RTX 5060 Ti eGPU (OCuLink) hard locks system on CUDA allocation (Ollama/llama.cpp) - Blackwell GSP firmware bug

[budionqc]

NVIDIA Open GPU Kernel Modules Version

580

Please confirm this issue does not happen with the proprietary driver (of the same version). This issue tracker is only for bugs specific to the open kernel driver.

• I confirm that this does not happen with the proprietary driver package.

Operating System and Version

Ubuntu Server 26.04 LTS (Resolute Raccoon)

Kernel Release

7.0.0-28-generic

Please confirm you are running a stable release kernel (e.g. not a -rc). We do not accept bug reports for unreleased kernels.

• I am running on a stable kernel release.

Hardware: GPU

NVIDIA GeForce RTX 5060 Ti (16GB VRAM, eGPU connected via OCuLink)

Describe the bug

Description of the bug

On an AMD Ryzen 7 Mini-PC running Ubuntu with an eGPU (NVIDIA GeForce RTX 5060 Ti connected via OCuLink), any sustained CUDA memory allocation (such as running Ollama / llama.cpp inference) instantly hard-locks the host system.
The crash is a silent, complete system freeze requiring a hard power cycle. No Xid error or kernel panic is logged before the freeze.

• Pure compute workloads like gpu_burn run fine at 100% load for several minutes without issues.
• The hard-lock happens immediately when the model begins to load into VRAM/bootstrap CUDA.
• Disabling the GSP Firmware (NVreg_EnableGpuFirmware=0) did not resolve the issue (system still freezes).

This issue is highly consistent with reports on Blackwell GPUs in external enclosures (OCuLink/Thunderbolt) experiencing GSP bootstrap or PCIe state sync crashes during dynamic memory allocation.

Hardware & Software Environment

• CPU: AMD Ryzen 7
• GPU: NVIDIA GeForce RTX 5060 Ti (16GB VRAM)
• Interface: OCuLink (PCIe external)
• OS: Ubuntu Server 26.04 LTS (Resolute Raccoon)
• Driver Version: 595.xx / 580.xx
• Container Environment: Podman (Rootless) running Ollama

Attached Logs

I am attaching the nvidia-bug-report.log.gz generated right after a clean boot.

To Reproduce

1. Connect RTX 5060 Ti eGPU via OCuLink.
2. Start Ollama container (Podman).
3. Run a model: podman exec -it ollama ollama run llama3.2
4. System hard-locks instantly when the model begins loading.

Bug Incidence

Always

nvidia-bug-report.log.gz

nvidia-bug-report.log.gz

More Info

No response