Skip to content

ERT_CMD_STATE_ERROR on Strix Halo with firmware 1.1.2.65 #103

Open

Description

@nickbaxter18

Describe the Issue
We are experiencing an ERT_CMD_STATE_ERROR (specifically: aie2_send_mgmt_msg_wait: command opcode failed, status 0x4000003, followed by aie2_hwctx_start: Map host buffer failed, ret -22) when attempting to execute operations on the Strix Halo NPU.

Hardware & Environment

  • Hardware: Strix Halo
  • Firmware Version: 1.1.2.65 (amdnpu/17f0_11/npu.dev.sbin)
  • Driver / Error log context:
amdxdna 0000:c6:00.1: aie2_send_mgmt_msg_wait: command opcode 0x115 failed, status 0x4000003
amdxdna 0000:c6:00.1: aie2_map_host_buf: hwctx 15 addr 0x7422cc000000 size 0x4000000
amdxdna 0000:c6:00.1: aie2_hwctx_start: Map host buffer failed, ret -22

Question
Is firmware version 1.1.2.65 known to have compatibility issues with current IRON deployments? What specific firmware version is recommended/needed for Strix Halo platforms to resolve this mapping buffer failure?

Activity

  1. AngryLoki commented on Apr 17, 2026

    @AngryLoki
    Contributor

    1.1.2.65 firmware user here - works fine for me for IRON. Kernel 6.19.11-gentoo-dist, Strix, XRT master build (~2.23.0), with master xdna-driver (driver + shim + firmware).

    Regarding diagnosis, I highly recommend to run and add xrt-smi validate (not just xrt-smi examine) command to README (similar to what we see in flm guide, even though they have their own validator). It is pointless to expect IRON to work if XRT infra setup is not complete.

    Regarding combinations of working setups:

    1. All-master setup usually works (unless there is some transition which you can observe in recent commits, i. e. temporary breakage). Don't forget to reload the driver!
    2. Last time I checked, IRON had issues with in-source driver, but I guess that's a common knowledge.
    3. XRT 2.21.75 + shim from XRT 2.21.75 + xdna-driver from 2.21.75 tag works.
    4. But apparently that's not how it was intended by xdna-driver developers - looks like the driver should be taken from 1.7 branch, not 2.21.75 tag. But anyways, now it works.

    flm has slightly better "validation":

    • Checks kernel version and requires kernel 6.17+ or 7.x+.
    • Iterates over /dev/accel/accel0 through /dev/accel/accel15. For each device found:
      • Queries the DRM version - requires amdxdna version >= 0.6.
      • Queries the NPU firmware version - requires firmware >= 1.1.x.
      • Check the NPUs has at least 8 columns (Phoenix and Hawk Point will fail)
    • If the memlock limit is below 100 MB, it attempts to auto-raise it to infinity via setrlimit. If that fails, it reports an error.
  2. andrej commented on Apr 17, 2026

    @andrej
    Collaborator

    Agreed with @AngryLoki, if xrt-smi validate does not work this is IRON-unrelated. To test if this is a memory limit issue, try running as superuser (sudo) or raise the limit using ulimit.

    The CI runners currently have XDNA at commit beb9e450fe123ecdf395453971576179cedcf1dd (and the corresponding firmware and XRT submodule) installed, so this is a known-good version.

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