CUDA Toolkit 13.4, available as version 13.4.1, adds Windows on Arm support for NVIDIA RTX Spark systems and introduces developer-preview support for Rubin GPUs. For developers, the immediate significance is a broader native CUDA toolchain ahead of RTX Spark’s expected October arrival.
CUDA 13.4 Brings Windows on Arm Support to RTX Spark
NVIDIA’s CUDA 13.4 release notes identify Windows on Arm support for RTX Spark devices. The component table lists Windows Arm64 versions of NVCC, the CUDA runtime, cuBLAS, cuFFT, and Nsight development tools.
This extends CUDA’s existing Arm support beyond Linux to supported Windows systems. It does not mean CUDA can now run on every Windows on Arm laptop: execution still requires a supported NVIDIA GPU and a compatible driver.
The toolchain covers several essential development tasks. NVCC compiles CUDA code, the runtime and mathematical libraries support application execution, and Nsight tools help developers investigate performance. NVIDIA also confirms Windows on Arm support for the N1X laptop ecosystem in its core CUDA math libraries.
Native toolkit availability removes an important adoption barrier, but application portability depends on the rest of the software stack. Developers still need compatible Arm64 dependencies, build configurations, and any third-party libraries or extensions their applications use.
RTX Spark Requires a Separate Windows Driver
NVIDIA specifies Windows driver 616.41 or later for RTX Spark in the CUDA 13.4 release notes. Developers should use a driver that supports their specific system and intended CUDA features.
The GPU driver is a separate installation requirement. According to NVIDIA’s Windows installation guide, Windows CUDA Toolkit installers have not bundled the NVIDIA driver since CUDA 13.1. Installing the toolkit alone therefore does not complete the setup.
RTX Spark Combines Grace CPUs and Blackwell RTX Graphics
NVIDIA’s RTX Spark specifications describe N1X systems combining Grace CPU cores, Blackwell RTX graphics, and unified LPDDR5X memory. The published configurations distinguish between laptops and compact desktops:
| Specification | Laptop 6,144-core GPU | Laptop 5,120-core GPU | Desktop 6,144-core GPU |
|---|---|---|---|
| GPU architecture | Blackwell RTX | Blackwell RTX | Blackwell RTX |
| Grace CPU | 20 cores | 18 cores | 20 cores |
| Unified memory | Up to 128 GB LPDDR5X | Up to 64 GB LPDDR5X | Up to 128 GB LPDDR5X |
| Published TDP | 45–80 W | 45–80 W | 140 W |
These figures describe platform specifications. Actual memory options and system configurations depend on the manufacturer. Core counts and TDP ratings also do not establish application performance or measured wall-power consumption.
Unified Memory Offers Capacity, with Workload-Dependent Benefits
A shared CPU-GPU memory pool can be useful for applications with large working sets. It may reduce the need to maintain separate copies of data and accommodate models or datasets that exceed the local memory capacity of some discrete graphics cards.
However, unified memory does not automatically eliminate data movement or synchronization costs. The benefit depends on memory bandwidth, access patterns, allocation behavior, and how the application coordinates CPU and GPU work.
The advertised capacity is also shared with the operating system and other applications. A system with 128 GB of unified memory does not make that entire amount available exclusively for model weights or GPU allocations.
ITD Insight
RTX Spark’s development story depends on both hardware and software. A large shared memory pool can accommodate demanding workloads, while native CUDA tools provide a path to building and optimizing applications. The practical test is whether the complete application stack works reliably on Windows Arm64 and delivers useful sustained performance.
Rubin Support Remains a Developer Preview
Separately, CUDA 13.4 introduces developer-preview support for NVIDIA’s Rubin GPU architecture, including the sm_107 target. This gives developers an early path to prepare code and architecture-specific build configurations.
NVIDIA explicitly states that the Rubin preview is not intended for benchmarking, performance analysis, or production deployment. It should therefore be treated as preparation support rather than evidence of finished-platform performance.
This announcement is distinct from RTX Spark’s hardware specifications. RTX Spark uses Blackwell RTX graphics; the addition of Rubin support to the same toolkit does not imply that RTX Spark contains Rubin silicon.
October Availability and the Next Development Steps
NVIDIA’s September 3 announcement says RTX Spark Windows PCs are coming in October 2026. Buyers should check individual manufacturers for model-specific launch dates, regional availability, configurations, and pricing.
For developers evaluating the platform, the most useful next steps are to:
- Check dependencies: Confirm that required libraries, frameworks, and extensions support Windows Arm64.
- Validate the toolchain: Build a representative application and verify the required Nsight profiling and debugging workflows.
- Install a supported driver: Follow NVIDIA’s RTX Spark driver requirements and the system manufacturer’s guidance.
- Measure realistic workloads: Test memory usage, responsiveness, throughput, and sustained power behavior with the intended application.
CUDA 13.4 gives RTX Spark a more concrete software foundation ahead of launch. Its Windows on Arm support is a meaningful compatibility milestone, while Rubin remains an early development target. The next measure of progress will be how well complete applications perform on shipping RTX Spark systems.