TLDR
-
Open-source CLI that reads NVIDIA GPU performance counters directly to show real compute utilization, not just kernel-busy status.
Key Takeaways
-
nvidia-smi and nvtop report 100% utilization even when a workload uses a tiny fraction of actual hardware capacity; Utilyze exposes the gap.
-
Reads GPU performance counters via CUPTI 12+ and NVIDIA Perf SDK; requires
sudo or CAP_SYS_ADMIN unless NVreg_RestrictProfilingToAdminUsers=0 is set.
-
Computes an attainable compute SOL (Speed of Light) ceiling per model and hardware by detecting running vLLM inference servers; SGLang support coming.
-
Current hardware coverage: H100-80G and A100-80G, up to 8 GPUs per node; Ampere or newer required (A100, H100, H200, B200, RTX 3000+).
-
Attainable SOL feature anonymously sends GPU config data to Systalyze servers; disable with
UTLZ_DISABLE_METRICS=1.
Hacker News Comment Review
-
No substantive HN discussion yet.
Original | Discuss on HN
