Skip to content

arm64: dts: apple: t6050-k116s: High Power mode - #2214

Open
axboe wants to merge 1 commit into
aurora-silicon:integration/aurorafrom
axboe:k116s-high-power
Open

axboe wants to merge 1 commit into
aurora-silicon:integration/aurorafrom
axboe:k116s-high-power

Conversation

@axboe

@axboe axboe commented Nov 18, 2026 •

Copy link
Copy Markdown
Contributor

v3 (2026-11-22): #2784 merged the binding and the driver side (apple,high-power-mode on the CLPC node, set for the J616s and the J714s). What is left here is one commit: the K116s sets the property and keeps the CLPC's apple,smc link the record is sent over. Rebased on integration/aurora bb9f8ec.

Without it the SMC's fan policy on the K116s keeps its Automatic ceiling whatever the profile: under a sustained load both fans level off at 62 % of their maximum while the PMP throttles to hold the die.

Measured on a K116s (firmware 27.2) with the first two versions of this PR, which differ from #2784 only in how the mode is selected. Four minutes of openssl speed -multi 18 -evp aes-256-gcm per profile:

low-power balanced performance
Fans 0 rpm 3900 / 4200 rpm 6250 / 6750 rpm (maximum) after about 90 s
Package power (PHPC) 14.0 to 14.8 W not limited not limited
PMP effort, lane 2 (0 to 255) 0 57 to 66 37 to 44
Throughput, sustained 50 to 54 GB/s 111 to 113 GB/s 116 to 118 GB/s
TCMb 50 to 62 °C 108 °C 108 °C
System power 21.5 W 84 to 89 W 94 to 96 W

On #2784's driver (bb9f8ec plus this commit and my K116s stack, 2026-11-22): the profile registers as apple-clpc with low-power, balanced and performance. In performance, two minutes of the same load took the fans from 0 to 6135 / 6683 rpm (maximum 6250 / 6750), past the 62 % they level off at without the record. The table above was not rerun.

dtbs_check adds no warning for the K116s.

@RyanTheTide

Copy link
Copy Markdown
Contributor

Thanks for bringing High Power mode to the K116s.

We select behaviour by a hardware description rather than a list of machines, and this turns the single machine string into a list: apple_clpc_t6050_high_power[] = { "apple,j714s", "apple,k116s" } (drivers/soc/apple/clpc.c:2032-2033, used at :2049 and :1509-1510 through of_machine_compatible_match()). Every new board would need another driver edit.

Whether a machine offers High Power mode is a property of the board (its cooling and power envelope), so it belongs in the board's device tree. Could you replace the list with a boolean on the clpc node, for example apple,high-power-mode, set from t6050-j714s.dts and t6050-k116s.dts, and keep the existing apple,smc link check? That also lets the other T6050/T6051 boards that macOS gives High Power mode opt in with a DT line only.

@RyanTheTide

Copy link
Copy Markdown
Contributor

Following up on the DT property note above with the test that would show the mode working on the K116s, since High Power mode changes behaviour only under sustained load:

  1. /sys/class/platform-profile/*/choices lists performance; write it to profile, and the clpc debugfs state shows smc high power: on with the record count going up by one. Setting balanced again shows off.
  2. 10 minutes of stress-ng --cpu 0 in balanced, then 10 minutes in performance from a cool start, logging every 5 s: fan speeds (macsmc hwmon fan*_input), the package power the driver reads (PHPC) and scaling_cur_freq of each cluster.
  3. Pass: over minutes 5-10, performance runs the fans above the balanced plateau and the mean cluster frequency is higher, with no thermal shutdown and no SMC write errors in dmesg.

The same log on a J714s would confirm the existing behaviour there is unchanged. Removing the &clpc { /delete-property/ apple,smc; } hold in t6050-k116s.dts also turns the low-power profile's package power limit on for this board, so a short run of step 2 in low-power (package power held near 15 W) would cover that too.

Once that is in we can run no-regression rows on our A18 Pro and M4 Pro, which use the same driver.

@axboe

axboe commented Nov 19, 2026

Copy link
Copy Markdown
Contributor Author

Ryan, thanks. v2 is pushed: apple,high-power-mode on the performance controller's node (binding added), set in t6050-j714s.dts and t6050-k116s.dts, and the driver reads it in place of the machine list. The apple,smc link check stays. The commits are ordered binding, device trees, driver, so the J714s keeps the mode at every step.

Tests at the v2 head (K116s, firmware 27.2):

  • The switch. choices lists low-power balanced performance. From the controller's debugfs status after each write:
    performance: smc high power: on (3 records sent)
    balanced:    smc high power: off (4 records sent)
    performance: smc high power: on (5 records sent)
    
    No SMC or clpc line in dmesg --level=err,warn.
  • low-power under load. Four minutes of openssl speed -multi 18 -evp aes-256-gcm, sampled every 5 s: PHPC between 14.0 and 14.8 W against the 15 W limit, the clusters' ceilings moving between 1644 and 1992 MHz, fans at 0 rpm, TCMb 50 to 62 °C. So the limit the SMC link turns on holds on this board.

balanced and performance under load were measured on the first version (same load, table in the description): fans 3900 rpm against 6250 rpm, PMP effort 57 to 66 against 37 to 44, throughput up about 4 %. I did not repeat them at v2: the only change is how the mode is selected, and the switch check above covers that.

I used a fixed openssl load for four minutes per profile in place of ten minutes of stress-ng: the fans and the PMP's effort reach their plateau within about 90 s, and it gives a throughput figure. If you want the ten-minute form as well, say so.

I have no J714s to run the no-regression row on.

@RyanTheTide

Copy link
Copy Markdown
Contributor

Thanks, v2 is exactly the shape we asked for, and the switch and low-power numbers cover the selection change. Accepting the openssl load and the v1 balanced/performance table is fine.

One coordination note: #2310 adds the same apple,high-power-mode property (same binding file) and sets it on every T6050 MacBook Pro, plus a per-SoC flag for whether the SMC takes the record. The two overlap, so whichever lands first, the other rebases to its remaining hunks. If this one goes first, #2310 just adds the other boards; if #2310 goes first, this PR is the K116s device tree change (the apple,smc link). Either way we will put this up for review.

For the J714s, the property keeps today's behaviour; a J714s owner's switch check (smc high power: on/off in the clpc status) would confirm it, but it does not need to come from you.

macOS offers High Power mode on the 16-inch M5 Pro. Say so on the CLPC
node, and keep the CLPC's SMC link the record is sent over, which was
held back here until it had been tried.

On a K116s the SMC on firmware 27.2 takes the record as the J714s's
27.0 does: in the performance profile the fans run up to their maximum
under a sustained load, where they level off at 62 % of it otherwise.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
@axboe axboe changed the title soc: apple: clpc: High Power mode on the K116s arm64: dts: apple: t6050-k116s: High Power mode Nov 22, 2026
@axboe

axboe commented Nov 22, 2026

Copy link
Copy Markdown
Contributor Author

Rebased on bb9f8ec after #2784: the binding and driver commits are gone, and what is left is the one K116s device tree commit (the property, and keeping the CLPC's SMC link). Description updated.

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.

2 participants