p6\n\nProbably Programmed Properly, Pending Physical Proof

A next-generation, cross-vendor, zero-dependency GPU hardware diagnostic suite built in safe Rust using explicit Vulkan APIs.

p6 is designed to be a modern replacement for aging CUDA-based tools like gpu-burn. By leveraging Vulkan, it supports Intel, AMD, and NVIDIA hardware completely natively without proprietary SDKs. It isolates and tortures physical GPU silicon blocks that standard Compute APIs cannot reach.

Features

• Total Hardware Saturation:

  • FP32 & FP64 Compute: Standard math loops designed to maximize streaming multiprocessor utilization and catch ALU bitflips.
  • AI Tensor Cores (--tensor): Leverages VK_KHR_cooperative_matrix to target proprietary Matrix hardware (Intel XMX, Nvidia Tensor, AMD Matrix) for immense FP16 MMA stress.
  • Hardware Ray Tracing (--rt): Manually builds Bounding Volume Hierarchies (BLAS/TLAS) and spams Ray Queries (VK_KHR_ray_query) to torture dedicated optical traversal silicon.
  • PCIe Interconnect & Resizable BAR (--pcie): Saturates the Host <-> Device motherboard bus by streaming direct to CpuToGpu mapped VRAM. Identifies faulty PCIe riser cables and degraded Gen4/5 lanes.

• Advanced Chaos Vectors:

  • Voltage Droop Transients (--transient): Rapid pulse cycling of workloads to induce massive di/dt power draw spikes to crash unstable VRMs and capacitors.
  • Cache Thrashing (--thrash): Forces chaotic memory bounds to defeat L1/L2 caches and overwhelm the Translation Lookaside Buffer (TLB).

• Enterprise / Server Farm Ready:

  • Hardware Fault Detection: Defeats driver Dead Code Elimination (DCE) to natively catch ALU bitflips and memory corruption via atomic error buffers.
  • Multi-GPU Parallelism (--all-gpus): Asynchronously dispatches threads to stress every physical graphics device on the system simultaneously.
  • Sequential Jobfiles (--jobfile sequence.toml): Supports stress-ng style TOML execution sequences. The engine cleanly drops and recreates the Vulkan context between jobs to guarantee zero VRAM leaks across hours-long validation suites.
  • JSON Export (--json out.json): Dumps structured telemetry (PetaFLOPs computed, TeraRays cast, Petabytes moved, Faults) directly to disk for automated server ingestion.

Installation

cargo build --release
./target/release/p6 --help

Usage Examples

Run a standard Compute Burn on GPU 0 for 60 seconds:

p6 --c -t 60

Stress the Ray Tracing Cores across all installed GPUs for 15 minutes:

p6 --rt --all-gpus -t 900

Diagnose an unstable PCIe Riser cable:

p6 --pcie -t 10

Export metrics to JSON for CI/CD:

p6 --tensor -t 30 --json telemetry.json

Sequential Jobfiles

Create a sequence.toml file to run multiple workloads back-to-back:

[[jobs]]
workload = "pcie"
duration_s = 60

[[jobs]]
workload = "rt"
duration_s = 60

[[jobs]]
workload = "compute"
duration_s = 60

Execute the sequence:

p6 --jobfile sequence.toml --json full_report.json