[joske]

On the M4 Max, a sustained all-core load under Linux runs into the critical trip and powers the machine off. My J614c (MacBook Pro 14", M4 Max) did that during a Mesa build. Two things combine:

1. No CPU energy model. The T6041 CPU OPP tables have no opp-microwatt, so the power allocator has no power actors, and cpu-die-thermal never throttles anything. That applies to the new GPU zones (arm64: dts: apple: t6041-j614c-j616c: add the GPU temperatures and zones #1447) too. On the J614c the clocks stayed at 4.1 GHz up to the 114 °C test kill. This is the T6040 problem that 76815fa fixed for the M4 Pro.
2. Trips that aren't Apple's. The J614c/J616c zone used a passive trip at 113 °C and a critical trip at 118 °C. Both ADTs carry the M4 Pro's die limits: temp-sensor and fast-die-ctrl-loop are byte-identical to the J614s's, with a 110 °C target and a 125 °C maximum (sochot). The PMP fast-die loop is off on these boards, so this zone is the CPU's only throttle.

Commit 1, t6041: add the CPU energy model: this takes T6040's measured energy model, since the cores are the same. Three P states that only T6041 has (3828, 3888 and 3948 MHz) are interpolated in frequency, and the tables stay monotonic. The model is in t6041.dtsi, so it also covers the J575c, which hasn't been tested.

Commit 2, t6041-j614c-j616c: hold the CPU die at the ADT target: this runs the power allocator the way the M4 Pro zones do, at the ADT target. Switch-on is at 75 °C and control at 110 °C, with a 20 s integral constant. sustainable-power is 48 W. Hot is 120 °C, report only, and critical is the 125 °C sochot level after three samples.

Why 110 °C and 48 W, measured on the J614c. I ran the same all-core NEON FMA load on macOS and Linux and read the whole SMC key table on both. On Linux I used a read-only debug module.

	macOS	Linux, old zone + energy model	Linux, this PR
CPU power (PHPC)	48–51 W	~33 W	—
TCMz	105–110 °C	105–116 °C, swinging	109–110 °C, steady
Fans	~4950 rpm	2317 rpm (minimum)	4000–4350 rpm

The SMC fan loop follows the dDT* hot spot, not TCMz. dDT* reads up to 16 °C above TCMz at high power. If the zone holds the CPU around 33 W, dDT never gets hot enough for the fans to ramp, and the low airflow then keeps the power down. With control at 110 °C and macOS's sustained power, the fans ramp on their own, as they do on macOS.

Testing on the J614c. linux-aurora was built from a82a1eb plus #1442. The DTB was built from this branch and installed through update-m1n1. There were no sysfs overrides.

• Four minutes of 16-thread NEON FMA load: TCMz was held at 109–110 °C from 30 s on, and no sample went above 110 °C. The fans ramped to 4000–4350 rpm, and the hot and critical trips were never reached.
• Before this PR, without the energy model, the same load ran into the 118 °C critical trip and the hardware-protection power-off.
• With only the energy model, the old 113 °C trip with a 48 W budget overshot to 114–116 °C, and the fans stayed at their minimum.

Known limits:

• Uneven throttling. The single die zone throttles the clusters unevenly: under load P1 stayed at 4.1 GHz while P0 dropped to 2.4–2.9 GHz. The M4 Pro's per-cluster zones avoid that, and that's a follow-up.
• 48 W is a 14" measurement. The J616c has more cooling, so 48 W is on the safe side for it.
• J616c and J575c are not booted with this. For the J616c, its ADT's temp-sensor and fast-die-ctrl-loop match the J614c's.
• The PMP. With the M4 Pro's PMP and PMC nodes enabled, the PMP starts fine on the J614c, and aura passes 90 tunables. But it doesn't feed the SMC's power keys, and it doesn't change the fans. It's not part of this PR.

🤖 Generated with Claude Code

[joske]

Added a third commit, t6041-j614c-j616c: throttle each CPU cluster on its own sensors (58d5f14). With one die zone controlling all three clusters, the allocator split the budget unevenly under an all-core load: P0 ran at 2.4–2.9 GHz, P1 at 4.1 GHz and E at 1.02 GHz.

What it adds. There is now one zone per SMC core-die sensor, as on the M4 Pro, and each one throttles only its own cluster.

• I loaded one cluster at a time on the J614c to map the sensors: Tpx0-5 follow P0, Tpx8-D follow P1 and Tex0-3 follow E.
• The new zones control at 110 °C. The die zone stays as a backstop, controlling from 115 °C.
• Budgets are 24 W per P cluster and 2 W for E. The two P clusters share one rail meter (PHPC), which reads 25 W with either cluster loaded alone.

Same 4-minute test on the J614c:

• P0 and P1 both held at 3.3–3.8 GHz, and E at 2.59 GHz.
• TCMz stayed at 109–111 °C, with single samples up to 114.
• No hot or critical trip was reached.
• The fans ramped to 4100–4450 rpm.

🤖 Generated with Claude Code