DEV Community

Ismael Alex Otero
Ismael Alex Otero

Posted on

Two faults, one stutter: MacBookPro16,1 on Omarchy

Field note from one machine, 17 November 2026. The desktop felt like it was
dropping frames. Two separate faults were doing it at once. The CPU was pinned
at 800 MHz by a stuck external PROCHOT pin, and Hyprland was drawing every
frame on the Intel GPU and copying it onto the Radeon, which is the chip the
panels are actually wired to.

This is one afternoon on one laptop. It is not a general recommendation to
disable PROCHOT, and it is not a recommendation to keep the Radeon awake on
every T2 Mac.

Machine

Model Apple MacBookPro16,1 (16-inch, 2019, T2)
CPU Intel Core i7-9750H, 6 cores / 12 threads, 800 MHz–4.5 GHz
iGPU UHD Graphics 630, PCI 00:02.0, 8086:3e9b
dGPU Navi 14, PCI 03:00.0, 1002:7340 (the 5300M in this model)
Internal panel Apple Color LCD, 3072×1920 @ 60 Hz, on the Radeon (eDP-1)
External panel Acer XV271, 1920×1080 @ 240 Hz, on the Radeon (DP-5)
OS Omarchy, kernel 7.2.15-arch1-1-t2
Desktop Hyprland 0.56.2, started by uwsm, SDDM autologin

Both connectors that were lit, the internal panel and the Acer, belonged to
the AMD card. The Intel card had no connected display.

Fault 1: the CPU would not leave 800 MHz

Every thread sat at 800 MHz. A 12-thread busy loop did not move them. The
package was 66°C. The performance power profile was selected, turbo was
allowed, the frequency cap was 4.5 GHz, and the AC adapter was online. Chassis
fans were already around 4400 RPM, so it looked like a thermal problem.

It was not the CPU's own throttle. MSR_POWER_CTL (0x1FC) read 0x6c005d.
Bit 0, the bidirectional PROCHOT enable, was set. On all 12 threads,
IA32_THERM_STATUS (0x19C) and IA32_PACKAGE_THERM_STATUS (0x1B1) had
bit 2 set (PROCHOT is asserted) and bits 0 and 4 clear (the CPU is not at its
own thermal or critical limit). Package power limits were 100 W and 125 W.
Clock modulation was off. Hardware P-state request already allowed the top
ratio.

Clearing bit 0 of 0x1FC on every logical CPU, and leaving every other bit
alone, let the same busy loop reach about 3.5 GHz within a second. After that
the cores idled and boosted in the 3.6–4.3 GHz range at 74–79°C. One package
sample read 100°C during the jump and then fell back with the rest. The CPU's
own thermal-status bit stayed clear.

Firmware puts bit 0 back on boot and again on resume, so the clear has to be
reapplied. The hook used here refuses to write if bit 0 or bit 4 of either
thermal MSR is set. That leaves the CPU's own overheat brake in place. It does
not leave the external pin in place: a real PROCHOT from the SMC, a VRM, or
the GPU will also be ignored until the next boot if those thermal bits are
clear. That is the tradeoff, and it is why this should stay a per-machine
workaround rather than a default.

The T2 Linux hybrid-graphics guide
describes freezes, a fan ramp, and sudden shutdown (CPU CATERR) on this
generation when the AMD GPU is unhappy. A stuck PROCHOT pin is one of the ways
a hot or confused SMC gets the CPU's attention. Clearing it removes that
signal. Watch the temperatures.

Fault 2: Hyprland drew on the Intel GPU

At startup Aquamarine enumerated one GPU, /dev/dri/card1, driver i915, and
made it primary. A moment later udev added /dev/dri/card2, driver amdgpu.
The log then said the Radeon backend was starting with primary /dev/dri/card1.
The late card never became the renderer. It only received the finished frames,
because that is where the connectors are.

Over one second, Hyprland's Intel render engine accumulated more time than the
wall clock, while the Radeon's graphics engine was around 11% and its shader
clock sat on the 200 MHz step. The desktop was being painted on the small GPU
and copied across the PCI bus onto the chip that scans out, including onto a
240 Hz panel. The Radeon looked idle because it was.

Hyprland chooses that primary from the DRM devices that exist when it starts.
Listing the Radeon first in AQ_DRM_DEVICES does nothing if the node is not
there yet. The fallback becomes primary, and the hotplug stays a secondary
display device.

Telling them apart

These two show up together and feel like one stutter.

The CPU fault is a clock that will not rise. Read /proc/cpuinfo during a
short all-core burn. If every thread is still at the minimum (800 MHz on this
chip) and the package is far from 100°C, read 0x19C and 0x1B1. Bit 2 set,
with bits 0 and 4 clear, is an external PROCHOT pin. Bit 0 or bit 4 set means
the CPU itself is hot, and the pin should be left alone.

The GPU fault is in the Hyprland log. "Found 1 GPUs" followed by an Intel
backend, then an amdgpu backend "with primary" pointing at the Intel card,
while hyprctl monitors shows the live panels on the AMD connectors. The
compositor is copying.

