Skip to content

Pixel 8a (akita): CPU DVFS with the 8a's battery-limiter caps, and the GPU - #264

Open
aninmucek wants to merge 3 commits into
derekross:omamobilefrom
aninmucek:omamobile-zuma-akita-speed
Open

aninmucek wants to merge 3 commits into
derekross:omamobilefrom
aninmucek:omamobile-zuma-akita-speed

Conversation

@aninmucek

@aninmucek aninmucek commented Nov 15, 2026 •

Copy link
Copy Markdown
Contributor

The Pixel 8a's CPU frequency scaling, with its own battery-limiter caps, and its GPU. These are the board-file parts of akita's M5 (speed) in omarchy-mobile, plus the GPU table moved into zuma.dtsi for both boards. They follow zuma-husky, on top of #158 (CPU DVFS, thermal) and #157's GPU commits.

  • zuma-akita: Add CPU DVFS with the 8a's caps.
    • The levels: the CPUs point at zuma.dtsi's three tables. The levels and their powers are the 8a's own: stock's pixel-em default profile on this phone lists the same frequencies and microwatts.

    • The 8a's caps: as you said on Pixel 8 Pro CPU DVFS: zuma's CPU levels from the ECT and CPU thermal zones (stacked on #157) #158, its battery current limiter is lower than the 8 Pro's: 1475, 1549 and 2147 MHz on VDROOP1, VDROOP2 and the SMPL warning (Google's zuma-akita-bcl.dtsi). So the board file adds turbo-mode to the six levels above those caps, by node path:

      • little 1548 and 1704 MHz;
      • mid 1572, 1836 and 1945 MHz;
      • big 2294 MHz.

      With boost off, the tops are 1425, 1418 and 2147 MHz.

    • The board file can only add marks, not remove them, so this works because the 8a's caps are below the 8 Pro's on every cluster. zuma.dtsi's comment now says so.

  • zuma: Share the GPU's levels. The GPU's 13 levels to 890 MHz move from zuma-husky.dts into zuma.dtsi, next to the CPU tables, so both boards point at one table. The 8 Pro's DTB is the same before and after: decompiled with dtc -s, identical, phandles included; only the root's node order in the blob differs. Not booted on an 8 Pro.
  • zuma-akita: Enable the GPU to 723 MHz.
    • The change: &cmu_g3d and &gpu as in zuma-husky.dts: ACPM's paired "gpu" clock and CMU_G3D's two gates, pointing at the shared table.
    • The cap: stock's 8a GPU table has the same shader rates as the 8 Pro's, but zuma-akita-bcl.dtsi caps the 8a's GPU at 649 (batoilo), 723 (vdroop1), 580 (vdroop2) and 649 MHz (smpl_gm), all below 890. With no BCL driver nothing runs above stock's caps, so the board file deletes 807, 850 and 890 MHz by path, leaving ten levels from 150 to 723 MHz.

Tested on a Pixel 8a from slot B (7.3.0-rc6; steps 1 to 4 on 317d87dc2 with the commits as first posted, step 5 on 26867fc4e with these three), one change per boot:

  1. omamobile as it is (…g317d87dc2db0):

    • gs101-acpm-tmu: 4 thermal zones, the first time the 8a reads its own SoC temperatures;
    • about 45 °C at idle;
    • no new error lines.
  2. CPU DVFS (…g03e86b1eb814):

    • three cpufreq policies (0-3, 4-7, 8) on schedutil, boost 0;
    • scaling_max_freq 1425, 1418 and 2147 MHz, and the levels above them only under scaling_boost_frequencies;
    • three CPU cooling devices.
    • One core per cluster (openssl speed sha256, against the bootloader's clocks): little +0 %, mid +33 %, big +84 %. Each cluster went to its top under load and back down within a second.
  3. All nine cores at full load for 180 s, logged every second, with a stop at 80 °C:

    • the tops held the whole time;
    • the zones peaked at 71/69/68 °C (big/mid/little), so no trip was reached;
    • the battery went from 33.4 to 35.2 °C;
    • the phone was in hand and "warm but comfortable to hold".
  4. The GPU (…g0b04e364a09c), with omarchy-mobile's package setting clk_acpm zuma_gpu_max_khz=890000 as first posted:

    • panthor binds the Mali-G715 (shader_present=0x1110055, CSF interface v2.8.0) and gives renderD128;
    • devfreq moves between 150 and 890 MHz;
    • Hyprland and the shell render on it;
    • tools/phone-smoke reads GPU ok, with 3 known error lines and 0 new. no regulator (mali) is one of them, as on the 8 Pro.
  5. The GPU at the 8a's cap (…g9cf0719104a1, this head's tree, with omarchy-mobile#832's package at zuma_gpu_max_khz=723000):

    • devfreq offers ten levels, 150 to 723 MHz;
    • under swiping the GPU went to 723 MHz and no higher;
    • tools/phone-smoke reads GPU ok, with 0 new error lines.

