NVIDIA has opened the door to running CUDA natively on Windows machines built around Arm chips for the first time. The change landed in a blog post from NVIDIA on September 9, 2026, titled “CUDA Toolkit 13.4 Adds Windows on Arm Support and Greater Control Over Shared GPUs.” It is a narrow-sounding update with a wide blast radius: developers writing GPU-accelerated code for Arm-based Windows PCs no longer have to fake it through emulation or skip Windows entirely and reach for Linux.
CUDA, NVIDIA’s parallel computing platform, has powered the vast majority of AI training and inference work for more than a decade. Until this release, that power came with a catch on Arm-based Windows systems: the toolkit simply wasn’t built for them. NVIDIA’s own announcement acknowledges the gap directly, noting that CUDA applications had long been supported on Arm platforms through Linux, and that CUDA Toolkit 13.4 extends that capability to the Windows on Arm platform for the first time.
This is a hardware story before it’s a software story. Windows on Arm has struggled for years to become a serious platform for AI development, mostly because the GPU tooling that data scientists and ML engineers rely on day to day never showed up. CUDA Toolkit 13.4 changes that calculus, at least on paper, and it does so at a moment when NVIDIA is pushing hard on Arm-based compute across its entire product line, from data-center Grace CPUs down to desktop-class RTX Spark devices.
What NVIDIA Actually Announced on September 9
The announcement itself is compact. In its September 9 blog post, NVIDIA states plainly that “CUDA Toolkit 13.4 adds support for Windows on Arm.” The same post frames the change historically, noting that “CUDA applications have long been supported on Arm platforms through Linux; this release extends that capability to the Windows on Arm platform.” NVIDIA’s own CUDA Toolkit product page reinforces the point, describing CUDA 13.4 as bringing “a host of advancements to the platform including CUDA support for the Windows on Arm ecosystem and developer support for the NVIDIA Rubin Architecture.”
The CUDA documentation is more specific about which hardware benefits first. NVIDIA’s CUDA documentation states that “the CUDA Toolkit is now supported on Windows on Arm platform for RTX Spark devices,” and the CUDA Toolkit 13.4 release notes echo that same line almost verbatim: “CUDA 13.4 adds support for Windows on Arm on RTX Spark devices.” RTX Spark is NVIDIA’s compact desktop AI workstation platform, and tying the initial Windows-on-Arm rollout to it suggests NVIDIA is treating this as a targeted enablement for a specific device category rather than a blanket claim that every Arm-based Windows laptop on the market instantly gets full CUDA support.
The release notes for CUDA NVCC 13.4.59 list supported architectures as x86_64, arm64-sbsa, and arm64 (Windows), spanning both Linux and Windows. That’s a meaningful detail for anyone tracking the toolkit’s evolution: arm64-sbsa (Server Base System Architecture) has been NVIDIA’s Linux-side Arm server target for a while, and arm64 (Windows) is the genuinely new column added in this release.
Beyond those specific statements, the public record thins out fast. There is no confirmed detail yet on exactly when general availability lands beyond the developer-preview label attached to the release, nor firm figures on how many RTX Spark units are in developers’ hands. Treat any claim you see elsewhere about adoption numbers or a hard GA date with some skepticism until NVIDIA publishes something more concrete.
Why Windows on Arm Needed This
Windows on Arm has existed in some form since Microsoft’s Surface RT experiment more than a decade ago, and it has spent most of that time as a compromise platform: good battery life, mediocre app compatibility, and a persistent absence of the professional tooling that developers actually reach for. GPU compute was one of the biggest holes. CUDA never ran natively there, which meant anyone doing serious AI or scientific computing work on an Arm Windows machine had two bad options: run x86 code through emulation and eat the performance penalty, or dual-boot into Linux and lose the rest of the Windows ecosystem.
NVIDIA’s own framing of the CUDA 13.4 release leans into this history. By explicitly calling out that Arm support already existed on Linux, the company is signaling that Windows on Arm was the missing piece, not Arm computing generally. NVIDIA has shipped Arm-based data center silicon for years through its Grace CPU line, and Grace already pairs with Blackwell GPUs in superchip configurations used in AI infrastructure. What changed on September 9 is that a Windows-native build path for that same Arm+CUDA combination now exists, at least for RTX Spark hardware.
That timing lines up with a broader industry push toward Arm-based PCs. Qualcomm has spent several product cycles trying to convince Windows OEMs and developers that Arm laptops are viable daily drivers, and Apple’s own shift to Arm-based Apple Silicon proved that a mainstream OS vendor could make the architecture switch work at scale. NVIDIA extending CUDA to Windows on Arm doesn’t erase the years of fragmented tooling that came before it, but it does close one of the more consequential gaps: the one that kept AI developers away from Arm Windows machines specifically because their GPU libraries wouldn’t compile.
Inside CUDA Toolkit 13.4: What’s New for Developers
For developers who actually have to build against this toolkit, the practical question is simple: what changes in my workflow? Based on NVIDIA’s release notes and documentation, the arm64 (Windows) target now sits alongside the existing x86_64 and arm64-sbsa targets across the CUDA C++ Core Compute Libraries, meaning components like Thrust, CUB, and libcu++ list Windows Arm64 as a supported build target. NVIDIA’s tooling around Visual Studio integration also picks up arm64 (Windows) as a listed platform in the 13.4 release notes, which matters for anyone doing CUDA debugging work inside a native Windows IDE rather than a cross-platform editor.
NVIDIA’s blog post frames the change as extending a capability, not inventing one from scratch: CUDA applications had long compiled and run on Arm through Linux, and 13.4 essentially opens a second front on Windows. For teams already maintaining CUDA codebases that target Arm-based Linux servers, that should mean less rework than starting a Windows Arm64 port cold, though NVIDIA has not published detailed porting guidance beyond the toolkit documentation itself.
The other half of the September 9 announcement, tucked into the same blog post title, covers greater control over shared GPUs. NVIDIA didn’t pair those two features by accident. RTX Spark and similar compact AI workstations are often used by multiple developers or workloads sharing one physical GPU, and better partitioning controls are the kind of feature that becomes more urgent exactly when you’re trying to get more people building on a new platform. NVIDIA has not published granular technical specifics on the shared-GPU controls beyond the announcement itself, so treat any detailed breakdown of scheduling behavior as speculative until the documentation catches up.
CUDA’s Platform History: From Linux-Only Arm to Windows
CUDA launched in 2007 as an x86-first platform, tightly coupled to NVIDIA’s desktop and workstation GPU lineup. Arm support came later and stayed Linux-bound, aimed mostly at embedded and edge devices (NVIDIA’s Jetson line) and, more recently, Arm-based Linux servers running NVIDIA’s Grace CPUs. Windows, by contrast, remained an x86-only story for CUDA for essentially the platform’s entire history.
That split made sense when Arm-based Windows PCs were a niche curiosity. It stopped making sense once NVIDIA started building its own Arm+GPU superchip hardware and marketing it as a developer workstation rather than a data-center-only product. RTX Spark is the clearest expression of that shift: a compact machine pairing an NVIDIA GPU with a Grace-derived Arm CPU, aimed at developers who want data-center-style AI hardware on a desk rather than in a rack. Once that hardware exists and ships with Windows as an option, the absence of native CUDA support on Windows Arm64 becomes a much harder gap to justify.
The table below lays out how CUDA’s platform coverage has evolved at a high level, based on NVIDIA’s own toolkit documentation and release notes.
| Platform | Architecture | Status Before CUDA 13.4 | Status as of CUDA 13.4 |
|---|---|---|---|
| Linux | x86_64 | Fully supported | Fully supported |
| Linux | arm64-sbsa | Supported (server-class Arm) | Supported |
| Windows | x86_64 | Fully supported | Fully supported |
| Windows | arm64 (Windows on Arm) | Not supported | Supported for RTX Spark devices (developer preview) |
RTX Spark: The Hardware Anchor for This Release
Every specific reference to Windows on Arm in NVIDIA’s CUDA 13.4 documentation ties back to RTX Spark devices by name. That’s a deliberate scoping choice, and it tells you something about how NVIDIA is rolling this out: rather than declaring open-ended Windows on Arm support across any Arm chip that happens to run Windows, NVIDIA is anchoring the initial support window to its own hardware platform, where it controls the GPU, the driver stack, and increasingly the CPU architecture as well.
RTX Spark has already been through several rounds of coverage on this site, from its initial multi-OEM rollout to pricing leaks and sellout reports as demand outpaced supply. What’s new here isn’t the hardware, it’s the software layer finally catching up to support the full Windows-based developer experience on that hardware’s Arm CPU side, not just its GPU. For developers who bought into RTX Spark expecting a Linux-first workflow, CUDA 13.4 is the first sign that NVIDIA intends the platform to work equally well for people who want to stay on Windows.
It’s worth being precise about what’s confirmed versus what’s reasonable inference here. NVIDIA’s own materials confirm the Windows on Arm CUDA support is tied to RTX Spark devices and is currently a developer preview. Specific driver version numbers, exact CUDA core counts, and detailed compute-capability figures tied to future Rubin-architecture hardware are the kind of granular specs that shift between preview and general availability, so this piece sticks to what NVIDIA has put on the record rather than projecting forward.
