness by reducing color and disocclusion noise."

The researchers used an Nvidia RTX 5880 to test it out over four scenes, and that's not a typo but rather a workstation card that's somewhere between an RTX 4080 Super and RTX 4090 in performance.

Taking the important parts out from the nitty gritty of the results, all the improvements the researchers have made to ReSTIR give a 2.74x performance increase, and even with some other quality improvements thrown in (noise reduction, etc), it's still 2.3x faster than the original ReSTIR.

Overall, "our enhanced ReSTIR PT reduces correlation artifacts, color and disocclusion noise. It exhibits better quality thanks to our new reconnection criteria… and the unification of direct and indirect lighting, yet runs significantly faster, reaching 2.08x-3.05x speedup over [the original ReSTIR]."

Which ultimately means, as the researchers say in the paper abstract, that ReSTIR PT is "closer to production-ready." And this means we could be a whole lot closer to actually usable, mainstream path tracing in games.