What isn't covered:

  • The thermal throttling itself: nothing reached the 80 °C trip in 3 minutes.
  • Boost: it stays off, since there's no BCL driver.
  • The display: the session is clearly faster, but still not smooth. Hyprland scans out through simpledrm, which copies every frame with the CPU. That needs akita's display driver, the 8 Pro's Pixel 8 Pro display: gs-decon, DSIM0 and the HK3 panel at 120 Hz #160 work for the ak3b panel. It's the next milestone.

tools/kernel-pr-check --build OUT --merge origin/omamobile, with the dtschema validator, passes everything:

  • a clean merge onto 26867fc4e;
  • the sign-offs: 3 commits, each signed off by its author, none by an AI;
  • checkpatch --strict with no errors and no warnings;
  • the policy checks;
  • the 10 Google device trees build with W=1 and pass dtbs_check, with no new warnings.

🤖 Generated with Claude Code

@derekross

Copy link
Copy Markdown
Owner

Thank you, and sorry this sat for two days. It's a clean pair, and the order of your runs, with thermal on its own boot before anything could raise the clocks, is how I'd want it done. I checked the caps against Google's zuma-akita-bcl.dtsi on the akita branch: cpucl0/1/2 are 1475, 1549 and 2147 MHz on batoilo, vdroop1, vdroop2 and smpl_gm alike, so your six marks leave exactly 1425, 1418 and 2147 MHz with boost off. Our gate passes on the merge with today's omamobile, CI ran dtbs_check with no new warnings, and only the 8a's DTB changes.

On your two questions, and one more:

  1. (a) or (b) for the CPU tables: (a), as you have it. While every 8a cap is at or below the 8 Pro's it gives the same DTB as tables of its own would, with one table to maintain. Please change the sentence in zuma.dtsi in the same commit: "A board whose BCL caps are lower needs tables of its own". As it stands it says the opposite of what the akita board now does. Something like "A board whose caps are lower marks its extra levels turbo-mode by path, as zuma-akita.dts does" would do.
  2. Where the GPU table lives: in zuma.dtsi, once, with both boards pointing at it, the same way the CPU tables are shared. Two identical copies of 13 levels will drift. Please check that the 8 Pro's DTB comes out the same before and after the move (decompile both and diff), and say so in the commit message, since I won't have an 8 Pro boot of it.
  3. The GPU's top. The same freq_qos blocks also cap the 8a's GPU: 649 MHz on batoilo, 723 on vdroop1, 580 on vdroop2 and 649 on smpl_gm. Your CPU commit takes the view that with no BCL driver nothing should run above stock's caps, and I agree with it. The GPU table and omarchy-mobile#832's zuma_gpu_max_khz=890000 let the GPU run to 890 MHz under all four. The 8 Pro's 890 MHz at least equals one of its own caps (vdroop1); on the 8a it is above every one, on a smaller battery. So the 8a's top is 723 MHz, the highest value any of its four rows allows (vdroop1); I've asked for the same in omarchy-mobile#832's zuma_gpu_max_khz. Here, please take the levels above 723 MHz out for the 8a in the board file, so the DTB holds the limit by itself whatever the module parameter says, and say in the commit message that it follows the vdroop1 cap.

One small thing: the test kernels are ...g03e86b1eb814 and ...g0b04e364a09c, and I can't find either commit on your fork. Could you push them, or give me the blob id of zuma-akita.dts in the GPU build? This head's is 210154a07fac. And if you touch the commits anyway, the trailer form our tree's coding-assistants.rst gives is Assisted-by: LLM.

zuma.dtsi has the three clusters' CPU levels (from the ECT) and their
powers, but only zuma-husky.dts points its CPUs at them, because the
tables' turbo-mode marks are the Pixel 8 Pro's battery current limiter
(BCL) caps. The Pixel 8a's levels and powers are the same as stock's on
this phone (its pixel-em default profile), but its BCL caps all three
clusters lower: 1475, 1549 and 2147 MHz on batoilo, vdroop1, vdroop2
and smpl_gm alike (Google's zuma-akita-bcl.dtsi). With no BCL driver,
mark the levels above those caps turbo-mode too (1548 and 1704 MHz
little, 1572 to 1945 MHz mid, 2294 MHz big), so that with boost off the
tops are 1425, 1418 and 2147 MHz, and point the CPUs at the shared
tables. While every 8a cap is at or below the 8 Pro's, that gives the
same levels as tables of its own would, so zuma.dtsi's comment now says
a board with lower caps marks its extra levels by path.

