Intel Arc Pro Battlemage — Linux fan control, GPU tuning & overclocking

A complete Linux control panel for the Intel Arc Pro B60 / B70 (Battlemage) — custom fan curves, power/clock tuning, voltage-frequency-curve overclocking, a live metrics dashboard, a fan-guarded stability test, multi-GPU support, and kernel-update resilience — where the stock xe driver exposes only read-only fan RPM and no voltage control. Everything is driven from a native GTK4 desktop app (xe-gpu-gui) and scriptable CLI helpers.

Overclocking now works too. The VF (voltage-frequency) curve — the one control Windows has and stock xe lacks — is unlocked by a small xe_gt_oc patch that issues the PCODE begin → write → end transaction the stock driver omits. Read/undervolt/overvolt the 86-point curve from sysfs. It was never a fused-off Pro-SKU lock — just a missing driver transaction. See docs/OVERCLOCKING.md.

Key finding: the Arc Pro B60 (8086:e211) does not need the missing MEI late-bind fan firmware for manual control. Intel's fan-control patch (series 168027) programs the user fan table straight over the FAN_SPEED_CONTROL PCODE mailbox, and the Pro card's PCODE accepts it directly. Confirmed working end-to-end on a real B60 — see docs/EVIDENCE.md.

Related upstream issue: intel/compute-runtime #885.

What works (verified on Arc Pro B60, 8086:e211, kernel 7.0.0 / Ubuntu 26.04)

Capability	Stock xe	With this toolkit
Read fan RPM / temps	✅	✅
Write fan speed (pwm1)	❌	✅
10-point fan curve (pwm1_auto_point*)	❌	✅
Graphical fan-curve editor	—	✅ built-in (xe-gpu-gui, draggable points)
GPU power cap (TDP)	✅ sysfs	✅ + persistent helper
GPU clock min/max limits	✅ sysfs	✅ + persistent helper
Voltage-frequency curve (undervolt/overclock)	❌	✅ xe_gt_oc patch + xe-gpu oc
VRAM (GDDR6) memory-speed overclock	❌	✅ xe_gt_oc patch (oc/mem_speed) + xe-gpu oc mem
GPU temperature (throttle) limit	❌ (read-only)	✅ xe_gt_oc patch (oc/temp_limit) + xe-gpu oc temp
Idle power/heat optimization	❌ (idles at 1200 MHz)	✅ (idles at 400 MHz)
All-sensor temp/health monitor	raw sysfs	✅ xe-gpu-temps (table/watch/json)
Single-command status dashboard	—	✅ xe-gpu (fan+clocks+power+temps)
Native desktop GUI	—	✅ xe-gpu-gui (GTK4: animated metric dashboard, fan-curve editor, VF-curve overclock tab, stability test)
Live VRAM-usage metric	⚠️ per-client only (fdinfo); device-total needs root debugfs	✅ xe-gpu-vram.service → device-total GUI tile, no root
Overclock benchmarking	—	✅ FPS, VRAM bandwidth, compute (TFLOPS) + real LLM tokens/sec (OpenVINO GenAI), compared to a saved stock baseline
Silent-corruption guard	—	✅ LLM-output coherence check flags an unstable memory OC that keeps tok/s high but produces gibberish
No repeat password prompts	—	✅ optional polkit rule — the GPU-control helpers run without a prompt for a local admin
Multiple GPUs	—	✅ per-card selector; all tabs + writes re-target the chosen card; stock baseline + benchmarks keyed per-card
Survives reboots	—	✅ systemd + failed-boot watchdog (a bad OC can't cause a boot loop)
Survives kernel updates	—	✅ auto-rebuild hook

Desktop app — xe-gpu-gui

A native GTK4 / libadwaita control panel (xe-gpu-gui, or "Arc GPU Dashboard" in your apps menu). Three tabs; all writes go through pkexec. install.sh also drops in a scoped polkit rule so those helpers run without a password prompt for a locally logged-in admin (delete /etc/polkit-1/rules.d/49-xe-gpu.rules to go back to per-action prompts). Full walkthrough in docs/GUI.md.

