[Laptop 13 Pro] Core Ultra Series 3: EC power table sets PL4 = 80 W in all modes, clamping the Arc B390 at 1900 MHz

[bpavlo]

System Model or SKU

Framework Laptop 13 (Intel® Core™ Ultra Series 3)

BIOS Version

3.02

DIY Edition Information

Memory: Micron CT32G75C2LP5X 32GB
Storage: Crucial T710 1TB

Port/Peripheral Information

Peripheral vendor and name:

1. [Fill in device / expansion card]
2. [Fill in device / expansion card]
3. [Fill in device / expansion card]
4. [Fill in device / expansion card]

Standalone Operation

Mainboard only, Standalone mode enabled

Describe the Bug

The Arc B390 never exceeds ~1900 MHz under GPU load although the driver reports rp0_freq = 2500 MHz.
The only throttle reason the SoC raises is PL4; measured package power is 15–30 W and PL1, PL2,
thermal and PROCHOT are never active. The cause is the EC power table for this board, which writes
PL4 = 80 W over PECI in every power state.

EC console (ectool console) after PD negotiation:

PL1:35, PL2:47, PL4:75, PSYSPL2:75 updated success    # 79 W supply, no battery
PL1:35, PL2:56, PL4:80, PSYSPL2:89 updated success    # 96 W Apple adapter, no battery

These match zephyr/program/framework/sakura/src/cpu_power.c (branch fwk-sakura-20260429,
commit 853d767b "fwk: sakura: Add power limit setting"):

/* DC mode Battery only */     pl1_watt = 35; pl4_watt = 80;
/* Standalone mode AC only */  pl2_watt = MAX(40, MIN(60, active_power*60/100));
                               pl4_watt = MIN(80, active_power*95/100);
/* AC DC mode */               pl1_watt = 35; pl2_watt = 60; pl4_watt = 80;

GPU behaviour under a 60 s Vulkan load (vkmark 1920x1080), 1 Hz samples of
/sys/class/drm/card0/device/tile0/gt0/freq0/ plus RAPL package energy:

supply	EC PL4	GPU act_freq	package power	throttle/reasons
79 W	75 W	1750–1850 MHz	11–16 W	pl4 (60/60 samples)
96 W	80 W	1850–1900 MHz, peak 2050	13–17 W	pl4 (59/60 samples)

Intel enforces PL4 preemptively (datasheet 15.1.1.1: "the PL4 power limiting algorithms will
preemptively limit frequency to prevent spikes above PL4"), i.e. as a frequency ceiling derived from
worst-case power, which is why the clamp holds at 15 W of real draw. A 16-thread CPU load in the same
session reaches 36–40 W at 96 °C, so PL1/PL2 and cooling behave as configured; PL4 is the only limit
that is mis-sized for this SoC.

Comparison with the Core Ultra Series 1 table (marigold/src/cpu_power.c, same branch):

power state	Series 1 (marigold) PL4	Series 3 (sakura) PL4
battery only	80	80
adapter only, no battery	0.95 × adapter (95 on 100 W)	min(80, 0.95 × adapter)
adapter ≥ 55 W, battery ≥ 30 %	120	80
adapter ≥ 55 W, battery < 30 %	min(0.9 × adapter + 80, 120)	80

Steps To Reproduce

1. Connect the 100 W (or any ≥ 60 W) USB-C supply, with or without battery.
2. sudo ectool console | grep PL1 → shows PL4:80 (or PL4:75 on a ~80 W supply).
3. Start any GPU load (vkmark, or a game) and run
   watch -n1 cat /sys/class/drm/card0/device/tile0/gt0/freq0/act_freq /sys/class/drm/card0/device/tile0/gt0/freq0/throttle/reasons
4. Observe act_freq ≤ ~1900 and reasons = pl4, while sudo turbostat (or RAPL energy_uj)
   shows package power far below 80 W.

Expected Behavior

PL4 sized to what the source can deliver, as on Series 1: ~95 W on a 100 W contract without battery,
and 120 W with adapter + charged battery, so the B390 can reach its rated frequency. A BIOS toggle
would also be much appreciated.

Screenshots

report-20260910-201810.log
report-20260910-203004.log

Operating System

NixOS 26.11

Linux Kernel Version

7.2.3

Additional Context

No response