Panther Lake Xe3 iGPU (8086:B03E) — Level Zero GPU driver segfaults on initialization

[eckmannmiles91]

Environment

• CPU: Intel Core Ultra 7 366H (Panther Lake)
• iGPU PCI ID: 8086:B03E (Xe3 architecture, Panther Lake [Intel Graphics])
• OS: Proxmox VE 8.x (Debian Bookworm base), services running directly on host (not LXC)
• Kernel: 6.17.2-1-pve (pinned — cannot upgrade without breaking NPU driver)
• compute-runtime / intel-level-zero-gpu: 26.14.37833 (libze_intel_gpu.so.1.15.37833)
• intel-level-zero-npu: v1.32.1 (working, separate device)
• NPU: 50 TOPS Panther Lake NPU at /dev/accel/accel0 — fully functional via Level Zero

Problem

The Panther Lake Xe3 iGPU is recognized by the xe kernel driver and /dev/dri/ device nodes exist, but the Level Zero GPU driver (libze_intel_gpu.so) crashes during initialization with segfaults and general protection faults. sycl-ls does not enumerate a level_zero:gpu device — only opencl:cpu appears.

The NPU on the same system works perfectly via intel-level-zero-npu, confirming that Level Zero infrastructure is functional.

Expected Behavior

The Panther Lake Xe3 iGPU should appear as a level_zero:gpu device in sycl-ls, similar to Arrow Lake Xe2 iGPUs which work correctly with the same compute-runtime + oneAPI + llama.cpp SYCL stack.

Diagnostics

lspci

0000:00:02.0 VGA compatible controller: Intel Corporation Panther Lake [Intel Graphics]

/dev/dri

crw-rw---- 1 root video  226,   0 Apr 22 22:00 card0
crw-rw---- 1 root render 226, 128 Apr 22 22:00 renderD128

dmesg (xe driver + crash traces)

[    4.281929] [drm] Initialized xe 1.1.0 for 0000:00:02.0 on minor 0
[    5.782058] xe 0000:00:02.0: [drm] Allocated fbdev into stolen
[    7.281350] xe 0000:00:02.0: [drm] Reducing the compressed framebuffer size.
[    7.333858] xe 0000:00:02.0: [drm] fb0: xedrmfb frame buffer device
[72669.395838] llama-bench[235685]: segfault at 128 ip 73584d1b6a9e sp 7ffd791d5e60 error 4 in libze_intel_gpu.so.1.15.37833[+1b6a9e]
[72693.268360] llama-bench[236053]: segfault at 128 ip 7f38dd5b6a9e sp 7fff6e930ff0 error 4 in libze_intel_gpu.so.1.15.37833[+1b6a9e]
[73308.382664] xe 0000:00:02.0: Using 41-bit DMA addresses
[73308.443635] traps: llama-bench[238834] general protection fault ip:713a780e6104 sp:7ffe22900ad8 error:0 in libze_intel_gpu.so.1.15.37833[+6e6104]
[73335.765711] traps: llama-bench[239225] general protection fault ip:7266f0ae6104 sp:7fff2843e758 error:0 in libze_intel_gpu.so.1.15.37833[+6e6104]
[73360.915387] traps: llama-bench[239571] general protection fault ip:75eecf4e6104 sp:7ffc1fcf2ef8 error:0 in libze_intel_gpu.so.1.15.37833[+6e6104]
[73399.157700] traps: llama-bench[239979] general protection fault ip:7e820d0e6104 sp:7fff5c4e4cb8 error:0 in libze_intel_gpu.so.1.15.37833[+6e6104]
[73482.193347] traps: llama-bench[240532] general protection fault ip:7438e3ae6104 sp:7ffc31c92838 error:0 in libze_intel_gpu.so.1.15.37833[+6e6104]

Installed packages

ii  intel-level-zero-npu  1.32.1.20260422-24767473183~ubuntu24.04
rc  intel-opencl-icd       26.14.37833.4-0 (removed, was segfaulting)

sycl-ls output

[opencl:cpu][opencl:0] Intel(R) OpenCL, Intel(R) Core(TM) Ultra 7 366H OpenCL 3.0 (Build 0)

No level_zero:gpu device appears.

Kernel

Linux version 6.17.2-1-pve (gcc 14.2.0, GNU ld 2.44) #1 SMP PREEMPT_DYNAMIC PMX 6.17.2-1 (2025-10-21T11:55Z)

Context

• We run llama.cpp with the SYCL backend for local LLM inference on iGPUs. This works correctly on an Arrow Lake (Ultra 7 265) Xe2 iGPU using the same compute-runtime + oneAPI + llama.cpp SYCL stack (~10 tok/s on Gemma 4 E4B Q4_K_M).
• We attempted to run the same workload on the Panther Lake Xe3 iGPU but hit these initialization crashes.
• The kernel is pinned to 6.17.2-1-pve because newer PVE kernels (6.17.9+) cause NPU driver segfaults. We cannot upgrade the kernel to test if a newer xe driver resolves this.
• Both crash patterns (segfault at offset +1b6a9e and GPF at offset +6e6104) are reproducible across multiple attempts.

Possible Causes

1. PVE-patched kernel 6.17.2 may be missing Panther Lake-specific xe driver commits from mainline
2. Missing GPU firmware blob
3. compute-runtime 26.14.37833 may not yet have full Panther Lake Xe3 support

Request

Any guidance on what's needed to get the Panther Lake Xe3 iGPU working with Level Zero on this kernel would be appreciated. If a specific compute-runtime version, kernel version, or firmware file is required, we'd like to know so we can plan accordingly (we need to maintain NPU compatibility).

Related: #918 (filed earlier with less detail, no response)