Skip to content

j575c (Mac Studio M4 Max): test results #1498

Open

Description

@7robertodantas

Current status (2026-11-14): see #1498 (comment). Support page: aurora-silicon/aurora-silicon.github.io#39.

Mac and installed build

  • Mac Studio (2025), M4 Max, Mac16,9, board J575c, SoC T6041. 16 CPUs (4 E + 6 P + 6 P), 48 GB.
  • /proc/device-tree/compatible: apple,j575c, apple,t6041, apple,arm-platform
  • Aurora Installer v0.1.0-alpha.19, profile Arch (ALARM) minimal, encrypted root (installer default encrypt=1).
  • uname -r: 7.3.0-1-26-ARCH. Packages: linux-aurora 7.3.0.aurora1-26, aura 1.0.0b8-2,
    uboot-aurora 2026.07.aurora7-2 (U-Boot e1ef8b63), asahi-scripts 20261023-5.
  • Local changes (these results depend on them):
    1. Stage 2 (m1n1/boot.bin): U-Boot replaced with a local build of 250385f8 (same
      apple_m1_defconfig, built with a different compiler: Homebrew aarch64-elf-gcc 16.2). The stock e1ef8b63 panics on T6041 (see logs).
    2. Stage 2 DTB: linux-aurora 15f88e1c plus two t6041-j575c.dts changes (PR linked below): a DockChannel UART node, and gpio-hogs releasing the GL3590 hub.
    3. chosen.aura,serial-console=1 appended to Stage 2 (debug console).
    4. Workarounds: end0 held in promiscuous mode by a small systemd unit. On this test machine only, the LUKS root also unlocks with a key file on /boot.
  • Debugging was done over macvdmtool reboot debugusb from an M1 Pro MacBook Pro (USB 2 cable). Channel 1 is the aura/U-Boot/kernel console.

Both firmware versions and provisioning

asahi,os-fw-version: 26.6.2
asahi,system-fw-version: unknown   (aura logs System FW mBoot-20457.41.3, macOS 27.2.1 26B131)
Own-Mac provisioning: Wi-Fi/BT calibration loaded from platform data (contents not shared)

Actions and results

Feature / action Result Details
Boot, stock install fail U-Boot e1ef8b63: Unsupported SoC → panic before GRUB. No picture, so it looked like a dead keyboard
Boot, with U-Boot 250385f8 pass aura → U-Boot → GRUB → Linux → systemd → first-boot text wizard → login. 2 clean reboots + several VDM resets came back
First-boot wizard pass Completed through tty1. Disk re-keyed, user created. The wizard only runs on tty1 (invisible here, see display)
CPU pass 16 CPUs, 3 cpufreq policies: E 1020–2592 MHz, P 1260–4104 MHz
RAM / NVMe pass 46 GiB usable. APPLE SSD AP1024Z, SMART readable
PCIe enumeration pass BCM4388 Wi-Fi/BT, GL9755 SD, AQC113 10GbE, ASM3142 xHCI all enumerate
ASM3142 (port 3) pass Enumerates and xhci binds without a pwren-gpios/gP16 entry
GL3590 hub → back USB-A / front USB-C, stock DT fail Hub held in reset: ASM3142 root ports "not attached", nothing plugged in enumerates
Same, with the reset hogs (DT change 2) pass Hub 05ac:8009/05ac:800a enumerates at boot. USB mic and a USB HID device work through it. The mic records real audio (3 s, 48 kHz)
Back USB-C, port next to Ethernet (hpm0, back-right) partial Device mode only so far (debug cable to another Mac); macvdmtool debugusb works through it
Back USB-C 2nd (hpm1, back-right-middle) partial USB keyboard works (host). USB 2 only: USB3: apple,tunable-axi2af: -ENOENT, USB2 only
Back USB-C 3rd (hpm2, back-left-middle) partial A USB hub (with monitor hubs behind it) switches the port to host/source and enumerates at USB 2 high speed. No USB 3
Back USB-C 4th (hpm3, back-left) fail Power only: an iPhone charges, but no USB or Type-C event at all (webcam and iPhone not seen); only 3 Type-C ports registered. The port isn't described in the DTS (ADT hpm3 at 0x6 / usb-drd3 / atc-phy3, IRQs 89/95/97)
Back USB-A ×2 (GL3590 hub ports 3, 4) pass (with hub fix) USB mic records audio, HID device works
Front USB-C ×2 (GL3590 hub ports 1, 2) pass (with hub fix), USB 2 iPhone (high speed) on front-left = hub port 1, a USB HID hardware key (full speed) on front-right = hub port 2. USB 3 not tested (no USB 3 device yet)
Ethernet AQC113 (atlantic) fail, workaround Link 1000/full, TX reaches the LAN, but unicast/broadcast RX is dropped (no DHCP lease, ARP fails, rx_dropped grows). Works in promiscuous mode: DHCP, ping 0.8 ms, SSH
Wi-Fi BCM4388 partial Firmware and calibration load, scan sees networks. Connect not tested yet
Bluetooth partial Powered, scan finds devices. Pairing not tested
HDMI fail iBoot hands over a 640×1136 dummy framebuffer. aura DCP sees no display (connected:0), then did not receive HELLO. apple-t6041-display: no display subsystem
GPU fail Unsupported GPU firmware image (reported macOS version 26.6.2). The driver supports Apple GFX 25F84 for T6041
SEP / Touch ID fail (expected?) apple_sep: no platform profile for this SoC. Its units time out, adding ~3 min to every boot
Fans / thermal partial 60 s 16-thread load: die max 44 → 76 °C, 23 → 75 W total system power. Fans steady ~990 rpm. Ramp at higher temperatures not tested
Headphone jack not tested card registered (AppleJack), nothing plugged in
Debug console (DockChannel, DT change 1) pass earlycon + ttyDC0 over debugusb. Input works (U-Boot and GRUB were driven over it)
pacman keyring fail Fresh install: aurora/asahi-alarm repos fail signature checks. /etc/pacman.d/gnupg had no local key, and the trustdb marked the Aurora key ultimately trusted while the key was missing. Fixed with pacman-key --init + --populate

Graphics and session details

No graphical session (minimal profile, no display). Mesa 26.3.0-devel present but renders on softpipe/llvmpipe only; no DRM render node. Suspend not tested.

Relevant logs and related reports

# stock U-Boot e1ef8b63 (uboot-aurora 2026.07.aurora7-2)
U-Boot 2026.07-ge1ef8b636fc9 (Nov 08 2026 - 05:46:48 +0000)
Model: Apple Mac Studio (M4 Max, 2025)
DRAM:  47.2 GiB
Unsupported SoC
### ERROR ### Please RESET the board ###
# cause: build_mem_map() in e1ef8b63 matches apple,t6040 only. 250385f8 adds apple,t6041.

# aura 1.0.0b8 stage 1
display: waited 2000 ms for display status
display: connected:0 timing_cnt:0 color_cnt:0
FDT: Unable to find alias atcphy0   (… atcphy7, usb4-*)

# kernel
phy-apple-t6040-usb2 382a90000.phy: USB3: apple,tunable-axi2af: -ENOENT, USB2 only
aurora 28a600000.gpu: supported: Apple GFX 25F84 for T6041 (normalized loaded image)
aurora 28a600000.gpu: Unsupported GPU firmware image (reported macOS version Some([26, 6, 2]))
apple_sep: no platform profile for this SoC; the SEP stays disabled
NetworkManager: dhcp4 (end0): activation: beginning transaction (timeout in 45 seconds) … state changed no lease

# GL3590 hub control pins before the fix (AP pinctrl, ADT /arm-io/spi3/usb3hub-flash/usb3hub-fw)
pin 127 usb3hub_rst        output, low
pin 128 usb3hub_fwload_done input,  high  (goes low once the hub has loaded)
pin 130 usb3hub_flash_oe   output, low

DT changes tested (against linux-aurora 15f88e1c): both are in a PR, linked below.
aurora-diag check --expect / diff after the hub change: 8 improvements. The diff also lists 5 "regressions"; these
are gpiochip renumbering and the per-boot hash in the GPU firmware message, not functional changes.

Untested features and limitations

  • USB storage throughput and USB 3 speeds on any port, SD card, headphone jack, internal speaker, USB-C/TB display, Wi-Fi connection,
    Bluetooth pairing, suspend: not tested yet (results will follow as comments).
  • macvdmtool reboot serial (s5l UART): no output, but only USB 2 cables were available. Not conclusive.
  • The 4th back USB-C port (hpm3) is not in the DTS. Not tested.
  • U-Boot is a local build, not the official package binary.

AI disclosure: this testing and report were done with an AI assistant (Claude); the owner reviewed and ran the hardware steps.

Activity

  1. 7robertodantas commented on Nov 11, 2026

    @7robertodantas
    ContributorAuthor

    DT changes from this report (DockChannel UART, GL3590 hub reset hogs, and the pinctrl binding change they need): #1499

  2. 7robertodantas commented on Nov 11, 2026

    @7robertodantas
    ContributorAuthor
  3. 7robertodantas commented on Nov 11, 2026

    @7robertodantas
    ContributorAuthor

    GPU follow-up: with #1539 (J575c 25G83 firmware entry, plus the J614c 40-core fix already on integration/aurora) and mesa-aurora 20261110, the GPU works headless on this unit: vulkaninfo and eglinfo report Apple M4 Max (G16C B1), glmark2 --validate gives 0 failures, and vkcube runs. HDMI still has no picture: the display subsystem isn't described for the J575c yet (LPDPTX PHY, no T604x driver).

  4. 7robertodantas commented on Nov 11, 2026

    @7robertodantas
    ContributorAuthor

    Display and USB follow-up: with #1546 and aurora-silicon/aura#160 this unit gets a picture over USB-C (back port 2 from the right edge, USB-C to HDMI adapter, 2560x1440 Linux console from a cold boot and after a replug), USB3 on the four back Type-C ports (10 Gb/s and 5 Gb/s measured; SuperSpeed is orientation-dependent for some devices, see the PR), and a working fourth back port. HDMI and display on the other three back ports are not done yet.

  5. 7robertodantas commented on Nov 11, 2026

    @7robertodantas
    ContributorAuthor

    HDMI follow-up: with #1565 (on top of #1546) and aurora-silicon/aura#160, the HDMI port shows the Linux console at 2560x1440, from a cold boot and after a replug, and together with a second monitor on USB-C (back port 2). dcpext0 drives it through the LPDPTX PHY. Still to do: HDMI audio and displays on the other back USB-C ports.

  6. 7robertodantas commented on Nov 11, 2026

    @7robertodantas
    ContributorAuthor

    Audio and USB-A follow-up:

  7. 7robertodantas commented on Nov 11, 2026

    @7robertodantas
    ContributorAuthor

    CPU performance follow-up: under an all-core load this J575c ran at about half speed. The CPU clusters were held at 2352/1020/1260 MHz and the whole Mac at about 40 W, with the die at only 56 C. The cause was the board DT's guessed CPU thermal zone (a power allocator with no switch-on trip and no sustainable-power), with the PMP left off. #1599 turns on the PMP and uses the J614c's CPU thermal zones (same T6041). With it, all-core load holds 4104/2592/4104 MHz at 86 W with the die at 90 C, and throughput doubles (2.1x in stress-ng). glmark2 is up 10%, or 27% with the CPU loaded as well. Reboot and s2idle pass. Not reached: the fan ramp point (the fans stayed at their 1000 RPM minimum up to a 94 C die).

  8. 7robertodantas commented on Nov 12, 2026

    @7robertodantas
    ContributorAuthor

    USB3 follow-up: the "orientation-dependent SuperSpeed" limitation from #1546 is actually "attached at boot". A USB3 device that is plugged in while the Mac boots stays at USB2 (SuperSpeed stuck in RxDetect) until it is replugged; on a fresh plug both orientations train Gen 2. The Type-C retimers are set up by the port controllers when a partner attaches, so a connection inherited from the boot chain keeps a retimer set up before the host PHY ran. macOS avoids it by attaching the devices again as it starts. Details in #1670. #1671 renegotiates such connections at probe (the J716c mechanism, with the J575 key, gated on an attach-state byte), on the second to fourth back ports. With it, an SSD attached across a reboot and across a cold boot runs at Gen 2x1 (926 MB/s) with no replug; the keyboard hub, HDMI, USB-A and a USB-C DP adapter keep working. Not tested: the first back port (debug cable here) and USB4/Thunderbolt partners.

  9. 7robertodantas commented on Nov 12, 2026

    @7robertodantas
    ContributorAuthor

    Video follow-up. Hardware decode (AVD, VA-API through libva-v4l2-request-avd) works as shipped: H.264, HEVC Main and Main 10, VP9 profiles 0 and 2, and AV1 8/10-bit decode bit-exact against software (per-frame MD5, 1080p clips and libaom test streams). An AV1 stream that changes resolution mid-stream stops at the size change. The video encoder and both JPEG engines were disabled in the J575c DT. With them enabled the JPEG engines work as they are (47.6 dB at quality 90), but apple-ave refused the encoder firmware: the M4 Max gets its own build of AppleAVE2FW 9003.78.0 with a larger DATA segment and one TEXT word that changes every boot. #1677 accepts that build and enables the nodes. With it, H.264 and HEVC 1080p30 at 8 Mb/s give 43.7 and 44.5 dB, HEVC Main 10 from P010 works, and encoding works again after s2idle. Not tested: the second encoder (not described), long or concurrent encodes, and other M4 Max boards.

  10. 7robertodantas commented on Nov 14, 2026

    @7robertodantas
    ContributorAuthor

    Status as of 2026-11-14. Every J575c pull request from this report is merged. This table replaces the
    earlier follow-ups as the current state. The support page update is aurora-silicon/aurora-silicon.github.io#39.

    Tested on the J575c with integration/aurora (the GPU and ray tracing rows also on d89d8c0).