Fan Control	Overclock

• Dashboard — a live (2 s) monitor:
  • Specifications row — fixed values: device, VRAM size, power cap, power limit, clock limits, hardware range, power profile, fan mode, and Power/Temp/Voltage-limited ✓/✗ indicators.
  • Metrics — animated tiles with a custom icon (gauge / bolt / fan / RAM chip / thermometer), a value that tweens to each reading, a scrolling sparkline, and a radial ring for the percentage metrics. Covers GPU frequency (+actual), card & GPU power draw (derived from the energy counters), power %, fan speed & duty, temperature %, and VRAM used / usage.
  • Temperatures — every sensor (GPU / VRAM / Mem-ctrl / PCIe + all VRAM channels) with its own sparkline; the number + line turn red near the crit limit, 🔥 on the hottest.
  • Metrics filter — a modal to show/hide any metric or sensor; the choice is saved across launches.
  • GPU selector — on a multi-GPU box, a header dropdown picks which card everything targets.
• Fan Control — a draggable graphical fan-curve editor (10-point, presets, live temp marker) plus Auto (stock table) and Max buttons.
• Overclock (needs the xe_gt_oc patch) — a VF-curve editor in two modes (uniform offset or per-point drag-the-nodes curve), with aligned slider+spinbox controls for voltage offset, voltage limit, power, memory speed and temp limit; a colour-zoned curve graph; preset profiles (Stock / Efficient / Balanced / Performance); save/load your own named profiles; and a Stability test that runs a GPU load with the fan ramped to max and auto-reverts if the card hangs.
  • Benchmark (opt-in) — the test can also measure FPS, VRAM bandwidth, compute (TFLOPS) and real LLM tokens/sec (prefill + decode, via OpenVINO GenAI on the GPU), then show a table comparing this run to your saved stock baseline (▲/▼ per metric). A Stock bench button records that baseline; the app offers to run one if you benchmark an OC without it. A coherence check on the LLM output catches an unstable memory overclock that keeps tok/s high but silently corrupts results — treated as a failure and reverted.
  • Stock defaults, the stock baseline, and every benchmark are keyed per-card, so the tool works across different GPUs (it reads each card's own stock values instead of assuming the B60's).

How it works

• Fan control is Intel's kernel patch series 168027 (drm/xe/hwmon, by Karthik Poosa) — not yet in mainline. This repo bundles the CachyOS-7.1.2-adapted patch (which also applies to Ubuntu 7.0.0 with fuzz) plus automation to build it as an out-of-tree xe.ko, and userland helpers on top. Fan curve = FAN_SPEED_CONTROL PCODE: per-point FSC_WRITE_FAN_TABLE (0x1) with temp | speed<<8, then FSC_WRITE_NUM_FAN_CONTROL_POINTS (0x0) to commit; pwm1_enable selects full-speed(0)/manual(1)/auto-stock(2).
• Power/clock tuning uses only driver-exposed sysfs (.../gt0/freq0/*, hwmon power1_cap) — no patch, no PCODE poking, safe.
• Overclocking (VF curve) is a separate, self-contained xe_gt_oc patch (kernel/ + scripts/apply_xeoc.sh) that adds .../gt0/oc/vf_curve. Writing the curve is a PCODE transaction (begin 0x5f/2 → 86× point write 0x5d/0xa → end 0x5d/0xb); the stock driver omits the begin, so the writes are rejected. Details + how it was derived: docs/OVERCLOCKING.md.

Quick start (Ubuntu / Debian-ish)

# 1. build & install the patched xe module for your running kernel
#    (verified recipe — uses the kernel's real config + correct vermagic so the
#     module BINDS; full detail in docs/LINUX-BUILD.md, DKMS auto-rebuild-on-
#     kernel-update in dkms/README.md)
sudo bash scripts/build-xe-module.sh --build-only   # optional: build + verify only
sudo bash scripts/build-xe-module.sh && sudo reboot # build, install, reboot to activate

# 2. install the userland helpers + GUI (one command; re-run after every `git pull`)
sudo bash install.sh
# ...or install pieces individually:
sudo install -m755 scripts/xe-fan-curve.sh  /usr/local/bin/xe-fan-curve
sudo install -m755 scripts/xe-gpu-tune.sh   /usr/local/bin/xe-gpu-tune
sudo install -m755 scripts/xe-gpu-temps.sh  /usr/local/bin/xe-gpu-temps
sudo install -m755 scripts/xe-gpu.sh        /usr/local/bin/xe-gpu
sudo install -m755 scripts/xe-gpu-oc.sh     /usr/local/bin/xe-gpu-oc      # overclocking CLI
sudo install -m755 scripts/xe-fan-rebuild.sh /usr/local/sbin/xe-fan-rebuild

# optional: unlock voltage-frequency-curve overclocking (adds .../gt0/oc/vf_curve)
sudo bash scripts/apply_xeoc.sh && sudo reboot   # see docs/OVERCLOCKING.md
sudo install -m755 kernel-hook/zz-xe-fan-rebuild /etc/kernel/postinst.d/zz-xe-fan-rebuild
sudo cp systemd/etc/*.conf /etc/
sudo cp systemd/*.service /etc/systemd/system/
sudo systemctl daemon-reload
sudo systemctl enable --now xe-fan-curve.service xe-gpu-tune.service
# if you installed the OC patch (below), also persist the voltage curve across reboots:
sudo systemctl enable --now xe-gpu-oc.service

Full step-by-step (with the module install/reload, .zst gotcha, and verification): docs/INSTALL.md.

