Pantheon runs about fifty targeted workloads against NVIDIA and AMD GPUs to find out whether a card is healthy: memory diagnostics adapted from the DRAM testing literature (march tests, disturb patterns, retention bakes), power viruses aimed at individual subsystems, interconnect tests and AI shaped compute, with ECC and RAS counters recorded around every run. The kernels ship as source and compile on first run, so anyone can read exactly what will touch their hardware. GPUs fail quietly. A card can pass the vendor's tools while corrupting data or losing throughput, and the people renting GPUs by the hour, buying them used or running small clusters have no independent way to check. Pantheon is that check. It is packaged on PyPI, a signed apt repository, Fedora COPR and container images for CUDA and ROCm; a daily CI job reinstalls from every channel; and results from cloud and community systems feed a public, comparable database. Funding pays for the unglamorous work that keeps a hardware testing tool trustworthy: maintaining the release channels and CI, running the results database and its privacy rules, renting hardware across GPU generations so releases are validated on real cards, and answering the people who find bugs.
Fund this project