| On this page |
Overview ¶
Houdini has several nodes that allow ML models to be run in Houdini. Running a model is also known as “inferencing”. Performing inference on the CPU is typically very slow. Inferencing is usually a lot faster on the GPU. Houdini can use any GPU that supports CUDA to accelerate inferencing. (For hardware that doesn’t support CUDA, there are other options to accelerate inferencing: look at the docs for the Execution Provider parameter on the inference node you're using.)
Use the following steps to install CUDA and cuDNN on Windows or Linux. See Troubleshooting if you encounter issues while installing the tools.
-
Determine whether your hardware supports CUDA. This typically means that you are on windows or linux and have a modern NVIDIA graphics card. There is no point trying to continue these installation steps if you hardware doesn’t support CUDA, for example, because you're on macOS or because you have an AMD graphics card.
-
Ensure that the latest NVIDIA graphics driver is installed. After you install a new graphics driver, restart your machine before continuing.
-
Install the CUDA toolkit. It is very important that you install version 12.8 of the NVIDIA CUDA Toolkit, otherwise CUDA inference won’t work. Please refer to the following installation guides for windows and linux:
-
Install the cuDNN library. It is very important that your current a correct version of cuDNN, otherwise CUDA inference won’t work. You should be able to use any cuDNN version 9.x.y with x >= 7. For example, cuDNN versions 9.7.0, 9.10.2 should both work. Please refer to the following installation guides for windows and linux:
-
Restart Houdini. You should now be able to use CUDA as an execution provider to accelerate inferencing.
If these installation steps were succesfully completed, all ML inference nodes (any Houdini node with “Inference” in its name) should be able to use CUDA. Make sure your Execution Provider on your inference node set the right way. If it’s set to “Automatic” (the default), it will use CUDA if CUDA is available. You can also set the Execution Provider parameter explicitly to “CUDA”.
Troubleshooting ¶
Windows
-
You may receive a warning that says
Suspiciously long error messages truncated. This warning is usually caused by CUDA running out of memory. Restarting Houdini may resolve this issue. -
You may receive a warning that says
Failed to load any of the following: "cublas64_12.dll", "cublasLt64_12.dll", "cufft64_11.dll", "curand64_10.dll". This warning is usually caused by the CUDA installation. Make sure you install the correct version of CUDA to prevent this issue. Runnvcc --versionin the command line or power shell to verify your installation. -
You may receive a warning that says
Failed to load cudnn64_9.dll. This warning is usually caused by issues with the cuDNN installation. Make sure you install cuDNN in the correct location and that the PATH variable points to a folder containing the.dllfile. -
You may receive a warning that says
Failed to load cudnn_cnn_infer64_9.dll. This warning can be caused by an issue with the cuDNN installation. Make sure you install cuDNN in the correct location.
Linux
-
You may receive a warning that says
Suspiciously long error messages truncated. This warning is usually caused by CUDA running out of memory. Restarting Houdini may resolve this issue. -
You may receive a warning that says
Failed to load any of the following: "libcublas.so.12", "libcublasLt.so.12", "libcufft.so.11", "libcurand.so.10". This warning is usually caused by the CUDA installation. Make sure you install the correct version of CUDA to prevent this issue. Runnvcc --versionin the command line to verify your installation. -
You may receive a warning that says
Failed to load libcudnn.so.9. This warning is usually caused by issues with the cuDNN installation. Make sure you install cuDNN in the correct location. -
You may receive a warning that says
Failed to load libcudnn_cnn_infer.so.89. This warning can be caused by either an issue with the cuDNN or zlib installation. Make sure you install cuDNN and zlib in the correct locations.