joeasycompute/benchmark ? reverse-engineered prompt
Reverse engineered prompt
Build me a simple script driven GPU benchmark kit for a Linux machine that I can run with one command. I want it to compare GPUs across a few realistic workloads, synthetic LLM training, LLM inference, Stable Diffusion image generation, and Blender rendering. It should work on either NVIDIA or AMD if the right drivers are already on the host, and let me choose auto detect or force the backend in a shared config file.
Please make the main flow create a timestamped results folder, capture machine info, validate the config and system requirements up front, run each benchmark one after another, save logs and structured outputs, then combine everything into easy CSV summaries for comparing runs later. I also want a quick smoke mode that uses much smaller settings so I can verify the harness without waiting forever.
The first run should set up a local virtual environment automatically if needed. Blender can stay as a separate host install and should just be skipped if missing. Look up current docs online if you need to fill in details.
Have a live product UI? Try website reverse