New in this release

Target support

• Radeon GPUs - @gusteivos
• Tenstorrent accelerators - @OctopusET
• Iluvatar CoreX GPUs (ixML) - @lxzlxzliuxuzhao
• Qualcomm NPUs - @Srinivas-Kandagatla
  • HVX/HMX utilization plots
• Ascend NPUs - @KaranocaVe, @yulink1981
  • DCMI telemetry backend
  • DDR memory support and 310P3 device enumeration
• MSM: more GPU IDs, current clock and temperature reporting - @robclark, @fromdisposition

Improvements

• NVIDIA:
  • NVLink support - @danbedford
  • ECC support for professional GPUs - @LH-and-FPGA
  • Extra clock domains not reported by NVML (XBAR, SYS, NVD) - @Panchovix
  • Real memory usage on unified-memory platforms - @sudoingX, @johnnynunez
• Interface:
  • PCIe RX/TX load in the utilization charts - @Basavaraja-MS
  • Dynamic memory units with -u/--dynamic-memory-units - @unstable-code
  • Configurable plot line colors - @graysky2
  • Number keys to select a signal after F9 - @abpaudel
• Snapshot:
  • GPU architecture in JSON output - @Mahsuba

Fixes

• NVIDIA:
  • Correct MIG mode handling for process utilization - @Henry-ZHR
  • UMA memory fallback and unified-memory accounting - @redyuan43, @sudoingX, @johnnynunez
  • Do not report a fake PCIe link on unified-memory SoC GPUs - @Code4me2
  • Show 0% instead of N/A when there are no new utilization samples - @wuzhonglijz
  • Fix extra-clock flicker and use-after-free on re-monitored GPUs - @intel00000
  • Fix startup segfault and NVIDIA detection under WSL2 - @zzccppp
• AMDGPU:
  • Count GTT in per-process memory on integrated GPUs - @rborkow
  • Fall back to hwmon junction temperature on SMU13.0.6 - @jason34105533
  • Fix divide-by-zero when the power limit is missing - @JoseOliveira-beep
• Intel:
  • Fix ~20 MB/hour heap growth from per-refresh udev devices - @bqv
  • Read the xe GT frequency from sysfs instead of a perf event - @stolk
  • Fixes for Intel Xe cards and memory frequency reporting - @Steve-Tech
• fdinfo:
  • Apply the graphical default per process, not per fd - @stolk
  • Cap the engine busy percentage instead of dropping samples, avoiding a division by zero - @ViderStar
• Interface and snapshot:
  • Make --no-processes work with --snapshot and --loop - @yinghao-w
  • Fix null cmdline in snapshot output - @gitfool
  • Fix window cleanup leaks - @danbedford
  • Respect the hide-processes option from the config file - @yinghao-w
  • Increase the elapsed-time buffer on the timeline - @aadamowski
  • Clamp GPU and memory clock percentages to 100 - @bqv
  • Fix the shader core count field - @Panchovix
  • Guard clock-rate plots against a zero maximum, drop stale processes of unmonitored GPUs
• Other:
  • MSM temperature and current clock reporting - @fromdisposition
  • Various Ascend, Qualcomm and V3D fixes

Misc

• Dockerfile fix for Docker builds on Windows 11 - @piro4you
• Documentation for invalid TERM/ncurses errors - @raravind007
• Snap: terminfo packages for more terminals - @simonjgreen
• Snap: migrate to the core26 base - @Syllo
• AppImage: rebuild on AlmaLinux 8 without bundling glibc, add aarch64, and automate release builds - @Syllo
  • Requires glibc 2.28+ (Ubuntu 20.04+, Debian 10+, RHEL 8+)
  • RHEL 7 is no longer supported (EOL)

A huge thanks to all the contributors and testers 🎉

Full Changelog: 3.3.2...3.4.0