Installing CUDA Toolkit Fixed RTX 5090 Crashing on 8.5.5
Posted: Tue Jan 20, 2026 3:32 am
Try installing CUDA toolkit if you're having issues with Nvidia GPUs. This worked for me on a 5090 running v8.5.6 beta, and 2x 4080 and 5090 running v8.5.5.
I am noob so I had Gemeni put together an explanation of the exhausting troubleshooting process we went through:
The Issue: I recently attempted to set up a folding node on a cloud instance (vast.ai) using an NVIDIA RTX 5090. While the GPU was detected correctly by the v8.5.5 client, FahCore_27 (v8.2.1) would consistently crash immediately upon trying to initialize the CUDA platform.
Symptoms:
Client Version: v8.5.5
GPU: RTX 5090
Driver: 580.95
Error: Caught signal SIGSEGV(11) occurring specifically at the step Configuring platform CUDA.
Observation: The logs showed the Core detected the GPU correctly but crashed before processing any work. Attempts to force OpenCL via config.xml arguments were often ignored or overridden by the client defaulting back to CUDA.
Root Cause Analysis: It appears the bundled CUDA libraries included inside the downloaded FahCore_27 package are too old to support the Blackwell architecture (RTX 50-series) or the newer 580.xx drivers. When the Core attempts to link against these internal, outdated libraries to initialize the CUDA context on the 5090, it causes a Segmentation Fault.
The Solution: The fix was to manually install the full system-wide NVIDIA CUDA Toolkit. This likely provides modern shared libraries that override or supplement the incompatible bundled versions, allowing the Core to successfully initialize the CUDA context.
Fix Command:
Results: Immediately after installing the toolkit and restarting the client, the SIGSEGV disappeared. The log now shows Using CUDA on CUDA Platform and the node is folding successfully at 100% load.
Hopefully, this helps others trying to fold on 50-series hardware!
I am noob so I had Gemeni put together an explanation of the exhausting troubleshooting process we went through:
The Issue: I recently attempted to set up a folding node on a cloud instance (vast.ai) using an NVIDIA RTX 5090. While the GPU was detected correctly by the v8.5.5 client, FahCore_27 (v8.2.1) would consistently crash immediately upon trying to initialize the CUDA platform.
Symptoms:
Client Version: v8.5.5
GPU: RTX 5090
Driver: 580.95
Error: Caught signal SIGSEGV(11) occurring specifically at the step Configuring platform CUDA.
Observation: The logs showed the Core detected the GPU correctly but crashed before processing any work. Attempts to force OpenCL via config.xml arguments were often ignored or overridden by the client defaulting back to CUDA.
Root Cause Analysis: It appears the bundled CUDA libraries included inside the downloaded FahCore_27 package are too old to support the Blackwell architecture (RTX 50-series) or the newer 580.xx drivers. When the Core attempts to link against these internal, outdated libraries to initialize the CUDA context on the 5090, it causes a Segmentation Fault.
The Solution: The fix was to manually install the full system-wide NVIDIA CUDA Toolkit. This likely provides modern shared libraries that override or supplement the incompatible bundled versions, allowing the Core to successfully initialize the CUDA context.
Fix Command:
Code: Select all
apt-get update && apt-get install -y nvidia-cuda-toolkitHopefully, this helps others trying to fold on 50-series hardware!