Fan, Temp and GPU-Util not being reported

[WizardlyBump17]

davi@davi:~$ xpu-smi
+----------------------------------------------+----------------------------+------------------------+
|            Intel XPU-SMI v2.0    Driver: 1CF0B7A60BF9DB210B10E48    Level Zero: 1.28.0             |
+----------------------------------------------+----------------------------+------------------------+
| GPU  Name                    Persistence-M   | Bus-Id            Disp.A   |   Volatile Uncorr. ECC |
| Fan              Temp  Pwr:Usage/Cap         | Memory-Usage               |   GPU-Util  Compute M. |
+----------------------------------------------+----------------------------+------------------------+
|   0  Intel(R) Arc(TM) B580   Off             | 0000:09:00.0      Off      |               Disabled |
|      Graphics                                |                            |                        |
| N/A               N/A  35W / 190W            | 11610MiB / 12216MiB        |      N/A       Default |
+----------------------------------------------+----------------------------+------------------------+

+-------+--------+--------+------------------+-------------------------------------------------------+
| Processes:                                                                                         |
|   GPU      PID    Type    Process Name                                            GPU Memory Usage |
+-------+--------+--------+------------------+-------------------------------------------------------+
|     0     6971     C      ./llama-server                                                 11584 MiB |
+-------+--------+--------+------------------+-------------------------------------------------------+

As you can see, Fan, Temp and GPU-Util are not being reported. When running with sudo, Temp and GPU-Util gets reported. The temperatures and fan speeds are reported on /sys/class/drm/renderD128/device/hwmon/hwmon2, being readable by normal users. It would be nice if it didnt require sudo xpu-smi to read those values.

xpu-smi installed from https://github.com/intel/xpumanager/releases/tag/v2.0.1

[savery42]

Hi @WizardlyBump17 thanks for the report! A few interesting things -

---

It would be nice if it didnt require sudo xpu-smi to read those values.

This is an unfortunate consequence of the level-zero Sysman runtime Linux security practices

The default security provided by the accelerator driver is to permit querying and controlling of system resources to the UNIX user root, querying only for users that are members of the UNIX group root and no access to any other user. Some queries are permitted from any user (e.g. requesting current frequency, checking standby state).

We can certainly have a hwmon fallback for fan and temp. I'll create an internal ticket and would expect it to land in the next release (v2.0.2) on August 12th given it is trivial.

Also of interest - we do not have a fallback for fan speed in event that the max RPM is not available, such as the case with the B580 / GPUs with firmware-controlled fans. I'll also make the change to report current RPM if we cannot determine the %max fan speed instead of N/A as it's confusing to report that on systems with active cooling.

I'll also file an internal ticket to see if it is possible to read the VBIOS fan tables at init to be able to report utilization % properly to be consistent with the smi tools of the other vendors. Unsure if that will make it into the 2.0.2 release, but it will be slated for work.

---

If this is your personal system and you'd like immediate fixes for Temp/Pwr/Utilization% you can grant the following permissions to bypass the Sysman restriction:

# Give read access to users in the `video` group to fan/temp/pwr telemetry
sudo chown root:video /sys/class/intel_pmt/telem{3,4}/telem 
#  Allow xpu-smi binary to read system performance counters
sudo setcap cap_perfmon+ep $(which xpu-smi)`

The missing fan speed will still require a software change.

---

I may have some time today to provide a patch / alpha build and will update this issue in a new comment accordingly.

[savery42]

Hi @WizardlyBump17 this was a fun rabbit-hole - thanks!

The hwmon fallback patch will likely not land in out next release (v2.0.2) as we will target the fixes upstream in the level-zero Sysman runtime (as to avoid needing fallbacks for every shortcoming).

