CUDA 13.2 + cuDNN 9.23: PureLab GPU acceleration setup

CUDA 13.2 + cuDNN 9.23: PureLab GPU acceleration setup

PureLab now ships a newer ONNX Runtime GPU build. The runtime requirement is:

ort provides binaries for CUDA ≥ 13.2 and targets cuDNN ≥ 9.23. Make sure compatible versions of CUDA and cuDNN are installed and available on the PATH.

Follow the steps in order. Do not skip the first one.

0. Update PureLab first

Older PureLab builds are not supported. Install the current latest PureLab, then configure CUDA. Upgrading the runtime libraries without upgrading the app will still fail GPU acceleration.

CUDA 12.9, the old 12.9 bundle, and cudart64_12.dll no longer work.

Fast track: all-in-one package (recommended)

We provide an extract-and-run bundle: no installer, no extra components.

1. Download: Google Drive: CUDA 13.2 bundle | Baidu Netdisk: CUDA 13.2 bundle (code: w4m7)
2. Extract: Unzip it to a permanent folder (for example D:\CUDA_13.2_Express).
3. One-click config: Open PureLab → Settings → CUDA Settings, and point the configuration directory at that folder.

No system environment variables are required for the bundle. If you prefer the official toolchain, continue below.

Version requirements

Core idea: both DLLs must be found together

Installing cuDNN is not the same as the app being able to load it. Windows PATH also does not search subfolders recursively.

The current handshake needs both of these files:

• cudart64_13.dll (CUDA 13 runtime)
• cudnn64_9.dll (cuDNN 9 core)

On Windows, CUDA 13 DLLs may live in ...\CUDA\v13.2\bin or ...\CUDA\v13.2\bin\x64. The cuDNN 9.23 graphical installer usually lands in C:\Program Files\NVIDIA\CUDNN\v9.23\, while the real DLLs often sit in a CUDA-versioned folder such as bin\13.2 or bin\13.2\x64.

The reliable approach: find both files in Explorer, copy the cuDNN DLLs into the folder that already contains cudart64_13.dll, then point PureLab at that folder.

Do not copy CUDA 12 cuDNN files (for example from bin\12.9) into CUDA 13.

Step 1: Install CUDA Toolkit 13.2 or newer

Go to the NVIDIA CUDA Toolkit download page and choose CUDA Toolkit 13.2 (or a newer 13.x release, as long as it is ≥ 13.2).

We recommend the Express install. Remember the path. The default is usually:

C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.2

After setup, search that tree for cudart64_13.dll and note the folder that contains it. That folder is the CUDA Bin directory you will give PureLab.

Step 2: Install cuDNN 9.23 or newer (CUDA 13.x)

Go to the NVIDIA cuDNN download page.

• Version: 9.23 or newer
• Platform: Windows
• CUDA version must be 13.x, not 12.x

The graphical installer is the straightforward option. After setup, search for cudnn64_9.dll and confirm it lives under a CUDA 13 folder (the name should include 13.2 or another 13.x), not 12.9.

Copy the DLLs from that cuDNN bin folder (including the matching 13.x subdirectory) into the folder that already contains cudart64_13.dll.

Putting cudnn64_9.dll next to the CUDA 13 runtime is still the simplest way to fix a missing core library.

Step 3: Point PureLab at that folder

Open PureLab Settings → CUDA Settings and set CUDA Bin Directory to the folder that contains both cudart64_13.dll and cudnn64_9.dll, for example:

C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.2\bin

If your Toolkit put the runtime in bin\x64, use that x64 path. Do not point at an empty parent bin folder.

Step 4: Add the same path to the system PATH

ONNX Runtime expects CUDA and cuDNN to be available on PATH.

1. Press Win + S and search for “Edit the system environment variables”.
2. Click Environment Variables, find Path under System variables, and click Edit.
3. Add the folder confirmed above (the one with cudart64_13.dll; if cuDNN was not merged, also add the 13.x folder that actually contains cudnn64_9.dll).
4. If CUDA 12.9 is still on PATH: PureLab no longer uses 12.x, so remove it if nothing else needs it. If other software still needs 12.x, put the 13.x paths before 12.x.
5. Do not add a libnvvp path. Visual Profiler is not part of the CUDA 13 runtime you need.

Step 5: Restart and troubleshooting

Restart is required

After changing PATH or the PureLab directory, fully quit PureLab (reboot the PC if needed) so the new paths load.

Common issues

• Still on CUDA 12.9: diagnostics will ask for CUDA ≥ 13.2. Update PureLab and retarget the 13.2 folder.
• CUDA status FAIL / LoadLibraryExW failed: confirm cudnn64_9.dll is in the exact folder PureLab points to, and that it came from CUDA 13.x cuDNN, not 12.x.
• CUDA found, cuDNN missing: PATH is not recursive. Adding only C:\Program Files\NVIDIA\CUDNN\v9.23\bin is often not enough. Use the folder that actually contains the DLL, such as bin\13.2 or bin\13.2\x64.
• DirectML status FAIL: if CUDA shows PASS, the app prefers CUDA. You can ignore DirectML.
• TensorRT not installed: optional. Install it from NVIDIA only if you want that extra path, and add its library directory to PATH.

The CUDA 12.9 article is retired. Do not follow it: Deprecated CUDA 12.9 note