Cross-platform GPU stress testing and diagnostics
PANTHEON is a cross-platform GPU stress testing and diagnostics tool for Linux that validates compute, memory, cache, interconnect, and power behavior on NVIDIA (CUDA) and AMD (ROCm/HIP) hardware. It ships with 45 focused workload tests that range from compute viruses (FP64, tensor, integer, SFU) to memory aggressors, interconnect exercisers, and AI/ML workload suites.
Covers compute, memory, cache, interconnect, fixed-function accelerators, and AI/ML kernels—each designed to target a specific GPU subsystem.
Runs natively on both NVIDIA (CUDA) and AMD (ROCm/HIP) platforms, with automatic platform detection (no --platform flag needed).
Includes LLM decode, LLM prefill, KV cache churn, fused attention, RoPE stress, quantized GEMM, speculative decode, MoE router, transformer train step, and RAG embedding tests.
Tests include atomic virus, cache latency, aggressive/standard memory read and write, memory bank thrasher, cache fracturing, retention bake, and asymmetric thermal scenarios.
Covers P2P thrasher, all-reduce, TLB avalanche, PCIe thrasher, RAS validator, memory TSV thrasher, memory PC ping-pong, and scheduler virus.
Captures results locally, allowing comparison against the public performance database and export for further analysis.
Debian package for Ubuntu/Debian; portable bundle with install.sh for RHEL, Fedora, Rocky Linux, and AlmaLinux. Includes a complete-removal script that clears runtime-created files and caches.
First-run workload builds are cached under ~/.cache/pantheongpu/builds/, and the cache directory can be redirected via PANTHEON_BUILD_CACHE_DIR.
No. Pantheon automatically detects CUDA, ROCm/HIP, or mock mode. You run the pantheon command directly without passing --platform.
The Debian package is the simplest path for Ubuntu/Debian. A portable release bundle is available for RHEL-family systems and other Linux distributions, installed via ./install.sh.
Run sudo apt-get remove pantheongpu for package-managed files, plus curl -fsSL https://pantheongpu.com/uninstall.sh | sudo sh to also remove runtime-created files and the current user's compiled workload cache. This leaves CUDA, ROCm, system compilers, and benchmark reports outside Pantheon's directories untouched.
Yes, you can run pantheon --test baseline_metrics --duration 10 for a short hardware inventory test, which gives baseline metrics without lengthy stress runs.
First-run workload builds are cached under ${XDG_CACHE_HOME:-$HOME/.cache}/pantheongpu/builds/. You can override this with the PANTHEON_BUILD_CACHE_DIR environment variable.
Yes, Pantheon includes a public performance database for live benchmarks and comparisons, and results are stored locally for your own comparisons.