CUDA Toolkit 13.4 Adds Windows on Arm Support and Greater Control over Shared GPUs
CUDA Toolkit 13.4 extends CUDA application development to Windows on Arm, previously limited to Linux on Arm platforms Preview functional support for NVIDIA Rubin architecture (compute capability 107) enables early application porting ahead of general availability Multi-Process Service V3 introduces a modernized control layer with scriptable CLI, TOML configuration, and cgroup-integrated GPU memory limits for precise containerized GPU partitioning CUDA Compute Fabric Transport provides a transpo
65
Hot
70
Quality
75
Impact
Analysis
TL;DR
- CUDA Toolkit 13.4 extends CUDA application development to Windows on Arm, previously limited to Linux on Arm platforms
- Preview functional support for NVIDIA Rubin architecture (compute capability 107) enables early application porting ahead of general availability
- Multi-Process Service V3 introduces a modernized control layer with scriptable CLI, TOML configuration, and cgroup-integrated GPU memory limits for precise containerized GPU partitioning
- CUDA Compute Fabric Transport provides a transport-centric API for moving data across NVLink fabric using named logical endpoints and asynchronous operations
- CCCL 3.4 delivers warp-specialized cub::DeviceScan on Blackwell GPUs reaching 92% memory-bandwidth utilization, while CUDA Python expands with texture/surface programming and NUMA-aware managed memory
Why It Matters
This release significantly broadens CUDA's platform accessibility by bringing Windows on Arm support, enabling developers on a growing ecosystem of Arm-based Windows devices to leverage GPU acceleration. The Rubin preview and MPS V3 enhancements signal NVIDIA's push toward next-generation agentic AI workloads and more efficient multi-tenant GPU sharing in cloud and containerized environments.
Technical Details
- Windows on Arm Support: CUDA applications previously supported on Arm only through Linux; this release extends full CUDA development capabilities to Windows on Arm, opening a new deployment surface for enterprise and consumer Arm-based systems.
- NVIDIA Rubin Preview (CC 107): Early developer access to the next-generation Rubin architecture, positioned as the GPU architecture powering the era of agentic AI, allowing application porting before general availability.
- MPS V3: Modernized GPU resource management with scriptable CLI, named server instances, namespaces, TOML configuration, SM partition controls, and cgroup-integrated GPU memory limits for precise compute performance, memory boundaries, and execution priority in containerized environments.
- CUDA Compute Fabric Transport (CFT): A Driver API-level transport-centric approach for communication-library developers to issue asynchronous put, get, and reduction operations across NVLink fabric using named logical endpoints, reducing virtual-address pressure and supporting unicast/multicast patterns with completion and error reporting.
- CUDA Python & CCCL 3.4: cuda.core 1.1.0 adds texture/surface programming, NUMA-aware managed memory, and .pyi type stubs; cuda.compute 1.1 enables ahead-of-time compilation of CCCL algorithms. CCCL 3.4 features single-call CUB device-wide APIs, batched warp reductions, and cuda::std parallel algorithms.
- Locality Domains & Unified Memory Queries: Programmatic access to GPU locality domains for co-locating computation near memory access; cudaMemGetLocationInfo API enables residency queries for unified memory to optimize scheduling and avoid unnecessary page migrations.
- Driver-Toolkit Decoupling: CUDA SDK installers no longer bundle the NVIDIA driver, allowing independent update cycles for the toolkit and driver components.
Industry Insight
- The Windows on Arm expansion positions NVIDIA to capture the growing market of Arm-based Windows devices, including potential integration with Qualcomm and other Arm CPU vendors, broadening the CUDA ecosystem beyond traditional x86/Linux deployments.
- MPS V3's cgroup integration and precise GPU partitioning capabilities signal NVIDIA's focus on multi-tenant cloud and containerized AI workloads, making GPU sharing more viable for cost-sensitive inference and development environments.
- The Rubin preview and CFT API indicate NVIDIA is preparing the software stack for large-scale agentic AI systems that require efficient multi-GPU communication and next-generation architecture support, urging communication-library developers to evaluate CFT for custom fabric-level optimizations.
Disclaimer: The above content is generated by AI and is for reference only.
GPU Chip Deployment Product Launch
Related Articles
Deploying Qwen3.8-2.4T-A95B on Amazon SageMaker HyperPod with vLLM
When to Use Encode-Prefill-Decode Disaggregation to Accelerate Multimodal Model Serving
ICYMI: What landed for AI builders in August 2026
Chipmaker Patch Tuesday: Nvidia, AMD, Arm Issue Security Advisories
[GitHub] tensorflow/tensorflow