NVIDIA Launches CUDA Toolkit 13.4 with Arm Support

NVIDIA has officially launched CUDA Toolkit 13.4, bringing native application development to the Windows on Arm platform for the first time. The update also introduces early functional preview support for the upcoming NVIDIA Rubin GPU architecture, designated as compute capability 107, allowing developers to begin porting their applications ahead of general hardware availability. Notably, the CUDA SDK installers will no longer bundle the NVIDIA driver, requiring developers to install the nvidia-open driver or cuda-toolkit packages separately.

For system administrators and developers managing shared hardware, Multi-Process Service V3 introduces a modernized control layer featuring a scriptable CLI, TOML configuration, and cgroup-integrated memory limits for precise GPU partitioning. On hardware-coherent platforms like Grace Hopper, Grace Blackwell, and Vera Rubin, the driver now defaults to Coherent Driver-based Memory Management instead of NUMA. Additionally, the new CUDA Compute Fabric Transport API allows communication-library developers to move data across NVLink fabrics using named logical endpoints, reducing virtual-address pressure in massive multi-GPU setups.

The release updates CUDA Python to cuda.core 1.1.0, which adds texture and surface programming alongside .pyi type stubs for IDE autocompletion. Meanwhile, cuda.compute 1.1 enables ahead-of-time compilation of algorithms for multiple architectures, such as sm_80 and sm_90, without requiring a physical GPU. The CUDA C++ Core Libraries 3.4 package delivers a warp-specialized implementation of cub::DeviceScan for Blackwell GPUs. This update leverages the Tensor Memory Accelerator to boost memory-bandwidth utilization up to 92 percent, up from approximately 50 percent in previous versions.

Mathematical and profiling workflows also receive significant upgrades. The cuBLASLt library now dynamically schedules Grouped GEMM computations on Blackwell data center GPUs to minimize load imbalances in Mixture of Experts workloads, specifically optimizing calls with 32 or more groups. For profiling, Nsight Python 1.0 automates performance analysis, while Nsight Systems 2026.5.1 adds a PyTorch shape-tracing option to trace tensor shapes and training parameters. Host compiler compatibility has also been expanded to support GCC 16 and Clang 22.