Skip to content

A GPU task manager for Windows, Linux, and WSL: per-process VRAM, GPU activity, and LLM model/KV-cache insights (Windows AMD and NVIDIA).

Python

Clarify original NVML discovery benchmark hardware

Scott Hanselmana2e6579737ad
diff --git a/README.md b/README.md
index bd1ad57..39ff1b1 100644
--- a/README.md
+++ b/README.md
@@ -829,7 +829,7 @@ Identification is designed to stay cheap. Measured on one Windows machine:
|---|---|---|
| Process list, command line, listening ports | under 1 ms | Once per process / every 5 s |
| Windows adapter discovery (DXGI + exact-LUID PnP identity + PCI location, no telemetry sample) | ~0.86 ms median, 2.60 ms max in 30 warm reads, 26.1 ms first/cold call (Windows 11 build 28000, RTX Spark N1X, driver 32.0.16.2114); same machine without the PnP probe: ~0.36 ms median | Once at startup; TUI rediscovery every 5 s. PnP uses fixed buffers, one interface lookup per adapter and at most two name-property reads, no path/device enumeration or retries. Synchronous OS calls are not a hard latency bound. Other drivers, virtual displays and multi-GPU laptops unmeasured; provider reconstruction has additional driver-dependent cost |
-| Windows NVML handle reacquisition (no telemetry sample) | ~8.7 ms median, 9.3 ms max in 30 reads on the same machine | Every 5 s per active collector; retries initialization after NVML session-loss errors. Other drivers unmeasured |
+| Windows NVML handle reacquisition (no telemetry sample) | ~8.7 ms median, 9.3 ms max in 30 reads (Windows 11 build 26340, RTX 4080 SUPER, NVIDIA 630.20) | Every 5 s per active collector; retries initialization after NVML session-loss errors. Other drivers unmeasured |
| Memory-map scan (DLLs + mapped files) | 1-50 ms per process | New processes, a 64 MiB / 10% memory change, or every 2 min; at most 150 ms per refresh |
| GGUF header, model-store manifests | ms, cached | Once per file / every 60 s |
| Local API probes | ms, local only | llama-server `/v1/models` only when its model path changes |