Usage

# native desktop GUI (GTK4/libadwaita): live dashboard + graphical fan-curve editor
xe-gpu-gui             # or launch "Arc GPU Dashboard" from your apps menu

# one-stop dashboard + front-end (wraps the tools below)
xe-gpu                 # status: card + clocks + power + fan + temps in one view
xe-gpu watch           # live dashboard
xe-gpu fan set 45:80 55:130 65:180 75:220 85:255   # -> xe-fan-curve
xe-gpu tune set --power-w 150                        # -> xe-gpu-tune
xe-gpu temps                                         # -> xe-gpu-temps

# fan
sudo xe-fan-curve show
sudo xe-fan-curve set 45:80 55:130 65:180 75:220 85:255   # temp°C : pwm(0-255)
sudo xe-fan-curve auto        # hand back to stock auto table
sudo xe-fan-curve max         # full speed

# power / clocks
sudo xe-gpu-tune show
sudo xe-gpu-tune set --power-w 150 --clk-max 2000 --clk-min 400
sudo xe-gpu-tune reset

# overclocking — voltage/memory/temp (needs the xe_gt_oc patch, see docs/OVERCLOCKING.md)
sudo xe-gpu-oc read             # dump the 86-point curve + memory speed + temp limit
sudo xe-gpu-oc offset -25       # undervolt every point 25 mV (cooler / more efficient)
sudo xe-gpu-oc offset 25 1050   # overvolt +25 mV, capped at a 1050 mV ceiling
sudo xe-gpu-oc curve 0:820 10:880 ...   # write a full/partial custom curve (one transaction)
sudo xe-gpu-oc mem 20000        # GDDR6 memory speed, Mbps (20000 = 20 Gbps)
sudo xe-gpu-oc temp 95          # GPU thermal-throttle target, °C
sudo xe-gpu-oc profile save daily   # save current voltage/memory/temp as a named profile
sudo xe-gpu-oc profile load daily   # re-apply it (also: profile list / delete)
sudo xe-gpu-oc reset            # restore stock curve + memory + temp

# temperatures / health (read-only, no patch needed)
xe-gpu-temps            # table of every sensor + limits, fan, power
xe-gpu-temps watch 2    # live refresh every 2s
xe-gpu-temps json       # machine-readable (for scripts / dashboards)

Overclocks reset to stock on cold boot; choices persist to /etc/xe-gpu-oc.conf and are re-applied by xe-gpu-oc.service. On a multi-GPU box, set ARC_GPU_BDF=<pci-address> to target a specific card (the GUI does this automatically). VRAM usage needs xe-gpu-vram.service (installed + enabled by install.sh).

• GUI fan curves via the built-in editor (xe-gpu-gui) — see docs/GUI.md.
• GPU tuning details — see docs/GPU-TUNING.md.
• Persistent config: /etc/xe-fan-curve.conf, /etc/xe-gpu-tune.conf.

The full story

docs/NOTES.md — the project journal: how fan control and overclocking were reverse-engineered (DTrace on the live Windows driver → the missing PCODE 0x5f/2 begin transaction), every hardware fact learned (VF-curve monotonicity, the Vmax rail, the firmware VR lock, what telemetry Linux does/doesn't expose), the dead ends, and the traps.

Windows

A Windows port lives under windows/. It provides the same fan curves, power/temperature limits, VF-curve overclocking, live telemetry, and a boot service — but built on Intel's Graphics Control Library (IGCL / ControlLib.dll) instead of the Linux xe sysfs + PCODE patches, so there's no kernel driver to sign. It ships an arc-gpu CLI and an arc-fan-service Windows service. See windows/README.md to build/use it and windows/PORT.md for the Linux→Windows capability map and roadmap.

Kernel updates

The patched xe.ko is replaced by the stock driver on a kernel update. The included /etc/kernel/postinst.d/zz-xe-fan-rebuild hook (calling xe-fan-rebuild) rebuilds it automatically when the matching linux-source is installed. Major kernel jumps may need apt install linux-source-<newver> and a patch re-fuzz. (True DKMS is not possible — xe can't build against headers-only; it needs the full kernel source + i915 siblings.)

Contributing upstream

The real fix is series 168027 landing in mainline. The highest-impact thing you can do is add a Tested-by for your card to that series on intel-xe@lists.freedesktop.org, and confirm on compute-runtime #885. Templates + full workflow: contrib/.

Credits & license

• Fan-control kernel patch: Intel — series 168027, Karthik Poosa (intel-xe@lists.freedesktop.org). GPL-2.0. Bundled here under patch/ with its kernel lineage intact.
• B580 groundwork + CachyOS patch adaptation: PerkyZZ999/XeDriver_FanPatch.
• This repo's automation, helpers, tuning, and Pro-card (B60/B70) enablement: see LICENSE (MIT for the scripts). The bundled kernel patch remains GPL-2.0.

Not affiliated with or endorsed by Intel. Use at your own risk — see docs/EVIDENCE.md for exactly what was tested.