SerpApi.Org
NVIDIA has released CUDA Toolkit 13.4 as a developer preview — and with it, the company has done something it has never done before: shipped an official, first-party CUDA development package that natively targets Windows Arm64. The release closes a gap that made the entire Windows-on-Arm platform effectively invisible to the AI and scientific computing ecosystem, because every serious workload that touches CUDA — PyTorch, TensorRT, cuDNN, RAPIDS, most of the production AI stack — had no path to run on a Windows Arm machine.
What makes that gap significant is not just the inconvenience. CUDA is NVIDIA's proprietary parallel computing platform, and it is the exclusive computing substrate for most of the world's AI development and inference work. AMD ROCm and Intel OneAPI offer alternatives, but the overwhelming majority of AI frameworks default to CUDA. By extending CUDA to Windows Arm64, NVIDIA is not merely filling a feature hole — it is extending its proprietary software moat into the one remaining Windows device category where it had no presence. RTX Spark laptops and mini PCs will be the only Windows-on-Arm machines where that software stack runs without workarounds. That is an ecosystem positioning move as much as a toolkit release.
Two Concrete Build Paths, Available Today
CUDA Toolkit 13.4 gives developers two immediately usable routes to Windows Arm64 CUDA applications:
The first is a fully native Arm64 build: compile CUDA applications directly on a Windows Arm64 machine, with NVCC (NVIDIA's LLVM-based C/C++ compiler) targeting the Arm64 instruction set and producing code that runs without any emulation layer. The second is cross-compilation: build Windows Arm64 CUDA binaries from a conventional x86-64 Windows machine using the existing toolkit, which lets developers work from any desktop they already have while generating output for the new architecture.
NVIDIA is providing a local Windows Arm64 installer and binary archives containing CUDA headers, import libraries, runtime components, and the linking libraries required for full application builds. The toolkit includes NVCC 13.4.46, Nsight Compute, Nsight Systems, cuBLAS 13.7.0, cuFFT, cuRAND, cuSOLVER, cuSPARSE, and NPP — the same core library set available on other platforms.
A third intermediate path also deserves mention for developers with mixed codebases. Arm64EC (Emulation Compatible) allows native Arm64 code and emulated x64 code to run inside the same process, communicating through a shared ABI. NVIDIA explicitly supports this mode in 13.4, which means developers can convert their applications piece by piece — shipping a hybrid build that runs CUDA-accelerated paths natively while keeping legacy x64 components running through Windows' Prism emulation layer until a full port is complete.
Why the Driver Model Changed
CUDA Toolkit 13.4 also marks a structural change in how NVIDIA ships its developer tooling, one that developers should understand before installation.
Starting with this version, the GPU driver is no longer bundled with the CUDA Toolkit on Linux — it ships separately as NVIDIA's standalone Developer Driver (R616 and above). On Windows, this decoupling started earlier with CUDA 13.1, but the Linux change is new in 13.4. The preview driver, version 616.00, is already available through the NVIDIA Developer portal and is initially targeted at Microsoft Surface RTX Dev Box hardware.
The practical implication: developers cannot simply run the standard CUDA Toolkit installer and expect a working GPU environment. They must separately download and install an R616-series Developer Driver (version 616.00 or later) to access CUDA 13.4 features and the newly enabled Windows Arm64 platform. Existing CUDA 13.x applications will continue to run on any driver version at or above 580 through CUDA minor-version compatibility, but new platform-specific features require the R616 series.
What RTX Spark Actually Is
The CUDA 13.4 release is directly timed to NVIDIA's RTX Spark product roadmap, announced by CEO Jensen Huang during his Computex 2026 keynote in Taipei on June 1.
RTX Spark is a two-chiplet system-on-chip fabricated on TSMC's 3-nanometer process. A Blackwell GPU die carries 6,144 CUDA cores across 48 streaming multiprocessors — the same shader count as a desktop GeForce RTX 5070 — alongside fifth-generation Tensor Cores supporting FP4 (4-bit floating point) precision. A MediaTek-designed CPU die houses 20 Arm cores: ten Cortex-X925 performance cores and ten Cortex-A725 efficiency cores, a configuration drawn from high-end mobile silicon and scaled for PC workloads.
