Skip to content

arm64: dts: apple: t8122-j615: Add CPU thermal zones - #491

Open
ericjuta wants to merge 1 commit into
omacom:integration/aurorafrom
ericjuta:j615/cpu-thermal-zones
Open

ericjuta wants to merge 1 commit into
omacom:integration/aurorafrom
ericjuta:j615/cpu-thermal-zones

Conversation

@ericjuta

Copy link
Copy Markdown

Problem

Nothing limits the J615's CPU temperature. With all eight cores busy the die sensors pass 105 °C in 150 s and are still rising; a -j8 kernel build holds 110–114.6 °C for 22 minutes. The J613 got CPU thermal zones for the same reason (30db282). The J615 has the same die, the same six CPU die sensors in t8122.dtsi and no fan either, but no zones.

Change

One file, t8122-j615.dts: the J613's six CPU zones, unchanged. One zone per CPU die sensor, each throttling both clusters through the power allocator; switch-on 58 °C, control 95 °C, critical 110 °C, 6 W sustainable power, E cluster weighted three to one and capped at cooling state 3. The compiled thermal-zones node is identical to the J613's.

I copied the block rather than moving both boards onto a shared include, to keep this off the J613 file and leave each board free to tune its budget. Happy to switch to a shared .dtsi if you prefer.

Tested

On a J615 with these zones applied as a boot-time DT overlay (not yet booted as a kernel-built DTB):

Load Without zones 6 W zones
8 busy cores, 170 s 105.5 °C at 150 s, rising peak 87.7 °C, P cores 3.1–3.6 GHz
-j8 kernel build, 20 min 110–114.6 °C level 91–92 °C, peak 97.2 °C, 6 of 545 samples ≥ 95 °C, P cores 2.9 GHz

8 W was also tried: 96.3 °C peak under the eight-core load with no gain in P-core frequency, so the budget stays at the J613's 6 W.

  • t8122-j615.dtb and t8122-j615-25g83.dtb build with W=1; no new warnings.
  • The built node matches the live device tree of the machine running the overlay, apart from critical-trip hysteresis.
  • tools/aurora-ci/check-commits.py passes.

Not tested

  • CPU and GPU load together.
  • The 110 °C critical trip.

Context: #410 (comment). Independent of #410, which covers the GPU side.

Nothing limits the J615's CPU temperature: with all eight cores busy the
die sensors pass 105 C in 150 s, still rising, and a -j8 kernel build
holds 110-114.6 C for 22 minutes. The J613 got CPU thermal zones for the
same reason; the J615 has the same die, the same six CPU die sensors in
t8122.dtsi and no fan either, but no zones.

Add the J613's zones unchanged: one zone per CPU die sensor, each
throttling both clusters through the power allocator, switch-on 58 C,
control 95 C, critical 110 C, E cluster weighted three to one and capped
at cooling state 3. The compiled thermal-zones node is identical to the
J613's.

sustainable-power stays at the J613's 6 W after load tests on a J615
with these zones applied as a boot-time DT overlay: eight busy cores for
170 s peak at 87.7 C with the P cores at 3.1-3.6 GHz, and a 20 minute
-j8 kernel build holds a level 91-92 C with the P cores at 2.9 GHz (peak
97.2 C, 6 of 545 samples at or above 95 C). 8 W peaks at 96.3 C under
the eight-core load with no gain in P-core frequency.

Not tested: CPU and GPU load together, and the critical trip.

Assisted-by: Claude:claude-fable-5-2
Signed-off-by: Eric Juta <ericjohnjuta@gmail.com>

@scottjones scottjones left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks for this, and for re-measuring the power budget on a J615 rather than just copying it over. Once built, the zones match J613's exactly (only the phandles differ), and the sensor indices line up with the channel order in t8122.dtsi. The 6 W vs 8 W re-measure is a good reason to keep 6 W: 8 W ran hotter with no gain in P-core clock. Approving.

One note, not blocking. Combined CPU and GPU load is still the one way to reach the 110 C critical trip, since the GPU isn't a cooling device in these zones. That's the same as J613 today, and GPU throttling is handled separately by the m3_thermal governor, so it doesn't need anything here. Worth keeping in mind if you get a chance to run a combined load on the J615 later.

@freber

freber commented Nov 17, 2026

Copy link
Copy Markdown

Claude Code wrote this up from my tests.

I ran the combined load you mentioned on my J615 (15-inch M3 Air) with these zones in the boot tree. The kernel is bed544b with #410, this PR and the J615 display and memory floor patches. It also carries a T8122 GPU cap series that lets the GPU use states up to 8 under the m3_thermal governor. Its lowering threshold is left at the default 80 C. On AC with the Performance profile and the lid open.

The load was gpubench w10 at depth 2 on the GPU plus yes on all 8 cores for 300 s, starting from 65 C.

  • The hottest die sensor was Te0P on the E cluster. It was at 95 C within the first minute and stayed there. It peaked at 95.8 C and averaged 94.9 C after the first minute.
  • The zones took the P cluster from 4056 down to 1884 MHz (cpufreq-cpu0 state 15 of 19). Tp0b peaked at 93.7 C and was down to 89 C by the end.
  • The E cluster stayed at 2748 MHz the whole time (cpufreq-cpu1 state 0).
  • The GPU was at state 2 (618 MHz) from the first second and ran at 79.2 fps.
  • The session's kernel log had no asahi errors, faults or timeouts.

So in this case the 110 C trip wasn't close. Every zone has both clusters as cooling devices. The power allocator took the whole cut from the P cluster, so the E cluster zones hold 95 C by slowing the P cores. An earlier run stopped at 47 s because my own guard was set at 95 C, the same temperature the zones hold.

For a GPU-only load the m3_thermal governor does the work on its own. A 600 s w10 run stayed at state 2 and ended at 86.4 C.

@ericjuta

Copy link
Copy Markdown
Author

Thanks @freber, that is the run I was missing. I only tested CPU-only load here and had left combined CPU + GPU load as untested.

Your numbers line up with mine: the zones hold at the 95 C control point. My 20-minute -j8 build at 6 W averaged 91.4 C and peaked at 97.2 C. Your 300 s combined run held 95 C and peaked at 95.8 C, with the P cluster taking the whole cut. So on two J615s the 110 C trip was not close, with or without GPU load.

Two caveats so nobody over-reads it: freber's kernel also carried #410 and a separate GPU cap series, so it is not an isolated test of this PR, and the critical trip itself is still unexercised on J615.

@scottjones thanks for the review. Nothing further from my side; this is ready whenever CI has run.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants