Wenhao Chai has released RTX H100, a Blender Cycles project that reports 2.5× an RTX 5090’s sample throughput on a single NVIDIA H100. Its fastest configuration uses a lossy, scene-specific neural cache that stops light paths early. The sample counter is impressed; image quality has follow-up questions. SourceSource
The November 20 code release includes patches, scene configurations, cache code and reproduction scripts. The reported results come from Blender 5.3.0 on one H100 SXM 80GB, tuned to three Open Data benchmark scenes: monster, junkshop and classroom. SourceSource
The RTX 5090 reference is an Open Data median, not a same-system head-to-head test. Source
Two speedups, two different jobs
The exact build specializes GPU kernels—the programs doing the rendering work—for each scene. The project reports roughly twice stock H100 sample throughput while keeping images equal up to floating-point rounding. Its score of 14,106 samples per minute is still below the listed 5090 median. Source
The nrc build trains a small neural network offline on exact renders for each scene, freezes it, then queries the cache when a camera path first hits a surface. The path ends there instead of continuing through later bounces. As Chai explains, “one of our samples does far less work than a stock one.” SourceSource
The sample-rate lead doesn’t establish that the H100 performs equivalent full path tracing 2.5 times faster than the 5090. SourceSource
What the 9.7× figure actually measures
The project also measures how quickly images reach a numerical quality threshold: the , evaluated against a much longer 65,536-sample reference. By that measure, it reports for the neural build and for the exact version. This comparison is H100 against H100—not another claimed lead over the 5090.
Comments