It measures:
- CUDA memcpy throughput (H2D / D2H) - NVML PCIe bus-level traffic - Negotiated link state (pre/post load) - Replay counter deltas - AER error counters - Clock state transitions - NUMA affinity and system signals
The goal is to validate transport health from the negotiated PCIe link state down to actual transfer behavior.
This is not a compute benchmark. It focuses strictly on transport integrity and link behavior.
Built for Linux using CUDA + NVML.
gpu_systems•1h ago
It measures:
- CUDA memcpy throughput (H2D / D2H) - NVML PCIe bus-level traffic - Negotiated link state (pre/post load) - Replay counter deltas - AER error counters - Clock state transitions - NUMA affinity and system signals
The goal is to validate transport health from the negotiated PCIe link state down to actual transfer behavior.
This is not a compute benchmark. It focuses strictly on transport integrity and link behavior.
Built for Linux using CUDA + NVML.