For now, I have created custom, no guarantee deb packages (Ubuntu 24.04/26.04 base) to solve the problem without modifying system/application permissions. (the packages also include the fix in #145 for correct memory reporting in xpu-smi [ '' | 'ps' ])

24.04: xpu-smi_2.0.2~customhwmon+26.31.ecb87f2-1~24.04_amd64.deb.zip

26.04: xpu-smi_2.0.2~customhwmon+26.31.ecb87f2-1~26.04_amd64.deb.zip

> sha256sum output/*hwmon*                                               
43d72d62575b62984f4512e2be2fca04ea4d093e7bb402b0459897498302d985  output/xpu-smi_2.0.2~customhwmon+26.31.ecb87f2-1~24.04_amd64.deb
841f65e1f3bdbc15f346e7f850a7282426d676c2e388defbd8f0c8094020d705  output/xpu-smi_2.0.2~customhwmon+26.31.ecb87f2-1~24.04_amd64.deb.zip
4d482e5dbe26bc655bc6f230735ce275d5f899814e00759194dc1bddd5c16cb8  output/xpu-smi_2.0.2~customhwmon+26.31.ecb87f2-1~26.04_amd64.deb
38ca2692b7beacaed08411a1a6cfcb48048900d870e0137de543a315b6c2d30a  output/xpu-smi_2.0.2~customhwmon+26.31.ecb87f2-1~26.04_amd64.deb.zip

---

Side note: the fan speed may still not display properly on the B580, though it may be firmware and/or kernel dependent. I would see hwmon*/*/fan1_input reporting an uncorrected pulse count instead of the (expected) RPM - e.g. ~310,000 at 37C and ~291,000 at 75C.

As the current patch is just a 'for the fun of it' I didn't spend much time triaging that outside of internal notes. If you try the custom build and it shows 0 RPM, or you also see a very large number in your hwmon/*/fan1_input then I'll open up a ticket for a kernel patch - though for now I'm hoping it's only an issue with the custom kernel or firmware version that I am running, as the RPM count is working on a B70 Pro system of mine.

---

I am also happy to provide a .patch if you would like to maintain a custom build - just let me know :).

Otherwise, I will update this issue again once the changes (read fan/temp/power without elevated privs) land in an official release.

[thevisad]

@savery42 thanks for linking these — the temperature and fan parts of this issue are corrected by the two PRs on Battlemage/xe.

Mapping to what @WizardlyBump17 observed on the B580 (normal user: Fan / Temp / GPU-Util all N/A; with sudo: Temp and GPU-Util appear, Fan still N/A):

Temperature (xpum-temp-hwmon): Temp shows under sudo, so Level Zero has it but gates it behind privilege — a normal user gets N/A. This PR reads the world-readable PCI hwmon pkg/vram nodes when Level Zero exposes no matching sensor or denies it to an unprivileged process, so a normal user gets the reading without sudo. Genuine device-loss / runtime errors are still propagated (never masked with a stale value).
Fan RPM (xpum-fan-rpm): fan isn't reported even with sudo, i.e. Level Zero exposes no fan handle for this device at all — the hwmon fanN_input node is the only source. This PR reads it (per fan, keyed by id) and reports Fan Speed (RPM) / a fan.speed_rpm metric, for normal users. Validated on an Arc Pro B70 (same Battlemage family), which likewise exposes no Level Zero fan handle.
GPU-Util: not addressed by these two PRs. Utilization comes through a path that needs CAP_PERFMON, which is why it only appears under sudo — a privilege/packaging concern rather than a missing sysfs read. Options there (out of scope for these changes) would be setcap cap_perfmon+ep on the installed binary or a perf_event_paranoid adjustment.
Both PRs read the same world-readable hwmon node the report points at, are condition-based (not PCI-ID-gated), ship doctest unit tests for the source-selection logic and parsing, and are DCO-signed.

@WizardlyBump17 — if you can try these branches on your B580, an unprivileged xpu-smi stats should now show GPU Core/Memory Temperature and Fan Speed (RPM) matching the raw hwmon values (where stock xpu-smi shows N/A). Confirmation would settle the temp/fan portion of the issue.

Same caveat applies to #148 on the B580 — it reads fan1_input directly, so it surfaces that same uncorrected pulse count instead of true RPM. Without a known pulses-per-revolution/scale for that tach, the miscalibrated value can't be reliably converted in userspace, so agreed the correction belongs in the kernel/driver (it reads true RPM on the B70).