CUDA DL Release 26.07

The NVIDIA container image for CUDA Deep Learning is available on NGC.

.

Contents of the CUDA DL container

The container includes the following (not a complete list):

• Ubuntu 24.04 including Python 3.12
• NVIDIA CUDA 13.3.1
• NVIDIA cuBLAS 13.6.0.2
• NVIDIA cuDNN 9.24.0.43
• NVIDIA NCCL 2.30.7
• rdma-core 63.0
• NVIDIA HPC-X 2.50
• OpenUCX 1.20.0
• NVIDIA GDRCopy 2.5.1
• NVIDIA DOCA 3.4.0
• AWS OFI NCCL 1.17.3
• TensorBoard 2.16.2
• Nsight Compute 2026.2.1.5
• Nsight Systems 2026.3.1.117
• NVIDIA TensorRT™ 11.1.0.106

Driver Requirements

Release 26.07 is based on CUDA 13.3.1. For comprehensive and up-to-date driver compatibility information, please refer to the following documentation:

• NVIDIA CUDA Compatibility Guide - Compati.bility information between CUDA versions and driver releases.
• CUDA Toolkit Release Notes - Driver version requirements and compatibility matrices.
• NVIDIA Drivers Download - Latest NVIDIA drivers.

Key Features and Enhancements

This CUDA DL release includes the following key features and enhancements.

• CUDA DL container image is based on Ubuntu 24.04.
• NIXL, NVIDIA's high-performance network data transfer library, is now included in inference-level containers.
• The GCP NCCL plugin package is now installed in the CUDA base image. On Google Cloud Platform hosts (A3 Ultra, A4, A4X and A4X Max machine types), the container automatically detects and activates the GCP NCCL network, tuner, profiler, and environment plugins for optimized multi-node communication.
  • On GCP hosts with the GCP NCCL plugins installed, NCCL telemetry is enabled by default starting in this release.
  • A runtime advisory message is displayed when running on GCP hosts with the GCP NCCL plugins active, directing users to the GCP NCCL telemetry documentation.

Announcements

• Starting in 26.07, the CUDA DL images contain the GCP NCCL (gIB) plugin.
• Starting in 26.03, the CUDA DL Inference Devel container includes NIXL, for optimized cross-node data transfers, along with the nixlbench benchmarking tool.

• Starting in with the 26.01 release, the CUDA DL Base container is available in four variants optimized for both training and inference use cases

  • Introducing inference-runtime and inference-devel container variants, providing size reduction compared to full development images through stub library optimization for CUSPARSE, CUSOLVER, CUFFT, and NVJITLINK.

• Starting with 25.08 release, CUDA DL Image will be on CUDA 13.0
• Starting with the 25.03 release, NVIDIA provides two CUDA DL images: Developer and Runtime

Known Issues

• Certain distributed jobs may encounter UCX errors involving MLX5, in which case we recommend the use of the environment variable as a workaround: UCX_DC_MLX5_AR_ENABLE=n.