Breaking

NVIDIA CUDA 13.4 Adds Windows on Arm Support, Rubin Preview and MPS V3 Controls

October 7, 2026 at 14:08 EDT

Infra & Chips
Software Dev & Coding

NVIDIA has released CUDA Toolkit 13.4 with Windows on Arm support for RTX Spark devices, a developer preview path for the NVIDIA Rubin GPU architecture, and a redesigned Multi-Process Service control layer for shared GPUs. The update, announced on September 9, 2026, positions CUDA less as a single compiler drop than as a platform release for Arm PCs, next-generation data-center GPUs, and containerized GPU sharing NVIDIA Technical Blog.

Sep. 9, 2026 · NVIDIA

CUDA 13.4 broadens the AI compute platform

The release adds Windows on Arm support for RTX Spark, preview support for Rubin, and new shared-GPU controls for containerized infrastructure.

13.4

CUDA Toolkit release

107

Rubin compute capability preview

92%

Peak DeviceScan bandwidth use

616.41+

RTX Spark Windows driver floor

Blackwell scan throughput moves closer to the hardware ceiling

~50%

Previous implementation

up to 92%

CUDA 13.4 DeviceScan

The strongest reported performance figure is a memory-bandwidth utilization gain for large Blackwell scan workloads.

What changes for builders

Windows on Arm

Build native or cross-compiled CUDA apps for RTX Spark.

→

Rubin preview

Start porting to SM_107 before production support.

→

MPS V3

Partition shared GPU resources with cgroup-aware limits.

Why developers care

The release opens a Windows on Arm CUDA path, adds early Rubin build targets, and reduces CUB boilerplate with single-call APIs.

What to watch

Rubin support is preview-only, new features need R615-class drivers, and MPS memory partitioning excludes MIG devices.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.