Developers release Utilyze, an open source tool that corrects misleading GPU utilization metrics common across nvtop and cloud dashboards. Traditional metrics report kernel presence, not actual compute work, causing systems to appear saturated. This new approach measures hardware performance counters directly, giving engineers real throughput data relative to theoretical limits and exposing hidden bottlenecks.

Manya Ghobadi, MIT Professor and Systalyze CEO, explains existing tools like DCGM’s SM Active fail because resident warps may wait on memory. Utilyze implements the Speed-of-Light model, reporting separate Compute and Memory SOL % values. It samples counters to distinguish arithmetic work from idle waits, providing continuous observability without perturbing production workloads.

Validation compares Utilyze readings against direct FLOPS calculations for matrix operations, showing results within 2% accuracy for simpler kernels. For complex AI pipelines, direct calculation becomes impractical, so hardware-counter measurement becomes essential. Utilyze delivers constant, low-overhead visibility into true utilization, helping teams right-size infrastructure and avoid costly misallocations based on false data.