Competitive Landscape: Where This Leaves Apple, Qualcomm, and AMD
NVIDIA isn’t the only company with a stake in Arm-based Windows computing, and this announcement puts it in a genuinely different position from its neighbors. Apple made the cleanest Arm transition in the PC industry with Apple Silicon, but that ecosystem runs on Metal, not CUDA, and Apple has no interest in bringing NVIDIA’s stack to macOS. Qualcomm has spent multiple generations of Snapdragon chips trying to make Arm-based Windows laptops mainstream, but Qualcomm doesn’t build discrete GPUs at NVIDIA’s scale, so its Arm PCs have never been the natural home for CUDA workloads regardless of what Windows itself supports.
That leaves NVIDIA in an unusual spot: it’s extending a Windows-Arm capability that mostly benefits its own hardware, on a platform (Windows on Arm) that Microsoft has been pushing for years with mixed third-party enthusiasm. AMD, meanwhile, has no equivalent CUDA-style ecosystem lock-in to defend or extend, since its ROCm platform has taken a different path and doesn’t carry the same multi-year head start in AI tooling that CUDA has built up.
| Vendor | Arm Windows Strategy | GPU Compute Stack | Native CUDA on Windows Arm64 |
|---|---|---|---|
| NVIDIA | RTX Spark (Grace-based Arm CPU + NVIDIA GPU) | CUDA | Yes, developer preview (CUDA 13.4) |
| Apple | Apple Silicon, macOS only | Metal | No (different OS, no CUDA on macOS) |
| Qualcomm | Snapdragon-based Windows on Arm laptops | No discrete CUDA-class GPU stack | Not applicable to Qualcomm’s own silicon |
| AMD | Primarily x86 Windows PCs | ROCm | No (not a CUDA licensee platform) |
The practical upshot is that this isn’t really a four-way race for Windows-on-Arm CUDA support. It’s NVIDIA deciding whether its own Arm-based hardware gets full Windows tooling, and the September 9 announcement answers that question in the affirmative, at least for RTX Spark and at least in preview form.
Market Impact: What This Means for AI Developers and Enterprises
The most immediate beneficiaries are developers who already own or plan to buy RTX Spark hardware and want to keep working inside Windows rather than switching to Linux for CUDA compatibility. That’s a smaller group than “everyone building AI software,” but it’s not a trivial one. Enterprise developers, in particular, often can’t simply swap operating systems because of IT policy, existing tooling, or compliance requirements, so a platform that forced a Linux dual-boot to get full GPU compute support was effectively closed off to a chunk of the corporate developer base.
There’s a second-order effect worth watching too. NVIDIA pairing this release with developer support for its upcoming Rubin architecture suggests the company is using the Windows on Arm rollout as a proving ground ahead of a larger platform transition. Getting the toolkit, driver stack, and Visual Studio integration working cleanly on Windows Arm64 now, while the install base is still small and the stakes are lower, reduces the risk of larger problems surfacing once Rubin-based hardware ships in volume.
For competitors, the message is less about an immediate market shift and more about NVIDIA continuing to close gaps in its own ecosystem before rivals can exploit them. Windows on Arm has been an underserved developer platform for years. NVIDIA choosing to serve it, even narrowly, removes one more reason for developers evaluating Arm-based Windows hardware to look elsewhere for GPU compute support.
What’s Still Unconfirmed
It’s worth being straightforward about the limits of what’s public right now. NVIDIA has not confirmed whether CUDA 13.4 represents the first time any native CUDA support has existed on Windows on Arm across the platform’s history, versus simply the first time NVIDIA has formally documented and shipped it as a supported target; the company’s own wording focuses on extending existing Arm support from Linux to Windows, without directly addressing prior unofficial or experimental builds.
There’s also no confirmed general-availability date beyond the developer-preview status of the current release, no official pricing tied specifically to this software update (RTX Spark hardware pricing is a separate, previously reported topic), and no NVIDIA statement quoted from a named executive about the rollout’s strategic rationale. Readers should treat any article claiming otherwise with caution until NVIDIA publishes a follow-up with those specifics.
Historical Parallel: Windows RT and the Arm Adoption Problem
Anyone who’s watched Microsoft’s Arm ambitions over the years will recognize the shape of this problem. Windows RT, launched alongside the original Surface tablets, failed largely because developers wouldn’t build for a platform that couldn’t run their existing software or tools. Microsoft’s own Windows on Arm documentation has spent years chronicling the slow process of getting native Arm64 builds of everyday developer tools, browsers, and now GPU compute stacks onto the platform. Each missing piece became a reason for developers to stick with x86 hardware, and each piece that got filled in removed one more excuse.
CUDA support is arguably one of the more consequential pieces to fill, because AI and data science work has become such a large share of what professional developers actually build day to day. A platform that can’t run the dominant GPU compute toolkit isn’t just missing a nice-to-have, it’s excluded from an entire category of modern software development. NVIDIA closing that gap, even in a limited preview tied to specific hardware, changes the calculus for anyone deciding whether Windows on Arm is worth targeting.