The two dies connect via NVIDIA's NVLink-C2C chip-to-chip interconnect, running at 600 GB/s of bidirectional bandwidth between CPU and GPU — approximately five times the bidirectional bandwidth of PCIe Gen 5. This coherent interconnect allows the CPU and GPU to share the same memory address space directly, which is the architectural underpinning of the unified memory design.
That unified memory pool reaches up to 128 GB of LPDDR5X, with the memory subsystem delivering approximately 300 GB/s of bandwidth to the chip. It is worth being precise about these two figures: the 600 GB/s describes how fast data moves between the CPU and GPU dies over NVLink-C2C; the 300 GB/s describes how fast the chip can stream data out of the physical LPDDR5X memory chips. For large language model inference — which is fundamentally a memory-bandwidth-limited workload — the 300 GB/s figure governs token generation speed. NVIDIA rates the platform at 1 petaFLOP of AI compute, but that figure is measured at FP4 precision with sparsity enabled — a best-case theoretical ceiling, not a sustained general-compute figure.
NVIDIA claims the 128 GB memory pool is sufficient to run local AI reasoning models with up to 120 billion parameters without relying on cloud inference — a figure specific to the laptop SKU. The DGX Spark desktop module, which uses the same GB10 die, targets 200 billion parameter models at its full 128 GB configuration.
RTX Spark systems are expected to ship in fall 2026 across premium laptops and compact desktop mini PCs, with confirmed OEM partners including ASUS, Dell, HP, Lenovo, and MSI. Microsoft has also built kernel-level Windows 11 optimizations specifically for RTX Spark — changes that were not extended to Qualcomm's Snapdragon platforms — and is co-developing OpenShell, a pre-1.0 agentic AI runtime that provides kernel-level sandboxing, a YAML policy engine, and a Privacy Router for masking personally identifiable information before routing to cloud models.
Inverting the Platform Launch Sequence
NVIDIA's decision to ship the CUDA toolkit and developer driver before retail RTX Spark hardware represents a deliberate reversal of the typical platform launch sequence, and it matters for understanding why the release is structured as it is.
Historically, a new processor platform follows a predictable arc: hardware ships first, drivers follow weeks or months later, and developer tooling arrives last, often after early users have already had a disappointing experience. Windows on Arm's troubled history is largely a story of exactly this failure mode — each attempt at the platform has run into software ecosystem gaps that were not addressed before hardware landed in users' hands.
With CUDA 13.4, NVIDIA is compressing that sequence. According to NVIDIA's own guidance, developers do not need RTX Spark hardware to begin porting. Any existing Windows-on-Arm development system — a Surface Pro X, a Snapdragon X Elite laptop — can be used today to audit third-party dependencies for Arm64 compatibility, choose between a native Arm64 or Arm64EC build approach, and validate the build before RTX Spark devices ship. The R616.00 preview driver provides GPU testing capability for developers who already have a Windows-on-Arm machine with a compatible display.
Where RTX Spark Stands Against Its Competition
RTX Spark enters a market that Qualcomm's Snapdragon X Elite and X2 Elite have been developing for roughly two years. The competitive picture is uneven by workload category.
On GPU compute and AI inference, RTX Spark's advantage is structural: CUDA, RTX graphics with DLSS 4.5 support, and 1,000+ TOPS of AI performance (at FP4) against Qualcomm's approximately 80 TOPS NPU and Apple's 38 TOPS Neural Engine. Adobe is rebuilding Photoshop and Premiere Pro to run natively on RTX Spark, signaling that creative-professional workflows are being designed around the platform.
On CPU single-core performance, the picture is less clear. Qualcomm's custom Oryon Gen 3 cores in the Snapdragon X2 Elite Extreme match Apple M4 Pro in several single-core benchmarks, while RTX Spark's Cortex-X925-based CPU tile — co-designed with MediaTek rather than built in-house — has not yet been validated in shipping hardware. Apple's M-series chips also hold a memory bandwidth advantage: the M4 Max delivers approximately 546 GB/s of memory bandwidth against RTX Spark's 300 GB/s, which matters for large language model token generation speed even though RTX Spark offers more total memory capacity.
One limitation cuts across all Windows-on-Arm platforms, including RTX Spark: the OS cannot emulate kernel-mode drivers, meaning games and applications that rely on kernel-level anti-cheat software may not run at all — regardless of how capable the GPU is. NVIDIA's driver team's maturity and its history of shipping reliable GPU drivers on Windows gives RTX Spark an advantage over Qualcomm's Adreno on that dimension, but it does not solve the structural anti-cheat problem.
