RTX 5080 via Thunderbolt 5 eGPU: Hard lock on CUDA operations (nvidia-smi works at idle)

[roger-pmta]

NVIDIA Open GPU Kernel Modules Version

590.44.01 and 580.105.08, both from the NVIDIA's official CUDA rhel10 repo

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

Rocky Linux 10.1 (Red Quartz)

Kernel Release

6.12.0-124.13.1.el10_1.x86_64 (PREEMPT_DYNAMIC)

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 5080

Describe the bug

RTX 5080 connected via Thunderbolt 5 eGPU enclosure (Sonnet Breakaway Box 850T5) initializes correctly and is visible to nvidia-smi at idle. Any CUDA operation causes immediate system hard-lock requiring power cycle. No kernel panic, no Xid error logged, no SysRq response.

This appears related to closed issue #900 (RTX 5090 via OCuLink), which showed identical symptoms: GPU functional at idle, crash under load, GSP firmware bootstrap errors. That issue was resolved by switching OCuLink docks, but dock alternatives for Thunderbolt 5 are extremely limited.

PCIe link negotiates correctly at 16GT/s x4 (PCIe 4.0 x4) - optimal for Thunderbolt 4 host controller bandwidth. BAR allocation succeeds with hotplug resource reservation parameters. Driver loads and initializes without error.

Minimal reproducer:

nvidia-smi                    # Works - GPU visible, ~2W idle
python3 -c "import torch; torch.zeros(1, device='cuda')"   # Hard lock

Hardware:

• GPU: NVIDIA GeForce RTX 5080 (GB203)
• eGPU Enclosure: Sonnet Breakaway Box 850T5 (Thunderbolt 5)
• Host: Lenovo ThinkPad X1 Carbon Gen 11
• Thunderbolt Controller: Intel Raptor Lake-P Thunderbolt 4 (host), USB4/TB5 (enclosure)
• OS: Rocky Linux 10.1

Required kernel parameters: pcie_ports=native pcie_aspm=off pcie_port_pm=off pci=assign-busses,realloc

Without pcie_ports=native, GPU enters D3cold and driver fails with "Unable to change power state from D3cold to D0".

To Reproduce

1. Connect RTX 5080 to Sonnet Breakaway Box 850T5 (Thunderbolt 5 eGPU enclosure)
2. Connect enclosure to host via Thunderbolt cable
3. Boot system with kernel parameters: pcie_ports=native pcie_aspm=off pcie_port_pm=off pci=assign-busses,realloc
4. Confirm GPU detected: nvidia-smi (shows GPU at idle, ~2W, ~30°C)
5. Run any CUDA operation: python3 -c "import torch; torch.zeros(1, device='cuda')"
6. System hard-locks immediately. No kernel panic, no SysRq response. Power cycle required.

Bug Incidence

Always

nvidia-bug-report.log.gz

nvidia-bug-report.log.gz

More Info

Also reported at https://forums.developer.nvidia.com/t/590-release-feedback-discussion/353310/53, cross-posted to https://forums.developer.nvidia.com/t/580-release-feedback-discussion/341205/898, and and linux-bugs

[roger-pmta]

For folks who find this, @morgwai and I have been troubleshooting in the PR (#981) and the thread at https://egpu.io/forums/thunderbolt-linux-setup/rtx-5080-via-thunderbolt-5-egpu-hard-lock-on-cuda-operations-nvidia-smi-works-at-idle/ too.

[roger-pmta]

NVIDIA RTX 5080 eGPU on Linux - Working Configuration

Hardware:

• Laptop: Lenovo ThinkPad X1 Carbon Gen 11 (Thunderbolt 4)
• eGPU Enclosure: Sonnet Breakaway Box 850T5 (Thunderbolt 5, JHL9480)
• GPU: NVIDIA GeForce RTX 5080 (Blackwell, GB203)
• Device ID: 10de:2c02
• OS: Rocky Linux 10.1, Kernel 6.12.0-124.16.1.el10_1.x86_64

BIOS Settings:

• Thunderbolt 4: Enabled
• PCIe Tunneling: Enabled
• Kernel DMA Protection: Enabled (can stay enabled)

Kernel Parameters (add via grubby):

sudo grubby --update-kernel=ALL --args="pci=assign-busses,hpbussize=0x10,hpmmiosize=64M,hpmmioprefsize=384M,realloc pcie_port_pm=off pcie_aspm.policy=performance intel_iommu=off thunderbolt.clx=0"

Parameters explained:

pci=assign-busses      - Kernel reassigns PCI bus numbers (helps Thunderbolt)
hpbussize=0x10         - Reserve 16 bus numbers per hotplug bridge
hpmmiosize=64M         - Reserve 64MB non-prefetchable MMIO per hotplug bridge (for BAR0)
hpmmioprefsize=384M    - Reserve 384MB prefetchable MMIO per hotplug bridge (for BAR1+BAR3)
realloc                - Force PCI resource reallocation for hotplug
pcie_port_pm=off       - Disable PCIe port power management (prevents D3cold freeze on cold boot)
pcie_aspm.policy=performance - Minimize ASPM power state transitions
intel_iommu=off        - Disable IOMMU (reduces DMA latency for eGPU)
thunderbolt.clx=0      - Disable Thunderbolt low-power states on high-speed lanes