What to Expect Next
Based on the pattern of NVIDIA’s own release notes and the pairing of this update with early Rubin architecture developer support, a few reasonable expectations follow, though none of these are confirmed by NVIDIA and should be read as informed forecasting rather than fact.
- CUDA’s Windows on Arm support will likely broaden beyond RTX Spark to additional Arm-based Windows hardware as the toolkit moves out of developer preview, following the same pattern CUDA has historically followed on new platforms.
- NVIDIA will probably publish more detailed porting and compatibility guidance for existing CUDA codebases as adoption grows, since the current announcement leans on documentation rather than a dedicated migration guide.
- Expect closer alignment between CUDA’s Windows Arm64 rollout and the Rubin architecture’s public launch timeline, given that both were referenced together in the same release.
- Competitive pressure from Qualcomm and Microsoft’s own Windows on Arm push could accelerate NVIDIA’s timeline for moving this support out of preview status.
- Enterprise IT departments evaluating Arm-based Windows workstations for AI development teams are likely to treat this as a green light to pilot RTX Spark hardware, though full production rollouts will probably wait for a non-preview toolkit release.
How This Fits NVIDIA’s Broader Arm Strategy
NVIDIA’s Grace CPU line has already made the company a serious player in Arm-based data center compute, and pairing Grace-derived silicon with Blackwell and future Rubin GPUs has become NVIDIA’s signature move for high-end AI infrastructure. RTX Spark takes that same Arm-plus-NVIDIA-GPU formula and shrinks it down to desk-side hardware. CUDA 13.4’s Windows on Arm support is the software piece that makes that smaller-scale hardware fully usable inside the operating system most developers already run day to day.
Seen that way, this isn’t a standalone feature update so much as the next step in a strategy NVIDIA has been building for several product cycles: control the CPU architecture, control the GPU, and control the software stack that ties them together, on every scale from data-center rack to developer desktop. Extending that control to Windows, rather than leaving Windows-based developers stuck on x86 or forced into Linux, closes one of the last remaining gaps in that plan.
Whether that strategy pays off depends heavily on factors outside NVIDIA’s direct control, including how aggressively Microsoft continues investing in Windows on Arm as a platform, and how the broader Arm PC market (spanning Qualcomm-, Apple-, and increasingly NVIDIA-powered devices) shakes out over the next couple of product cycles. For now, the concrete fact is narrower and more verifiable: as of September 9, 2026, CUDA developers targeting RTX Spark hardware running Windows on Arm have a supported, if still preview-stage, path to build native applications without emulation or a Linux detour.
Frequently Asked Questions
What did NVIDIA announce about CUDA and Windows on Arm?
On September 9, 2026, NVIDIA published a blog post confirming that CUDA Toolkit 13.4 adds support for Windows on Arm, extending a capability that had previously existed only on Linux-based Arm systems.
Which hardware is supported first?
NVIDIA’s documentation and release notes specifically tie the new Windows on Arm CUDA support to RTX Spark devices. The company has not stated that all Arm-based Windows PCs gain full CUDA support through this release.
Is this a full, production-ready release?
NVIDIA has not described CUDA 13.4’s Windows on Arm support as a finished, general-availability feature; the toolkit page and release notes frame it within the broader 13.4 release, which also introduces developer support for the upcoming Rubin architecture, itself an early-stage capability.
Could CUDA run on Windows on Arm before this update?
According to NVIDIA’s own announcement, CUDA had long been supported on Arm platforms through Linux, but not on Windows running on Arm hardware. CUDA Toolkit 13.4 is the release that extends support to that specific combination.
Does this affect Apple Silicon or Qualcomm Snapdragon devices?
No. CUDA is an NVIDIA-specific platform tied to NVIDIA GPUs. Apple Silicon runs macOS with Apple’s Metal framework, and Qualcomm’s Arm chips for Windows PCs don’t include NVIDIA discrete GPUs, so neither platform is directly affected by this CUDA update.
What is the Rubin architecture NVIDIA mentioned alongside this release?
NVIDIA’s CUDA Toolkit page lists developer support for the NVIDIA Rubin Architecture as one of the advancements shipping in CUDA 13.4, alongside Windows on Arm support. NVIDIA has not published detailed public specifications for Rubin-based consumer hardware as part of this announcement.
Where can developers find the official documentation?
NVIDIA’s CUDA documentation site and the CUDA Toolkit 13.4 developer preview release notes, both linked above, contain the authoritative technical details on supported architectures and platforms.
Will this expand to more Arm-based Windows hardware later?
NVIDIA has not made that commitment publicly. Given the pattern of past CUDA platform rollouts, which typically start narrow and broaden over subsequent releases, wider hardware support is a reasonable expectation but not a confirmed roadmap item.
