Subwarp Interleaving GPU, NVIDIA is studying solutions to improve ray tracing performance

The hybrid rendering that displaces the most advanced video games today, in which rasterization is accompanied by some effects of real-time ray tracing (global illumination, shadows, etc.), leads to greater realism of the image but at the same time performance suffers.

It is for this reason that NVIDIA (DLSS) first then AMD (FSR/RSR) – without forgetting Intel with XeSS in the near future – have proposed compensatory solutions capable of preserving maximum image quality and restoring a performance gain. First introduced with GeForce RTX 2000 GPUs in late 2018, real-time ray tracing is not a static technology and like all others is under continuous study to increase its quality and minimize its impact.

An article published by NVIDIA researchers at this addressentitled “GPU subchain interleaving“, seems to be heading in this direction. According to reports, it is an “architectural improvement” that he would be able – on the basis of the first tests – to guarantee an average performance increase of 6.3% (up to 20%).

Everything is very beautiful, at least in appearance, but there is a good. The Underdistortion interlacingwrites NVIDIA, “has shortcomings that prevent its implementation in the short termAccording to the researchers, some improvements in the architecture must be implemented, otherwise the practical performance could be limited.

nvidia rtx ray tracing 26 01 2022

Ray-tracing cores strain GPUs in three main ways: are highly divergent, suffer from low occupancy, and warps (thread blocks) hang while waiting for long-latency operations to complete,” writes NVIDIA. cores often have few active distortions to mask latency.”

Subwarp Interleaving is a new technique that aims to reduce pipeline bubbles in ray tracing cores. When a long-latency operation blocks a warp and the GPU warp scheduler cannot find an active warp to switch to, a subwarp scheduler can instead switch execution from the current warp to another diverging subwarp.

Demonstrate the technique NVIDIA has created what it calls “architectural extensions” for a “Turing-like” GPU which takes advantage of “most of the features inherent in the base architecture to support thread-independent scheduling”. Subwarp interlacing reduces exposed dropouts by 10.5%“.

Tests have shown that the technique, while showing some interesting performance improvements (6.3% on average, 20% maximum) in some areas of ray tracing application, unfortunately sees “restricted usage and complexity of design that limit attractiveness”. for current graphics architectures”.

“However, the evolution of application requests and their behavior can motivate future consideration of latency tolerance and divergence mitigation approaches such as underdistortion interleaving“.

Subwarp interlacing is probably just one of many projects Nvidia is working on to improve future architectures in the management of ray tracing: it is not certain that its application will become reality, but it is precisely this process of study and constant analysis of new techniques that allows us to obtain products ever more efficient and innovative.