The performance results published in The CUDA Handbook were measured on 2012--2013 hardware. This page collects reader-submitted results from running the book's benchmarks on modern GPUs -- so the book's performance story stays alive, generation after generation.

Submitting results is a perk of membership: build the open source repository, run python3 bench/run.py --bin-dir build --suite all --submit, and enter the one-time code sent to your membership email. See Running and Submitting Benchmarks for the criteria a submission must meet and what happens after you send one. (A browser-based submission form is planned for those who'd rather not use the CLI.) Submissions are reviewed before they appear here; raw program logs are committed alongside every submission.

Application benchmarks

N-body

Chapter 14 · 1 metric · 1 GPU · view results

Histograms

Chapter 16 · 15 metrics · 1 GPU · view results

Microbenchmarks

Global memory bandwidth

Chapter 5 · 3 metrics · 1 GPU · view results

Copy/kernel overlap

Chapter 6 · 6 metrics · 1 GPU · view results

Kernel launch overhead

Chapter 6 · 4 metrics · 1 GPU · view results

Memcpy latency and bandwidth

Chapter 6 · 8 metrics · 1 GPU · view results

Hardware

GeForce RTX 3060 {#gpu-rtx-3060}
Ampere (SM 8.6) · 28 SMs @ 2160 MHz · 12 GiB @ 192-bit · L2 2304 KiB · 2 copy engines · PCIe 4.0 ×16 · driver 580.95.05 · CUDA 12.9 / 13.0 · 2 submissions

Newest hardware on record: GeForce RTX 3060 (Ampere).

Contributors

With thanks to: Nicholas Wilt.