Also blacklist nouveau in initramfs:

rd.driver.blacklist=nouveau rd.driver.blacklist=nova-core

Driver: Patched 590.44.01 Open (NVreg_ForceExternalGpu module parameter)

The stock NVIDIA driver does not recognize Thunderbolt 4/5 bridges as valid
eGPU enclosures. Without proper eGPU detection, power management code runs
that can cause PCIe link instability.

Required patch hacks around this by adding a NVreg_ForceExternalGpu module
parameter (long-term fix will likely require fixes in the GSP firmware)

• kernel-open/nvidia/nv-reg.h: Define ForceExternalGpu parameter
• src/nvidia/arch/nvalloc/unix/src/osinit.c: Check parameter and force eGPU mode
• See attached patch

When NVreg_ForceExternalGpu=1, the driver bypasses automatic Thunderbolt
bridge detection and treats the GPU as external, which:

• Sets PDB_PROP_GPU_IS_EXTERNAL_GPU property
• Skips Platform Request Handler power management
• Works with any Thunderbolt enclosure (TB3, TB4, TB5, future)

Setup

Step 1: Add CUDA repository

sudo dnf config-manager --add-repo <https://developer.download.nvidia.com/compute/cuda/repos/rhel10/x86_64/cuda-rhel10.repo>
sudo dnf config-manager --set-enabled cuda-rhel10-x86_64

Step 2: Install packages

sudo dnf install kmod-nvidia-open-dkms nvidia-driver-cuda nvidia-gpu-firmware

Installed packages:
  kmod-nvidia-open-dkms-590.44.01-1.el10.noarch
  nvidia-driver-cuda-590.44.01-1.el10.x86_64
  nvidia-driver-cuda-libs-590.44.01-1.el10.x86_64
  nvidia-gpu-firmware-20251008-15.8.el10_0.noarch
  nvidia-kmod-common-590.44.01-1.el10.noarch
  nvidia-modprobe-590.44.01-1.el10.x86_64
  nvidia-persistenced-590.44.01-1.el10.x86_64
  libnvidia-cfg-590.44.01-1.el10.x86_64
  libnvidia-gpucomp-590.44.01-1.el10.x86_64
  libnvidia-ml-590.44.01-1.el10.x86_64

Step 3: Build and install patched driver

Note: the official NVIDIA repo installs modules to
/lib/modules/$(uname -r)/extra which supersedes the default install location
of /lib/modules/$(uname -r)/kernel/drivers/video.

cd /path/to/open-gpu-kernel-modules
git apply 0001-nvidia-add-NVreg_ForceExternalGpu-module-parameter.patch
make -j$(nproc) DEBUG=1 modules
sudo make modules_install

Step 4: Configure nvidia module options

Create /etc/modprobe.d/nvidia.conf:

# Blacklist display modules for compute-only
blacklist nvidia_drm
blacklist nvidia_modeset

# Override system softdep - only load nvidia-uvm, skip drm/modeset (compute-only)
softdep nvidia post: nvidia-uvm

# Disable Resizable BAR (resize fails with small bridge windows)
options nvidia NVreg_EnableResizableBar=0

# Disable dynamic power management (prevents D3cold issues over Thunderbolt)
options nvidia NVreg_DynamicPowerManagement=0

# Force eGPU mode (bypasses Thunderbolt bridge detection for TB4/TB5 enclosures)
options nvidia NVreg_ForceExternalGpu=1

# Lock clocks after driver loads (prevents ASPM-triggered link degradation)
install nvidia /usr/local/bin/nvidia-egpu-setup.sh $CMDLINE_OPTS

Step 5: Enable nvidia-persistenced

sudo systemctl enable nvidia-persistenced

Step 5a: Set up automatic clock locking (prevents PCIe link degradation)

The PCIe link degrades from Gen4 to Gen1 due to ASPM common clock
reconfiguration during power state transitions. Locking GPU clocks prevents this
by keeping the GPU active.

Create /usr/local/bin/nvidia-egpu-setup.sh:

#!/bin/bash
# Clock locking wrapper for nvidia module load

logger -t nvidia-egpu "Install directive triggered"

# Load the actual nvidia module first
/sbin/modprobe --ignore-install nvidia "$@"
ret=$?

if [ $ret -ne 0 ]; then
    logger -t nvidia-egpu "modprobe failed with $ret"
    exit $ret
fi

logger -t nvidia-egpu "Module loaded, waiting for nvidia-smi..."

# Wait for nvidia-smi to respond
for i in {1..60}; do
    if nvidia-smi --query-gpu=name --format=csv,noheader &>/dev/null; then
        logger -t nvidia-egpu "nvidia-smi responding after ${i}s"
        break
    fi
    sleep 1
done

if ! nvidia-smi --query-gpu=name --format=csv,noheader &>/dev/null; then
    logger -t nvidia-egpu "Timeout - nvidia-smi not responding"
    exit 0  # Don't fail module load
fi

sleep 2

logger -t nvidia-egpu "Locking clocks..."
nvidia-smi -pm 1
nvidia-smi -lgc 3090,3090
nvidia-smi -lmc 15001,15001
nvidia-smi -pl 390
logger -t nvidia-egpu "Complete"

Then mark it as executable

sudo chmod +x /usr/local/bin/nvidia-egpu-setup.sh

The install directive intercepts modprobe nvidia, runs the script which loads
the module synchronously, then locks clocks. This avoids backgrounding issues
where scripts get killed before completion.

Logs go to syslog (journalctl -t nvidia-egpu).

Step 6: Enroll Thunderbolt device for auto-authorization

sudo boltctl enroll <UUID>  # Get UUID from: sudo boltctl list

The enrolled device, in my case:

UUID: e4ec8780-0079-3b72-ffff-ffffffffffff
Policy: auto

Step 7: Reboot with eGPU connected

Verification:

nvidia-smi
sudo lspci -vvs 14:00.0 | grep LnkSta  # Should show 16GT/s
lsmod | grep nvidia  # Should NOT show nvidia_modeset or nvidia_drm
dmesg | grep -i "external.*gpu"  # Should show eGPU detected

Current Status

• nvidia-smi: Working
• nvidia-persistenced: Enabled
• BAR1: 256MB (ReBAR disabled due to bridge window constraints)
• Thunderbolt: Auto-authorized via boltctl enrollment
• PCIe Link: Gen4 (16GT/s) x4 - maintained via clock locking
• Bandwidth: ~4 GiB/s (Host↔Device)
• Compute: ~40 TFLOP/s, confirmed with gpu-burn

Technical Details

Why Automatic eGPU Detection Fails (without NVreg_ForceExternalGpu):

The NVIDIA driver's RmCheckForExternalGpu() function checks if a GPU is
connected via a Thunderbolt eGPU enclosure by:

1. Looking for Intel vendor ID (0x8086) on upstream bridge
2. Querying closed-source approved TB3 bridge list
3. Checking for HotPlug+ AND Surprise+ slot capabilities

Thunderbolt 4/5 bridges fail because:

• They're not in the closed-source TB3 approved bridge list
• TB4/TB5 bridges may report HotPlug- Surprise+ (OR)

Without eGPU detection, PDB_PROP_GPU_IS_EXTERNAL_GPU is not set, causing the
Platform Request Handler to run power management code intended for internal
GPUs.

Why These Settings:

1. pcie_port_pm=off: Prevents Thunderbolt bridges from entering D3cold,
   which causes "Unable to change power state from D3cold to D0" errors
   and system freeze on cold boot.

2. hpmmiosize=64M, hpmmioprefsize=384M: Sized for actual BAR needs:

   • BAR0: 64MB non-prefetchable
   • BAR1: 256MB prefetchable
   • BAR3: 32MB prefetchable

   Larger values starve other bridges since 4 bridges share ~49GB. I believe this
   only applies to hotplug devices.

3. NVreg_EnableResizableBar=0: ReBAR resize fails with ENOENT because
   the bridge prefetchable window is too small after allocation.

4. NVreg_DynamicPowerManagement=0: Prevents GPU from entering low power
   states that cause link issues over Thunderbolt.

5. NVreg_ForceExternalGpu=1: Forces eGPU mode, bypassing automatic
   Thunderbolt bridge detection. Required for TB4/TB5 enclosures. See patch.

6. softdep nvidia post: nvidia-uvm: Overrides system default that loads
   nvidia-drm (which pulls in nvidia_modeset). Only nvidia and
   nvidia-uvm are needed for compute.

PCIe Link Speed Issue (Gen4 → Gen1 degradation):

Without clock locking, the kernel's ASPM code detects "inconsistent common
clock configuration" during device enumeration and triggers a link retrain.
The JHL9480 Thunderbolt 5 controller retrains at Gen1 instead of Gen4.

The message in dmesg is:

pcieport 0000:00:07.2: ASPM: current common clock configuration is inconsistent, reconfiguring

This happens in drivers/pci/pcie/aspm.c:pcie_aspm_configure_common_clock()
which calls pcie_retrain_link(). No kernel parameter exists to disable this.

Workaround: Lock GPU clocks to maximum immediately after driver bind. This keeps
the GPU in an active power state, preventing the ASPM reconfiguration from
triggering link speed degradation.

---

Last tested: 2025-12-10