Community benchmarks
What do AMD GPUs really run?
Measured results from real machines, ranked by decode speed. Every run keeps its stack, its memory phases and who vouches for it. Add yours with one command.
Leaderboard
Median decode tokens per second at batch size 1
Loading benchmark data…
Add your GPU
Runs on your machine. Nothing is sent unless you add --submit.
pip install torch transformers # use the ROCm build of PyTorch on AMD
python bench/rocm_bench.py --model Qwen/Qwen2.5-1.5B-Instruct --contexts 512,2048
# review amd_llm_lab_results.json, then share it:
python bench/rocm_bench.py --model Qwen/Qwen2.5-1.5B-Instruct --contexts 512,2048 --submit --api <hub-url>The results file is plain JSON you can read first. It contains the GPU name, memory, ROCm and torch versions, OS name and the measured numbers. No prompts, paths or usernames.
How to read the table
- Decode speed is tokens per second while generating, with prefill timed separately and removed.
- Memory is split into steady (held once loaded) and run peak. Older reference runs only have run peak, which for quantized models includes the loading spike.
- Bandwidth used compares decode speed with the GPU’s published memory bandwidth. A low number means the software stack, not the hardware, is the limit.
- Verified runs are checked by a maintainer. Community runs pass automatic physics checks. Runs that fail them are held back.