← All resources

CUDA Ray Tracing 2x Faster Than RTX: My CUDA Ray Tracing Journey

Auto-discovered

Blog Advanced

Karim Sayed

Visit resource ↗

Description

This is a long-form technical blog post by Karim Sayed documenting a CUDA path tracer based on Ray Tracing in One Weekend and comparing it with GPSnoopy's RayTracingInVulkan RTX sample on an RTX 3080.

It walks from a naive multi-second CUDA port down to roughly 8 to 9 ms through header-only inlining, replacing recursion with an explicit BVH stack, register and occupancy tuning with Nsight Compute, memory layout changes, and related CUDA pitfalls. It discusses why compute-based procedural sphere tracing can beat RT cores in this synthetic case, with caveats about AAA triangle workloads and fair Russian Roulette comparison adjustments that reduced an earlier 3.6x claim to about 2x.

It is for intermediate to advanced graphics programmers learning CUDA performance work or comparing software ray tracing with hardware RT pipelines. Familiarity with BVHs and basic path tracing helps.

Access is a free HTML article with tables, code excerpts, and a linked source repository and commit history.

It is useful as a concrete optimization narrative for CUDA ray tracing rather than as a claim that RT cores are obsolete in games.