Releases: Thenukegun10x/GPU-SMI
Release list
gpu-smi v1.2.0 - Hardened Headless API & Linux Backend Parity
gpu-smi v1.2.0 - Hardened Headless API & Linux Backend Parity
Security + Linux catch-up release. No breaking changes: loopback --serve stays zero-config.
What's New in v1.2.0:
- Bearer token auth for
--serve:--token/--token-file/GPU_SMI_TOKEN(constant-time compare)./healthstays open for probes;/+/metricsrequire the token when one is set. Non-loopback--listen(e.g.0.0.0.0for LAN Prometheus scraping) refuses to start without a token (fail-closed). - HTTP hardening: strict
METHOD PATHrouting with405/404(kills the old/metrics-evilsubstring confusion), loopbackHost:check against DNS rebinding, removedAccess-Control-Allow-Origin: *so random websites can't scrape your process list, 5s read timeout, proper Prometheus label escaping,nosniff+no-storeheaders. - DLL sideload guard:
System32/driver absolute paths are tried before bare names for ADL/HIP/system DLLs. - Linux backend parity: full sysfs sensors (edge/hotspot/VRAM temps, gfx/mem clocks via hwmon +
pp_dpmfallback,gpu_busy_percent, power + cap, PCIe width/speed, BDF, vbios, serial, driver version), per-process VRAM via/procfdinfo (deduped by client-id, same overcount guard as Windows), HIP merge for names/CUs, connector-device phantom-GPU fix. - Correct gfx IDs: shared PCI-device table fixes 7900 XTX ->
gfx1100, 7800 XT ->gfx1101, 6900 XT ->gfx1030, Raphael ->gfx1036, Phoenix ->gfx1103(were mislabeled). - Token setup template:
example.env(copy to.env,chmod 600);.gitignorenow covers real secrets.
Downloads:
gpu-smi-windows-x64.exe(1.0MB) - Portable standalone single executable (Windows x64), reports1.2.0.- Linux binary: build on your box with
cargo build --release(this release was cut from Windows, no Linux linker available here).
Quick Start:
gpu-smi-windows-x64.exe --serve 8080 # loopback, no auth (unchanged)
openssl rand -hex 32 > token.txt # see example.env for full setup
gpu-smi-windows-x64.exe --serve 8080 --listen 0.0.0.0 --token-file .\token.txt
curl -H "Authorization: Bearer $(cat token.txt)" http://<host>:8080/metricsgpu-smi v1.1.0 - VRAM Usage, OS Pinned VRAM & Process Tracking
gpu-smi v1.1.0 - Dedicated VRAM, OS Pinned Memory & Process Tracking
Major feature update bringing live dedicated VRAM utilization, OS pinned memory reporting, and per-process GPU tracking directly into the base output table.
What's New in v1.1.0:
- Per-Process GPU & VRAM Tracking: Real-time discovery and attribution of all GPU workloads (
PID,Name,TypeC/G,Memory Usage) using native Windows WDDM PDH counters mapped to hardware adapter LUIDs. Zero admin privileges or external tools required. - Dedicated VRAM Usage: Live dedicated VRAM consumption (
BytesResident) alongside total adapter memory and percentage utilization directly in the base table. - OS Pinned VRAM: Non-evictable kernel driver & display compositor pinned memory calculation (
BytesResident - BytesCommitted). - Prometheus Metrics: Added
amd_gpu_process_mem_mbandamd_gpu_vram_pinned_mbto the/metricsendpoint on--serve <PORT>. - JSON Output: Fully serialized
processesarray with PIDs, names, types, and usage for programmatic scripting and tooling.
Downloads:
gpu-smi-windows-x64.exe(1.0MB) - Portable standalone single executable (Windows x64).gpu-smi-linux-x64(963KB) - Portable standalone single executable (Linux x64).
Quick Start:
gpu-smi-windows-x64.exe # table with GPU stats & running processes
gpu-smi-windows-x64.exe --verbose # show all processes + VRM/hotspot temps & clock rates
gpu-smi-windows-x64.exe --json # full JSON output
gpu-smi-windows-x64.exe --serve 8080 # headless HTTP & Prometheus /metricsgpu-smi v1.0.0 - portable single-exe for AMD
gpu-smi v1.0.0 - first release
Portable single-exe GPU SMI for AMD GPUs (Windows+Linux). AMD-compatible, unofficial.
Download: gpu-smi-windows-x64.exe (1.0MB) - just run, no installer, no ROCm SDK required.
Quick start:
gpu-smi-windows-x64.exe # table gpu-smi-windows-x64.exe --watch 1 # realtime gpu-smi-windows-x64.exe --json # JSON pipe gpu-smi-windows-x64.exe --serve 8080 # headless HTTP: curl http://127.0.0.1:8080/
Features:
- Auto-discovery HIP (TheRock) -> ADL PMLog (temps/clocks/power/util) -> WMI -> sysfs
- VRAM type autodetect DDR5/DDR4/LPDDR/GDDR5/HBM2/GDDR6 (not hardcoded)
- --watch realtime, --serve headless API (GET / JSON, GET /metrics Prometheus)
- Memory-safe Rust, self-contained (kernel32+VCRUNTIME140 only)
Build from source: cargo build --release
Disclaimer: Not affiliated with Advanced Micro Devices, Inc. AMD/Radeon are trademarks of AMD.
See README for Python/Node examples.