Hi,
I am experiencing repeated GPU TDR / black-screen failures with a newly installed NVIDIA RTX PRO 4500 Blackwell 32GB under Windows 11. In the most recent failure, the RTX PRO 4500 became completely inaccessible and nvidia-smi reported “GPU is lost”, while a second NVIDIA GPU remained operational.
System configuration
- CPU: AMD Ryzen 7 5700X3D
- Motherboard: ASUS TUF GAMING X570-PLUS
- RAM: 128GB DDR4-3200
- GPU 0: GeForce RTX 5070 Ti 16GB
- GPU 1: RTX PRO 4500 Blackwell 32GB
- PSU: Super Flower Leadex VII Pro 1000W
- OS: Windows 11 Pro 25H2
- Drivers tested: 616.92 Studio Driver and 610.62
The RTX PRO 4500 was installed in September 2026. Before installing it, the same system had been running an RTX 5070 Ti + RTX 5060 Ti configuration for approximately three months, including local LLM and AI image-generation workloads, without these TDR/GPU-lost failures.
Symptoms and previous testing
The problem has occurred with both Unsloth/llama.cpp and LM Studio, using different llama.cpp binaries/CUDA runtime generations.
Windows Reliability Monitor has recorded multiple GPU-related failures, including:
LiveKernelEvent 117
LiveKernelEvent 141
LiveKernelEvent 1B8
with failure buckets pointing to:
nvlddmkm.sys_Blackwell
Examples include:
LKD_0x141_Tdr:6_IMAGE_nvlddmkm.sys_Blackwell
and:
LKD_0x117_Tdr:A_IMAGE_nvlddmkm.sys_Blackwell
The failures do not appear to correlate simply with VRAM usage, GPU temperature, or power consumption. For example, failures have occurred with approximately 17–22GB of VRAM in use, while other workloads using approximately 28–31GB have completed normally.
I also ran a ~600-second CUBLAS/VRAM stress test on the RTX PRO 4500 using approximately 29.8GB of its 31.9GB VRAM, with GPU load near maximum and approximately 180W power consumption. The test completed successfully.
I also tested a reduced 175W power limit, but the black-screen/TDR problem still occurred.
Most important new reproduction (September 26)
To determine whether the issue was caused by the RTX PRO 4500 simultaneously handling display output and CUDA workloads, I moved both physical monitors to the RTX 5070 Ti.
At the time of the latest failure: - RTX 5070 Ti was driving both monitors and the Windows desktop.
- RTX PRO 4500 had no physical displays connected.
- RTX PRO 4500 was being used only as a compute GPU.
- Driver: 610.62
- RTX PRO 4500 power limit: default ~200W
- Workload: local Qwen LLM / llama.cpp together with Blender MCP operations
- RTX PRO 4500 VRAM usage was only approximately 22GB / 32GB.
During the workload, the displays briefly went black several times but recovered through the RTX 5070 Ti. Windows remained usable.
However, the RTX PRO 4500 itself did not recover.
Reliability Monitor recorded:
LiveKernelEvent 117
with:
LKD_0x117_Tdr:A_IMAGE_nvlddmkm.sys_Blackwell
During the same failure sequence, Windows also recorded:
nvidia-smi.exe APPCRASH
Faulting module: nvml.dll
Exception code: c0000005
and the running llama-server.exe process terminated with:
BEX64
Faulting module: ucrtbase.dll
Exception code: c0000409
After the failure, nvidia-smi dmon no longer showed the RTX PRO 4500 normally.
I then manually ran nvidia-smi, which returned:
Unable to determine the device handle for GPU1: 0000:0A:00.0: GPU is lost. Reboot the system to recover this GPU
The RTX 5070 Ti remained operational and continued driving the Windows desktop. A reboot was required to recover the RTX PRO 4500.
This reproduction appears to rule out the requirement for simultaneous physical display output on the RTX PRO 4500, since the card failed while operating as a compute-only GPU.
What I am trying to determine
At this point I cannot determine whether this is: - a Blackwell driver / GSP / GPU recovery issue,
- a hardware problem with this particular RTX PRO 4500,
- a PCIe/platform interaction,
- or a specific CUDA workload triggering a lower-level GPU/driver fault.
Since the card can pass high-load CUBLAS/VRAM stress testing but can still enter a persistent “GPU is lost” state during real LLM/CUDA workloads, I would appreciate any guidance on:
- Whether this resembles a known Blackwell driver/GSP issue;
- Whether there is a recommended driver branch for the RTX PRO 4500 that I should test;
- Which NVIDIA diagnostic logs or additional information would be most useful to collect the next time the problem occurs;
- Whether the current evidence suggests that I should proceed with a hardware RMA.
I have preserved the relevant Windows Reliability Monitor event details, WATCHDOG files, and the exact nvidia-smi output from the failure and can provide the diagnostic information through another method if requested.
Thanks.