Repository navigation
Description
Hi everyone! Over the last week I went after idle power on my M1 Pro (J314s, t6000), comparing against macOS on the same machine, and I'd like to hand over what I found in one place. Part of it is in pull requests now. The rest is described here with a reference branch, in case it is useful for whoever picks up power on the T600x boards.
Where it stands: awake, on battery, nothing attached, screen off, the machine went from about 2.70 W to 1.00–1.02 W. macOS measured 0.87 W in the same state, back to back. The figures are the SMC's system power and battery current times voltage, which agree within a couple of percent, and the comparisons are alternating on/off windows, usually ten rounds of 45 s.
The largest item is the DP-to-HDMI converter. It is powered all the time and draws 0.62 W with nothing plugged in (2.71 → 2.09 W with dp2hdmi-pwren low). The connector's HPD still reaches the SoC with the converter off, so it can be powered only while a display is attached, which is what macOS does. That is #510, as a draft and limited to this board.
PCIe L1 substates on the internal links are worth about 0.32 W (3.02 → 2.70 W). That is the same finding as #489 for the J514S, and my numbers are in a comment there.
The AXI2AF bridges of USB-C ports that have not seen a cable since boot cost about 0.2 W after a boot with nothing attached (1.21 → 1.02 W). That one is #509, as a draft.
Wi-Fi power save is next. brcmfmac sets pm2_sleep_ret to 2000 ms when cfg80211 passes no timeout, so with power save on the radio stays awake for two seconds after every frame, which on a normal network means always. With 200 ms the machine drew 0.155 ± 0.02 W less (1.24 → 1.08 W). That is on per-window medians, where all ten rounds landed between 0.12 and 0.19 W. On window means my tool called it unclear (0.11 ± 0.11 W), because the 2 s samples have 2–4 W spikes in them. macOS has the link in deep sleep 61 % of the time. I only did this from userspace, with the wireless-extensions call that can set the timeout, and only on battery. I have no kernel patch for it.
macOS's PMGR driver applies bridge tunables to SIO and to the three display front ends (disp0, dispext0, dispext1) that Linux does not. Writing them gave 0.146 ± 0.04 W (1.43 → 1.28 W), and the hardware idle counters moved with it: AFI idle went from 0 to 70–77 % and the PMP from always on to 40–50 %. They are lost in suspend-to-RAM and have to be written again after resume.
CPU PD's 10 ms target residency costs 0.13 ± 0.03 W against 200 us. Same finding as #497, numbers are there.
Gating the ANE cluster while nothing uses it gave about 0.09 W, and powering off the media blocks that have no driver bound (venc, scaler, jpeg, scodec) about 0.045 W.
Releasing the USB-C PHY, dwc3 and display crossbar of an empty port through runtime PM measured 0.026 ± 0.011 W on the battery meter for the usb and common domains of three ports. The second meter did not reach significance (0.016 ± 0.017 W), so I read it as real but small. Switching every ATC domain off by hand gave an upper bound of about 0.065 W.
Withdrawing the DISPEXT request of an external display processor that has no display I could not separate from the noise between two boots (about 0.04 W), which fits what CoreyH saw for these requests on the same board in #274. It did turn up one thing worth knowing if someone does this on T600x: with the request withdrawn, the first USB-C plug after boot failed with dcp_dptx_connect: failed to assert Type-C HPD: -110. The firmware's first DPTX power-up answered that HPD after 1.5–1.8 s, past the 1 s the driver waits, where it took 0.3 s with the request held. Waiting up to 4 s for that one HPD fixed it.
That leaves about 0.14 W against macOS that I could not attribute.
The kernel changes I actually run are on a reference branch: https://github.com/Ante042/linux/commits/j314s-display-usbc-idle-reference. It sits on the 7.3 suspend-to-RAM reference tree I shared earlier (t600x-s2r-reference), so it does not apply to integration/aurora as it is. It has the HDMI converter change with its HPD and reclaim fixes, a DISPEXT request withdrawal for T600x, and release-on-unplug runtime PM for the ATC PHY, dwc3-apple and the display crossbar (off by default, one module parameter each). That last part has been my daily kernel since the 19th with two USB-C displays, a dock and HDMI, through suspend and resume. Thunderbolt and USB4 are untested because I have no such device, and in two of five sleeps the afnc1_lw1 domain logged that it stayed off after resume with everything working, which I have not explained.
The register-level items (bridge tunables, ANE gate, media blocks, Wi-Fi timeout) are done by a small script on my machine. I haven't written kernel patches for them, and the bridge tunables in particular probably belong in the pmgr driver at domain power-on. I'm happy to share the script and the register tables, measure anything on the J314s, or test patches from others.
The machine is a MacBook Pro 14" M1 Pro, apple,j314s / apple,t6000, asahi,os-fw-version 13.5 (asahi,system-fw-version reads unknown), Omarchy, m1n1-aurora 1.6.1.aurora23, running a local 7.3.0 kernel built from the reference tree above, so none of these numbers are from a released Aurora tag. Not tested: other T600x boards, Thunderbolt, and s2idle, since this kernel uses suspend-to-RAM. #290 looks at s2idle on the same board, and this is the awake side of it.
Thanks for all the work on this kernel!
Activity
Ante042 commentedon Nov 22, 2026
Hi! Kernel work has moved to aurora-silicon/linux, so I've moved this write-up there as aurora-silicon/linux#2838, with the two pull requests as aurora-silicon/linux#2836 and aurora-silicon/linux#2837. Closing it here so there is one place to follow. Thanks!