Known Limitations in the Preview
Several known issues come with this preview release. CUDA memory transfers using pageable (non-pinned) host memory may perform below expectations; NVIDIA recommends using page-locked (pinned) host memory wherever possible to work around this limitation. PyTorch CI/CD build workflows may trigger GPU timeouts or system instability when run on Windows Arm64 with the preview driver. During driver installation, the built-in display may remain blank for approximately two minutes — a known quirk of the R616 preview driver that should be expected rather than interpreted as a failure. Finally, Nsight Copilot, NVIDIA's AI-assisted profiling tool, is not available on Windows Arm64 in this release.
NVIDIA has not announced a timeline for when Windows Arm64 support will graduate from developer preview to a general-availability CUDA release. RTX Spark hardware itself does not yet have a confirmed retail release date beyond the fall 2026 window.
A Multi-Generation Bet
Jensen Huang has publicly committed to a multi-generation RTX Spark roadmap. A second-generation platform called Vera Rubin Spark will incorporate a new CPU architecture and LPDDR6 memory. A third-generation platform follows after that. The intent is to signal to OEM partners and software developers that the RTX Spark platform is a sustained commitment rather than a one-cycle experiment — the same assurance that Apple gave developers when it committed to Apple Silicon in 2020 and Amazon gave cloud customers when it committed to multi-generation Graviton roadmaps.
For now, the message CUDA 13.4 sends to the developer community is concrete: start porting. The hardware is coming this fall, NVIDIA is betting the ecosystem will be ready before it arrives, and the toolkit to begin that work is available today on the NVIDIA Developer portal.
Frequently Asked Questions
Do I need an RTX Spark laptop to start porting my CUDA application to Windows Arm64?
No. NVIDIA's guidance is that any existing Windows-on-Arm machine — including current Snapdragon X Elite or Surface devices — is sufficient to audit dependency compatibility, choose between a native Arm64 or Arm64EC build approach, and compile initial builds. You will need to install the separate R616 Developer Driver (version 616.00, available on the NVIDIA Developer portal) to access GPU-accelerated testing. RTX Spark hardware is needed only for final validation against the actual target platform, which NVIDIA expects will be available in fall 2026.
What is the difference between the 600 GB/s and 300 GB/s bandwidth figures for RTX Spark?
These describe two different parts of the chip's memory system and should not be conflated. The 600 GB/s figure is the NVLink-C2C interconnect bandwidth — the bidirectional data rate between the Grace CPU die and the Blackwell GPU die inside the package. The 300 GB/s figure is the LPDDR5X memory bandwidth — the rate at which the chip reads data from the physical memory chips. For large language model inference, where the bottleneck is streaming model weights from DRAM rather than moving data between CPU and GPU, the 300 GB/s figure is the binding constraint. NVIDIA's stated 1 petaFLOP AI compute rating refers specifically to FP4 precision with sparsity — a theoretical maximum, not a sustained general compute figure.
What does it mean that CUDA is proprietary — and why does that matter for RTX Spark?
CUDA is NVIDIA's parallel computing platform and is only available on NVIDIA GPUs. Unlike AMD ROCm or Intel OneAPI, which are open-source and can run on multiple vendors' hardware, CUDA cannot be implemented by third parties. This means the AI and scientific computing software that runs on CUDA — PyTorch's default GPU backend, TensorRT, cuDNN, RAPIDS, and most production inference pipelines — works only on NVIDIA hardware. By bringing CUDA to Windows Arm64, NVIDIA is making RTX Spark the only Windows-on-Arm platform where that dominant software stack runs natively. Developers choosing between an RTX Spark device and a Snapdragon X laptop are not choosing between comparable CUDA options — only RTX Spark has CUDA at all.
Is CUDA Toolkit 13.4 safe to install on my production development machine?
No — and NVIDIA states this explicitly. Toolkit 13.4 is a developer preview designated as pre-release software. It must not be used for production deployment, certification, benchmarking, or business-critical systems. Performance data collected with this release is non-representative and must not be published, compared, or used to characterize NVIDIA hardware or software. NVIDIA has not announced a timeline for a general-availability release of Windows Arm64 CUDA support.
Related Articles
