AMD Strix Halo ROCm Crashes: Firmware Fix Is the Key Update

If you own a Strix Halo system and tried to run ROCm workloads for local LLM inference, you probably ran into hard crashes, GPU hangs, or instant failures when loading models. Most users discovered quickly that Vulkan-based paths kept working, while ROCm was effectively unusable. That behavior was the clue to what was really broken.

This issue is now largely resolved, but the fix is more specific than just “update everything”.

The real cause: broken AMD GPU firmware

The crashes were caused by a bad AMD GPU firmware package that shipped in late 2025. A faulty MES firmware blob made its way into linux-firmware updates used by most major distributions. If that firmware is present on your system, the GPU can lock up under ROCm compute workloads even if your kernel and ROCm versions are perfectly fine.

This is why users saw identical failures across different distros and kernel versions, and why Vulkan continued to work.

The fix that actually matters

To resolve the problem, you must be running a fixed linux-firmware package. For Strix Halo systems, that means avoiding the late November 2025 firmware and using a known good version instead.

Stable options are firmware releases from January 2026 onward, starting with 20260110. These versions restore ROCm stability under heavy GPU compute loads.

If you update your kernel or ROCm but keep the broken firmware from late November 2025, the crashes will continue. Firmware is the deciding factor here.

Current state for local LLM users

As of now, updated firmware packages are flowing downstream into mainstream distributions. Once the correct firmware is installed, ROCm workloads on Strix Halo behave normally again, including sustained inference runs that previously caused page faults or system freezes.

Bottom line

If you are testing or deploying ROCm on Strix Halo for local inference, focus first on your linux-firmware version. With the fixed firmware in place, the platform is finally stable enough to evaluate performance, memory behavior, and real performance-per-dollar for integrated GPU LLM workloads.

Trending Stories