GPU PCIe Diagnostic & Bandwidth Analysis — v2.7.4
Initial public release of a deterministic, hardware-level PCIe diagnostic tool for NVIDIA GPUs.
This release provides a reproducible baseline for validating GPU PCIe link health and transport behavior using only observable data.
Capabilities
- Detect negotiated PCIe generation and link width (current vs max capability)
- Measure peak Host↔Device copy bandwidth via CUDA memcpy timing
- Measure sustained PCIe utilization using NVML TX/RX counters
- Compute efficiency relative to theoretical PCIe payload bandwidth
- Emit deterministic verdicts: OK / DEGRADED / UNDERPERFORMING
- Optional read-only PCIe AER counter inspection (if exposed by platform)
Design Principles
- No BIOS or firmware modification
- No registry hacks or tuning assumptions
- No inference when data is unavailable
- Reports observable facts only
Validation
Copy timing and telemetry behavior cross-validated using NVIDIA Nsight Systems.
Scope
This tool evaluates PCIe transport behavior only.
It does not profile kernels or applications.
Source-only release.
Build artifacts and runtime logs are intentionally excluded.