Assisted-by: LLM
Signed-off-by: DarekM <149845949+aninmucek@users.noreply.github.com>
The GPU's 13 levels to 890 MHz were in zuma-husky.dts. The Pixel 8a's
GPU DVFS table has the same shader rates, so move the table into
zuma.dtsi, next to the CPU tables, for both boards to point at, rather
than keep two copies that can drift. A board whose BCL caps the GPU lower
deletes the levels above its cap by path; the comment says so.

The Pixel 8 Pro's DTB is the same before and after: decompiled with
dtc -s, the two are identical, phandles included; only the order of the
root's nodes in the blob differs. The Pixel 8a's gains the table, unused
until its GPU is enabled.

Assisted-by: LLM
Signed-off-by: DarekM <149845949+aninmucek@users.noreply.github.com>
The Pixel 8a's Mali-G715 is the Pixel 8 Pro's, and its stock GPU DVFS
table has the same shader rates. Enable CMU_G3D and the GPU as
zuma-husky.dts does, with ACPM's paired "gpu" clock and zuma.dtsi's
levels. Google's zuma-akita-bcl.dtsi caps the 8a's GPU at 649 MHz on
batoilo, 723 on vdroop1, 580 on vdroop2 and 649 on smpl_gm, all below
the 8 Pro's 890. With no BCL driver, nothing should run above stock's
caps, as with the CPUs, so delete the levels above vdroop1's 723 MHz,
the highest of them, and the DTB holds that limit whatever the ACPM
clock's module parameter allows. That leaves ten levels, 150 to 723 MHz.
gpu-thermal throttles them.

Assisted-by: LLM
Signed-off-by: DarekM <149845949+aninmucek@users.noreply.github.com>
@aninmucek
aninmucek force-pushed the omamobile-zuma-akita-speed branch from 1184249 to 04945d1 Compare November 18, 2026 20:14
@aninmucek

Copy link
Copy Markdown
Contributor Author

Thank you for checking the caps. You're right about the GPU: I read only the cpucl lines of freq_qos. I've force-pushed the rework onto today's omamobile (26867fc4e), now three commits:

  1. zuma-akita: Add CPU DVFS with the 8a's caps: as before, (a). zuma.dtsi's comment now says "A board whose caps are lower marks its extra levels turbo-mode by path, as zuma-akita.dts does", and the message names all four events.
  2. zuma: Share the GPU's levels (new): the 13-level table moves from zuma-husky.dts into zuma.dtsi, next to the CPU tables.
    • The 8 Pro's DTB is the same before and after: decompiled with dtc -s, the two are identical, phandles included. Only the order of the root's nodes in the blob differs. The commit message says so.
    • No 8 Pro boot of it.
  3. zuma-akita: Enable the GPU to 723 MHz: akita points at the shared table and deletes 807, 850 and 890 MHz by path, citing all four GPU caps and vdroop1's 723 as the highest. That leaves ten levels, 150 to 723 MHz, so the DTB holds the limit whatever the module parameter says.

All three use Assisted-by: LLM now.

Booted on the 8a, with omarchy-mobile#832's package now at zuma_gpu_max_khz=723000:

  • The kernel: 7.3.0-rc6-omamobile-push1-g9cf0719104a1, this head's tree before my sign-off, on slot B by modules-push, then flash.
  • The parameter reads 723000.
  • devfreq offers the ten levels. Under 15 s of swiping the GPU went to 723 MHz and no higher.
  • The CPU tops are unchanged (1425, 1418, 2147).
  • tools/phone-smoke: GPU ok, 0 new error lines.

The earlier test kernels: the GPU build's zuma-akita.dts (0b04e364a) is blob 210154a07fac, the same as the previous head's, so I haven't pushed those commits.

tools/kernel-pr-check --build OUT --merge origin/omamobile passes everything on the three:

  • a clean merge onto 26867fc4e;
  • the sign-offs, each by its author, none by an AI;
  • checkpatch --strict with no errors or warnings;
  • the policy checks;
  • the 10 device trees with W=1 and dtbs_check, no new warnings.

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