Submit your benchmark

Your numbers help everyone with the same card. It takes about five minutes. No account needed.

  1. Get the script. rigbench_submit.py (one file, Python 3.8+, no dependencies; read it before running, it is short).
  2. Run it with llama.cpp's llama-bench or with Ollama: 
    python3 rigbench_submit.py --llama-bench /path/to/llama-bench --model /path/to/model.gguf
python3 rigbench_submit.py --ollama llama3.1:8b
    It detects your GPU, runs the benchmark, and writes rigbench-result.json.
  3. Submit. Add --submit to send it straight away, or pick the file below.

or paste it:

Each submission is checked automatically (format, a known card with a sourced bandwidth spec, and the physics limit below); anything unusual waits for a human. Approved results appear after the next daily refresh with their own public source page. Results that break physics (faster than the card's memory bandwidth allows) are rejected automatically.

Report a wrong number

What we collect

GPU/CPU model, VRAM, OS, backend and build, model file name and size, and the measured speeds. We also store the submission time and country (from the connection, not the IP address itself). No serial numbers, hostnames or IP addresses are kept with the result.