Skip to content

v2.2.0

Choose a tag to compare

@savery42 savery42 released this 16 Sep 19:57
· 467 commits to main since this release

xpu-smi

Bug Fixes

  • xpu-smi stats GPU utilization now reports the busiest engine rather than the average across all engines; a compute-only workload on a multi-engine device (e.g. Arc Pro B70) previously read ~11% instead of 100%
  • xpu-smi [ --query-gpu | dump --metrics ] energy.consumed (Card Energy Consumed) was previously read from the GPU power domain, which not every platform enumerates, causing N/A on hardware that does have a usable energy counter; it now reads the card-domain counter consistently. Added energy.consumed.gpu for the GPU-domain value
  • xpu-smi stats offline memory page values are now printed and failure is reported honestly instead of silently
  • xpu-smi dump metric values are no longer truncated when column-aligned output is used
  • xpu-smi stats PCIe read/write throughput now reports 0.0 for an inactive link instead of N/A
  • xpu-smi raslog non-destructive peek is now the default; use --drain to consume and clear records from the hardware buffer (previously every invocation drained the buffer)
  • xpu-smi vgpu -l -d <device> no longer fails on kernels that use the older VRAM debugfs path
  • xpu-smi discovery PCH SKU type is now read from firmware rather than returning a hardcoded string
  • xpu-smi RATL throttle reason is now mapped to a PSU alert event
  • xpu-smi config reset -d <bdf> no longer causes a host crash when active VFs are present
  • xpu-smi part number field is now only shown in discovery output when AMC is present
  • xpu-smi CPU affinity column now wraps correctly for devices with many cores
  • xpu-smi --display section list is now generated from the group table, fixing missing sections
  • xpu-smi fwupdate GFX data firmware update is now applied correctly
  • Resolved a false-positive linker warning with GCC 13.1 and LTO

New Features

  • xpu-smi raslog new command reads CPER (Common Platform Error Record) hardware error logs via the Level Zero sysman InfoLog extension (ZES_intel_experimental_driver_info_logs)
    • Default (peek) behaviour reads records without draining them; use --drain --file <path> to consume and persist
    • Note: currently based off of unstable Sysman APIs, this release will return UNSUPPORTED_FEATURE for all raslog command

  • xpu-smi crashlog new command drives the Intel Crash Log CLI (iclg) for platform crash log management: enable, disable, trigger, clear, extract, and decode, selecting a device by index (-d)
  • xpu-smi config now supports setting a device health status; the configured status is reflected in xpu-smi stats output
    • Note: configuring and reading device health status requires Level Zero ≥ 1.33.1 (spec v1.18). On older driver stacks the commands compile and run but return UNSUPPORTED_FEATURE; all other xpu-smi functionality is unaffected.

  • xpu-smi discovery now reports TDP information, memory vendor, fine-grained video memory specification, GFX Data firmware version, and kernel-mode driver srcversion
  • xpu-smi topology PCIe switches are now detected dynamically via PCIe port type; PCIe Gen 5 switch IDs are included in the fallback PCI database
  • PLDM type 5 extraction is now supported for unwrapping GPU firmware update binaries

xpumd

  • XPUMD metrics: New hw.gpu.ecc.state metric for the device memory ECC state
    • Breaking change: replaces the ecc_* states of the hw.status metric
    • The state name in the hw.state attribute has no ecc_ prefix (e.g. enabled)
    • Queries, dashboards and alerts need to be updated from hw_status{hw_type="gpu",hw_state="ecc_<state>"} to hw_gpu_ecc_state{hw_state="<state>"}
  • XPUMD metrics: New GPU hw.status states from the L0 device state extension: wedged, survivability and flash_override
  • Metric output fixes:
    • hw.power.limit was reported million times too large (mW → W conversion error)
  • XPUMD fixes: Crash on zero sampling_interval, and on missing parent metric in the intel_xpu_status processor
  • XPUMD dashboard: new metrics and usability improvements. Requires Grafana 13.x or newer
  • XPUMD privileges: Helm chart now uses PERFMON capability instead of SYS_ADMIN
    • This improves security, but some (typically very old) setups may still require the wider SYS_ADMIN capability for GPU utilization metrics access
  • XPUMD image:
    • L0 driver stack update to 26.31.39395.13
    • Libraries needed by the RAS metrics (libnl) and the info logs (libtracefs) added
  • Go dependencies: OTel collector update to v0.159.0 / v1.65 + other version updates

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

Supported Devices

  • Intel(R) Arc™ Pro B50 GPU
  • Intel(R) Arc™ Pro B60 GPU
  • Intel(R) Arc™ Pro B65 GPU
  • Intel(R) Arc™ Pro B70 GPU

Package verification

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