What was changed

PROCHOT, reapplied at boot and after wake:

  • /usr/local/sbin/disable-bd-prochot reads the two thermal MSRs and, when bits 0 and 4 are clear, clears only bit 0 of 0x1FC on each logical CPU.
  • disable-bd-prochot.service runs that at multi-user.target.
  • /usr/lib/systemd/system-sleep/disable-bd-prochot runs it on the post sleep action, because resume restores the firmware value.
  • /etc/modules-load.d/msr.conf loads the msr module so /dev/cpu/*/msr exists early enough.

GPU order, applied on the next Hyprland start:

# /etc/udev/rules.d/69-drm-gpu-names.rules
KERNEL=="card*", KERNELS=="0000:03:00.0", SUBSYSTEM=="drm", DRIVERS=="amdgpu", SYMLINK+="dri/radeon"
KERNEL=="card*", KERNELS=="0000:00:02.0", SUBSYSTEM=="drm", DRIVERS=="i915", SYMLINK+="dri/intel"
Enter fullscreen mode Exit fullscreen mode

/dev/dri/cardN numbers move between boots. /dev/dri/by-path/pci-0000:03:00.0-card
is stable and still unusable here: Aquamarine splits AQ_DRM_DEVICES on every
colon, and the PCI path contains two. The symlinks avoid both problems. On
this Hyprland 0.56.2 the symlink was resolved; the log said the explicit device
/dev/dri/card2 was found. An older report
(hyprwm/Hyprland#6997)
described symlinks being ignored. Check the log for "Explicit device … found"
after changing this.

AQ_DRM_DEVICES has to be in the environment before Hyprland starts.
hl.env() in the Lua config runs too late. Under uwsm the file is
~/.config/uwsm/env-hyprland, because XDG_CURRENT_DESKTOP is Hyprland:

export AQ_DRM_DEVICES="/dev/dri/radeon:/dev/dri/intel"
Enter fullscreen mode Exit fullscreen mode

The first entry is the renderer. The Intel node stays in the list so a missing
Radeon can still bring the session up.

A user drop-in waits for the name to appear, up to five seconds, then starts
Hyprland anyway:

# ~/.config/systemd/user/wayland-wm@hyprland.desktop.service.d/wait-radeon.conf
[Service]
ExecStartPre=%h/.config/uwsm/wait-for-radeon
Enter fullscreen mode Exit fullscreen mode

wait-for-radeon polls /dev/dri/radeon every tenth of a second and exits 0
whether or not it showed up. systemctl --user daemon-reload picks the drop-in
up. The value is not applied to a compositor that is already running. This
session was restarted with uwsm stop; SDDM's autologin brought it back.

Result

The new log enumerated both cards from the explicit list, started the amdgpu
backend first, and recorded "gpu /dev/dri/card2 becomes primary". The renderer
line was:

AMD Radeon Graphics (radeonsi, navi14, ACO, DRM 3.64, 7.2.15-arch1-1-t2)
Enter fullscreen mode Exit fullscreen mode

The Intel backend then started with the Radeon as primary. The same two panels
came back at the same modes. With the desktop idle, the Radeon sat near 7% on
its 200 MHz shader step, which is a quiet compositor rather than a copy engine.
The cores were boosting in the upper 3 GHz range. Subjectively the machine
felt much faster. That is the CPU no longer being stuck at its floor, and the
240 Hz panel no longer being fed through a copy from the other GPU.

What was not retested

The PROCHOT service was started once, after the bit had already been cleared,
and logged that BD PROCHOT was already disabled. A reboot, and a
suspend/resume cycle, were not done after the hook was installed. The sleep
hook is the standard post path, and the register is documented to come back
on those events, but this note does not claim those paths were observed.

No heavy GPU workload was measured after the switch. The claim is that the
Radeon became the renderer, which the log shows.

Prior art, and the opposite advice

Stuck BD PROCHOT at the minimum ratio is an old laptop failure. ThrottleStop
and several Linux MSR scripts clear the same bit. Nothing about that mechanism
is new.

The T2 Linux guide, and at least one Omarchy write-up for this exact model
(mbp2019-omarchy), go the other
way: force the Intel GPU primary with apple_gmux.force_igd so the Radeon can
sleep. That is the right goal when the dGPU is only burning power, and it is
the path those pages document for the freezes and shutdowns. On this boot the
panels were already on the Radeon, so it was powered either way. The waste was
painting on the other chip and copying. Forcing Intel primary would be a
different setup, with its own mux change, and was not tried.

Omarchy's own note on hybrid GPUs covers the colon-splitting trap in
AQ_DRM_DEVICES and the udev names. The piece that note does not spell out,
and the reason a correct AQ_DRM_DEVICES can still leave you on the Intel
renderer, is the probe order on this Mac: the Radeon appears after Hyprland
has already chosen.

Top comments (0)