Skip to content

v2.1.0

Latest
Latest

Choose a tag to compare

View all tags
@savery42 savery42 released this 12 Aug 15:25
v2.1.0
6468cec

xpu-smi

Bug Fixes

  • dump --device -1 now selects all available devices
  • xpu-smi [ --query-gp | dump ] --count now exit cleanly after the specified number of samples
  • dump -h metric field descriptions in the help output now wrap while maintaining formatting in smaller terminals
  • xpu-smi stats now reports 0.0 for PCIe read/write throughput when the link is accessible and idle instead of N/A
  • xpu-smi ps process name column now shows the program name only, not the full path
  • xpu-smi stats no longer cuts off values at 65 characters
  • ECC metrics now report correctable and uncorrectable error counts separately rather than as a combined total
  • xpu-smi [ --query-gpu | dump --metrics ] as a background job no longer causes the process to be suspended after the first sample. The interactive keyboard-quit thread is now skipped when the process is not in the terminal foreground
  • GPU pairs that share a NUMA memory domain but sit under different PCIe root complexes are now correctly reported as NODE rather than SYS
  • Serial number lookup now uses the IGSC library directly
  • Resolved a false-positive linker warning -Wfree-nonheap-object observed when compiling with GCC 15 + LTO
  • xpu-smi [ | ps] Process memory usage is now read directly from the kernel's per-process DRM client accounting, eliminating the previous multiplication by open file descriptor count. Forked processes no longer double-count inherited GPU contexts. iGPU memory regions (GTT / local) are now reported correctly alongside dGPU VRAM
  • xpu-smi config reset -d { bdf } no longer causes a host crash if there are active VFs
  • xpu-smi vgpu -j now properly outputs in JSON format

New Features

  • xpu-smi topology --p2p prints an N×N matrix showing whether each GPU pair supports peer read, peer write, MDF fabric, atomics, and PCIe access
  • All commands that accept --device, except for amc, now take a comma-separated list of indices or BDF addresses

xpumd

  • XPUMD config: Module name changes for readability / consistency
    • intelxpu => intel_xpu
    • intelxpuinfo => intel_xpu_info (also default name for gRPC local socket changed)
    • intelxpustatus => intel_xpu_status
  • XPUMD init: Do not fail if intel_xpu L0 Sysman init fails (fixes #129)
    • Avoids DRA driver exit when it's configured to use XPUMD for GPU health info
    • Allows intel_crashlog module to be used even when L0 driver stack finds no GPUs
    • Re-initializes Sysman after getting device detach / attach event
    • (fail_on_sysman_init_error config option restores earlier behavior)
  • Sysman Go bindings: Update to L0 API spec v1.16.24
  • XPUMD image: L0 driver stack update to 26.27.39122.11
  • Go dependencies: Security enhancements + other version updates (to gRPC, OTel etc modules)

Supported Operating Systems

  • Ubuntu 24.04
  • Ubuntu 26.04
  • Windows Server 2022 & 2025 (limited features, xpu-smi only):
    • GPU device information
    • GPU telemetry
    • GPU firmware update
    • GPU configuration

Package verification

Refer to *.sha256 files below for SHA256 checksums for Linux release files, and *.sig files for Windows release files.

Assets 10
Loading