Skip to content

v1.5.0

Choose a tag to compare

@bramvanoploo bramvanoploo released this 20 Nov 13:40
· 2 commits to main since this release

Changes

  • NVIDIA Tesla & Datacenter Hardening:
    • Expanded detection across Kepler (K10–K80), Maxwell (M4–M60), Pascal (P4–P100), Volta (V100 sm_70), Turing/Ampere/Ada/Hopper/Blackwell datacenter cards (T4, A2–A100, L4–L40S, H100–H200, B100–B300, GB200, GH200, CMP *HX, RTX PRO 6000 Server), NVSwitch fabrics, and passive thermal capabilities (refresh_nvidia_caps, refresh_nvidia_mig).
    • Added hardware-gated datacenter steps: nvidia_fabricmanager (NVSwitch Fabric Manager), nvidia_mig_disable (disable MIG partitioning via nvidia-smi -mig 0), nvidia_sm70_table (STRATA_SM70_TABLE=1 for V100 sm_70), nvidia_pstate (nvidia-pstated / Python nvidia-pstate idle P0→P8 fix with clean SIGTERM/revert restoration), and nvidia_power_cap (opt-in passive/datacenter thermal power cap preserving [a] mutual exclusivity).
    • Hardened clock locking (-lgc error-text check with -ac + --auto-boost-default=0 fallback), BIOS ReBAR capability detection (BAR1: display for legacy Tesla cards), non-systemd nvidia_persistenced fallback, consumer-only gpu_unlock / lact_overclock scoping, and CUDA 12 / R580 driver branch advisories.
  • Multi-GPU, DGX Spark, Jetson, Instinct & Intel Arc Expansion:
    • Added bare-metal multi-GPU pcie_acs_disable targeting PCIe bridges (0604) via setpci when NVLink/XGMI is absent.
    • Added NVIDIA DGX Spark (GB10) unified memory detection, NVIDIA Jetson Orin/Xavier/Thor/TX2 + L4T R35/R36 ina3221 telemetry & tegra_pwmfan support, AMD Instinct CDNA 1–4 (gfx908, gfx90a, gfx942, gfx950) PCI class 0x1200 + XGMI + amd_instinct_partition (SPX/CPX, NPS1/NPS4) + wave64 local HIP builds, and Intel Arc Battlemage/Alchemist/Flex/Max (B580, B570, Pro B50, Pro B60) VRAM tally, multi-GPU topology, openSUSE compute packages, and xe hwmon live telemetry & power sampling (while filtering out 0x1200 Intel NPU/QAT devices).
  • Index Legend Flags, Route Classification & Performance:
    • Added dynamic [!], [OPT], [EXP], and [L] legend rows in the TUI index with aligned cursor offset math and [a] / [l] / --check route filtering.
    • Consolidated nvidia-smi probing into a single query when supported, cached live nvidia-smi extra fields, eliminated hot-path subshells in run_as_user() / run_as_user_exec(), fixed \bati\b / \bArc\b word boundaries in GPU detection, updated documentation, and bumped manifest.json to 1.5.0.

Full Changelog: v1.4.0...v1.5.0