    Area Result Pull requests / notes
    Boot pass U-Boot 250385f8 needed instead of the alpha.19 payload's (separate report)
    CPU pass 16 cores, cpufreq, idle states (#1575); all-core load holds 4.1 GHz with the PMP and thermal zones (#1599)
    Display, USB-C pass all four back ports (#1546, #1871, aurora-silicon/aura#160); one DP adapter fails link training on the second port from the right, also under macOS
    HDMI pass 2560x1440 (#1565); once it stayed dark after a resume until replugged
    Speaker pass with speakersafetyd (#1573, aurora-silicon/alsa-ucm-conf-asahi#15, aurora-silicon/alarm-aurora#267)
    Headphone jack pass aurora-silicon/alsa-ucm-conf-asahi#15
    USB-A pass both ports, cold and warm boot (#1499, aurora-silicon/alarm-aurora#266)
    USB-C, USB 3 pass 10 Gb/s on all four back ports, both orientations, also with a device attached at boot (#1546, #1671)
    Front USB-C partial USB 2 works (#1499); USB 3 not tested
    Port 3 power (gP16) described #1857; the link is already up at handoff, so normal boots do not use it; no regression warm, cold or across suspend
    Ethernet pass no promiscuous mode needed (#1593)
    Wi-Fi pass joined on 5 GHz, DHCP, DNS, HTTPS over IPv4 and IPv6 with Ethernet unplugged
    Bluetooth partial powers up and scans; pairing not tested
    Sleep pass suspend to idle, keyboard and RTC wake (#1575); no deep sleep on T6041 yet
    Fans pass SMC ramps them from about 90 C (#1599)
    SD card pass SDR104, write protect; every hot removal logs "Card is consuming too much power!" but completes
    GPU, Vulkan pass glmark2, vkcube (#1539)
    Ray tracing partial ray queries from compute and ray-tracing pipelines render correctly (Aurora gpubench scenes). A ray query in a fragment shader loses the device: aurora-silicon/mesa#330, fixed by aurora-silicon/mesa#333 together with #1920 (both open). With both, the #837 CTS selection runs to the end: 51,190 pass, 1 crash, 2 timeouts, none of them memory (follow-up).
    Video pass decode H.264, HEVC, VP9 and AV1 (8/10-bit) bit-exact; encode H.264 and HEVC, JPEG (#1677)
    NFC probes #1858; identity only, no user-facing reader on this Mac
    Thunderbolt / USB4 not done

    A GPU note: on the -91 test kernel, ray-tracing benchmark processes run back to back lost the device once the
    lower VA window filled up (each one left about 335 MB behind). On d89d8c0 that case passes 10 of 10, so it is
    already fixed on integration.

    Thanks to everyone who reviewed and merged these.

  11. 7robertodantas commented on Nov 14, 2026

    @7robertodantas
    ContributorAuthor

    Ray tracing follow-up. The fragment-shader ray query that lost the device in the status table is
    aurora-silicon/mesa#330. On the 40-core GPU, the render ray-query scratch for both shader stages is one contiguous
    359 MiB block, but the largest free hole in the kernel's 1984 MiB lower window is 352 MiB even at idle. Long CTS
    runs also filled that window with idle compute pools (319 MiB each after a ray query). Two fixes, both open:

    With both, the #837 ray tracing CTS selection (CTS 1.4.7.1, 74,833 cases, 1 job) runs to the end in 45 minutes:
    51,190 pass, 23,640 skip, 1 crash, 2 timeouts. Before, it aborted. None of the remaining failures is memory:
    ray_query.stress.geometry_shader.triangles takes about 10.2 s and is cut off by the 10 s render watchdog
    allowance, and ray_tracing_pipeline.build.cpu.level_instances.{aabbs,triangles}_1048576_1_1 (host builds of a
    million instances) exceed the runner's 120 s timeout. The exact #1920 branch also passes a 2,893-case sample three
    times in a row with no failures. Not tested: other M4 Max boards and GPUs with fewer cores.

    Ray tracing stays partial in the table until both are merged.

  12. RyanTheTide commented on Nov 19, 2026

    @RyanTheTide
    Contributor

    @7robertodantas, thanks again for the J575c results here. One read-only ADT value would let us add the J575c's NOR flash/nvram node: the SPI NOR node under /arm-io/spi* (the aura proxy, print(u.adt["/arm-io/<node>"]), or macOS ioreg -l -w0 -p IODeviceTree -n <node>). With it, nvram listing on Linux (read-only) becomes a row we can ask for later.

    Two more J575c rows will follow once their device-tree changes exist: HDMI/DP audio (aplay -l listing both cards, a silent stream) and one ProRes encode on each engine. No need to do anything for those yet.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions