Nvidia developers have been experimenting with new ways to increase GPU ray tracing efficiency. GPU Subwarp Interleaving, according to a newly published paper noticed by 0x22h, can speed up raytracing by up to 20%.
But it requires micro-architectural changes to get that number. So Nvidia has to produce a new GPU with these changes baked in as architectural changes rather than using a tweaked GPU like Turing, which was used in this research. So the benefits of Subwarp Interleaving won’t show up till the next generation.
Real-time ray tracing will continue to see a rise in the GPU space. Nvidia is doing everything to stay ahead of the competition and a marketing boost by attacking the problem from different angles. With that in mind, a group of researchers including GIT professor Sana Damani, Nvidia’s Ram Rangan, Daniel Johnson, Rishkul Kulkarni, and Stephen W. Keckler has released a paper showing initial results in ray tracing micro benchmarking.
According to the report, new Nvidia GPUs are built to hinder their ability to achieve tasks. According to the study’s authors, cards use a single program counter (PC) to fetch and process threads in SIMT (single instruction, multiple threads) form. They can also mask delays on a GPU by using multiple warps simultaneously. Such design options hamper real-time ray tracing because of warp dispersion, warp-starved conditions, and reduction in graphics performance as the scheduler exhausts its threads and can’t mask any pauses anymore.
Add the Subwarp Scheduler
A suitable option for graphics cards that are overwhelmed with distortions and running out of threads like the above is GPU Subwarp Interleaving (GSI). Subwarp schedulers can be used instead of warp schedulers when a prolonged latency activity stalls the GPU’s scheduler and it can’t find an active warp.
A microarchitecture-enhanced Turing-like graphics card featuring Subwarp Interleaving delivered a “captivating improved performance” of 6.3% on average and nearly 20% in the best scenario in a suite of raytracing workloads.
As discussed earlier, existing GPU generations will not be able to take advantage of the RT performance upgrades. A driver upgrade of the current cards from Team Green will not give you an extra 20% performance gain. Upcoming GPU technologies will rely heavily on these side operations. GPU Subwarp Interleaving is likely only one of several initiatives Nvidia is concentrating on in the back for future progress.
