NVIDIA Releases CUDA 13.4 with Windows on Arm and Rubin

The toolkit splits driver installs, introduces container-level GPU slicing, and enables early compilation for upcoming Rubin silicon.

NVIDIA published technical release notes and documentation for CUDA Toolkit 13.4 on Wednesday, October 7, 2026, outlining the changes on its engineering blog and through an official broadcast on X. The core software platform behind modern artificial intelligence workloads focuses on two distinct areas: preparing compilers for upcoming data center microarchitectures and extending native support to personal computers beyond Linux.

From data centers to laptops: Windows on Arm and Rubin

The most immediate change for developer setups is native support for Windows on Arm. Prior to this release, developers working on Arm hardware had to run Linux distributions. The update directly targets the latest wave of enterprise Arm laptops, allowing engineers to compile, test, and run accelerated code locally without relying on remote virtual machines.

At the other end of the computing spectrum, the release includes an early developer preview for NVIDIA’s next-generation Rubin microarchitecture, which follows Blackwell. Under compute capability 107 (SM_107), engineers can compile code with the NVCC compiler before physical hardware reaches enterprise customers. NVIDIA has not published commercial shipping dates or silicon benchmarks for the upcoming chips.

Slicing and mathematical changes in CUDA 13.4

For multi-tenant cloud clusters, NVIDIA overhauled its Multi-Process Service with MPS V3. The update allows operators to partition shared graphics cards across isolated containers with explicit memory quotas and Streaming Multiprocessor (SM) limits, configured through TOML files and Linux cgroups. Inter-node networking also gains the Compute Fabric Transport (CFT) Driver API, designed for asynchronous transfers across NVLink meshes while minimizing pressure on virtual address space.

In the mathematical runtime, the cuBLAS library integrated the Ozaki Scheme II algorithm. The method splits calculations across lower-precision tensor units to emulate double-precision (FP64) floating-point arithmetic. This enables teams to run high-accuracy scientific simulations on hardware built primarily for deep learning, avoiding the premium required for dedicated FP64 supercomputing chips.

Workflow shifts and developer agent tooling

The release also alters deployment routines and execution speeds for frontier models. Starting with CUDA 13.4 on Linux (a standard in place on Windows since version 13.1), the standalone installer no longer packages display drivers. System administrators must install driver branch R615 or newer in a separate step.

For Mixture-of-Experts (MoE) models, which activate specific subsets of parameters per token, cuBLASLt received dynamic scheduling for Grouped GEMMs. Independent benchmarks published by FullStack Labs reported throughput increases of up to 20% on data center Blackwell GPUs when processing multiple workload groups.

Beyond low-level libraries, NVIDIA introduced an official Model Context Protocol (MCP) server. The tool connects autonomous coding agents directly to verified CUDA documentation and code examples, helping assistants generate valid syntax within local programming environments.

Practical impact across the AI stack

The updates in CUDA 13.4 address practical friction points across real-world AI operations. Native support for Arm laptops keeps engineers inside NVIDIA’s development ecosystem from their personal machines. At the same time, tighter GPU partitioning and MoE performance optimizations lower day-to-day hosting bills while enterprises wait for next-generation silicon to arrive.