Skip to content

PANTHEON

Pantheon Logo

Universal GPU Stress & Diagnostics Suite

View on GitHub Live Benchmarks


Why Pantheon?

Pantheon is a cross-platform (CUDA/ROCm) stress testing tool designed to isolate and hammer specific GPU subsystems. Unlike generic benchmarks, Pantheon targets specific silicon limits to expose hardware degradation, thermal throttling, and architecture bottlenecks.

  • VRAM & Infinity Fabric --- Target HBM/GDDR memory and interconnects with aggressive crosstalk patterns to detect bit flips and signal integrity loss.

    Test: hbm_write_agg, hbm_tsv_thrasher

  • Physical Matrix Cores --- Push hardware Tensor and Matrix cores to their absolute thermal limits using universal FP16/WMMA intrinsics.

    Test: mma_virus, tensor_virus

  • VRM Transients (dI/dt) --- Oscillate maximum load at 10Hz to induce high current transients, testing your power supply and voltage regulators.

    Test: pulse_virus, voltage

  • MMU & TLB Avalanche --- Force near 100% Translation Lookaside Buffer misses with random page-boundary jumps to choke hardware page-table walkers.

    Test: tlb_avalanche

  • ECC & RAS Validation --- Continuously read pristine patterns to expose the latency jitter of active ECC scrubbing and silent data corruption.

    Test: ras_validator

  • Hardware Scheduler --- Spam micro-kernels across 64 concurrent streams to force the GPU dispatcher into multiplexing mode and trigger context switch locks.

    Test: scheduler

Quick Start

# Clone the repository
git clone https://github.com/saqibkh/pantheon.git
cd pantheon

# Run the full suite (30 seconds per test)
python3 pantheon.py --test all --duration 30