ArcGIS Pro 3.6 Deep Learning and Blackwell GPU (sm_120): Solving the CUDA Mismatch

The release of the NVIDIA Blackwell architecture introduces Compute Capability 12.0 (sm_120), a significant jump in processing power for spatial AI. However, early adopters of ArcGIS Pro 3.6 may encounter "CUDA Error: no kernel image is available for execution" or general tool failures. This occurs because the pre-packaged Deep Learning Libraries for ArcGIS Pro often target older CUDA versions (e.g., sm_70 to sm_90). To leverage a Blackwell GPU, the underlying PyTorch and TorchVision environments must be forced to recognize the new instruction sets. This tutorial provides a workflow to bridge the gap between Esri's library installers and the cutting-edge Blackwell hardware.

Table of Content

• Purpose of CUDA Alignment
• Common Use Cases
• Step by Step: Patching the Environment
• Best Results for AI Performance
• FAQ
• Disclaimer

Purpose

The primary purpose of this guide is to achieve Binary Compatibility between the ArcGIS Pro Deep Learning Framework and Blackwell's sm_120 architecture. Without a workaround, the arcgis.learn module fails because the pre-compiled CUDA kernels provided in the standard installer do not contain the specific machine code required for Blackwell's tensor cores. By updating the Conda environment and utilizing NVIDIA's Forward Compatibility layers, we can enable GPU acceleration for imagery analysis without waiting for a major software patch.

Use Case

This workaround is essential for GIS professionals using Blackwell-based workstations (e.g., RTX 50-series or B-series data center GPUs) for:

• Automated Feature Extraction: Using MaskRCNN or DeepLab for large-scale building footprint and road digitizing.
• Change Detection: Running multi-temporal satellite imagery through U-Net models for disaster response mapping.
• Point Cloud Classification: Processing LiDAR data using PointCNN to differentiate vegetation from urban infrastructure.
• Real-time Object Detection: Deploying YOLO-based models within the ArcGIS Video AI extension.

Step by Step

1. Install the ArcGIS Pro Deep Learning Libraries

Download and install the official "Deep Learning Libraries Installer for ArcGIS Pro 3.6" from the Esri support site. This provides the base deeplearning Conda environment, even if it is initially incompatible with Blackwell.

2. Clone the Default Environment

Never modify the base environment. Open the Python Command Prompt (found in the ArcGIS folder) and clone the environment:
conda create --name arcgis_blackwell --clone arc_pro_deeplearning_env

3. Force-Update PyTorch for sm_120 Compatibility

The core issue is usually the PyTorch version. Blackwell requires PyTorch 2.4+ and CUDA 12.4+. Update your cloned environment using the following command:
conda install -n arcgis_blackwell pytorch torchvision torchaudio pytorch-cuda=12.4 -c pytorch -c nvidia

4. Set Environment Variables for JIT Compilation

If binary kernels are still missing, you can force Just-In-Time (JIT) Compilation. This allows the GPU driver to compile the kernels on the fly for sm_120. In Windows System Environment Variables, add:
Variable: CUDA_MODULE_LOADING
Value: LAZY

5. Point ArcGIS Pro to the New Environment

Open ArcGIS Pro, go to Project > Options > Python. Click the "Manage Environments" button and select arcgis_blackwell. Restart ArcGIS Pro to apply the change.

Best Results

FAQ

Why does ArcGIS Pro say "GPU not found" despite having an RTX 50-series?

ArcGIS Pro 3.6 checks for a specific version of nvcuda.dll. If your driver is outdated, the check fails. If the driver is current, the issue is that the library torch-cuda in the default Esri environment doesn't include the Blackwell compute capability. Updating the environment as shown in Step 3 fixes this.

Will this break other Geoprocessing tools?

By using a Cloned Environment, you protect the core ArcGIS Pro installation. If tools start failing, you can simply switch back to the default arcgispro-py3 environment in the Python options.

Is FP8 training supported in ArcGIS Pro on Blackwell?

While Blackwell supports native FP8, arcgis.learn is currently built on a PyTorch implementation that defaults to FP32 or Mixed Precision (FP16). You will benefit from Blackwell's speed, but full FP8 utilization may require custom training scripts outside of the standard Pro UI.

Disclaimer

Manually updating Conda packages within ArcGIS Pro can lead to dependency conflicts (DLL hell). This workaround is intended for advanced users and developers. Esri does not officially support Blackwell sm_120 architecture until the release of specific library patches or ArcGIS Pro 3.7/4.0. This information is current as of 2026 hardware/software configurations.

Tags: ArcGISPro, BlackwellGPU, DeepLearning, CUDA