Share your results

Numbers from other people’s P100s show whether a recipe holds up on cards, drivers and hosts it was not measured on. This page collects what makes a result useful, and says where to post it.

You need a model running and benchmarked: see Benchmark a running server.

1. Benchmark as JSON

Run the benchmark with its defaults, so your numbers compare with everyone else’s, and save the JSON:

tensorless bench iamMrDuncan/Qwen3.8-27B-PN3-GGUF --json > bench.json

It holds every run, the medians, and the difference from the recipe’s receipt.

2. Record the launcher and the card

tensorless version
nvidia-smi --query-gpu=name,memory.total,driver_version,power.limit --format=csv

tensorless version names the release and commit. The nvidia-smi line gives the card’s exact model, its memory, the driver version and its power limit, which can all change the speed. tensorless capabilities --json adds the recipes your launcher knows, if someone asks.

3. Note how you ran it

Say, in a sentence or two:

• whether you used podman or docker;
• whether you passed --context or arguments after -- (an untuned run measures something else);
• whether anything else used the GPU while it ran;
• anything unusual about the host: a PCIe riser, a card in an x4 slot, a cooling mod.

The JSON’s endpoint shows the address you benchmarked; check that you are happy to share it.

4. Post it

• Discord: join the tensorless server and post your results with the files above. It is the place for questions, results and ideas.
• GitHub: if your numbers are far from the receipt, or something failed, open an issue with the same details, so it can be traced.

Check that it worked

Your post includes bench.json (or its median and vs_receipt_percent), the output of tensorless version, the nvidia-smi line, and how you ran it. That is enough for someone else to repeat your measurement.