NVIDIA RTX Spark Review: Windows on Arm Gets Serious Graphics and Local AI

NVIDIA’s RTX Spark gives Windows on Arm the graphics and shared memory needed for demanding creative work and local AI. In the HP OmniBook Ultra 16 RTX Spark Edition I used for about a week, the top configuration combined a 20-core Arm CPU, a Blackwell GPU with 6,144 CUDA cores and 128GB of unified memory in a 15.73mm-thin chassis. That’s a compelling package for creators and developers who need substantial computing power away from a desk, though Windows on Arm compatibility still calls for some care.

This was a hands-on review of a chip inside a laptop, rather than a full review of the platform across multiple systems. I spent my time in DaVinci Resolve, Affinity Photo, ON1 Photo RAW, LM Studio, Helium, Steam and Office apps. That practical mix made the strengths clear: RTX Spark brings serious GPU capability to a slim Arm laptop, while software translation and game support remain the areas to check before buying.

A shared memory pool with room for demanding workloads

RTX Spark is an Arm-based superchip developed by NVIDIA with MediaTek. Its CPU and GPU connect through NVLink-C2C, and the system uses unified LPDDR5X memory shared between them. The laptop version comes in configurations with an 18-core CPU and 5,120-core GPU, or a 20-core CPU and 6,144-core GPU. Memory can reach 128GB on the larger configuration; the 18-core version tops out at 64GB.

That shared pool is the central reason this design stands out for local AI and large creative projects. The source lists support for models of up to 120 billion parameters and a context window of up to 1 million tokens. It also describes running 70-billion- and 120-billion-parameter models locally, without relying on a cloud service. For 3D work, it says the system can load scenes as large as 90GB in Blender. These are capabilities the article attributes to the platform; they aren’t results from a detailed benchmark suite in this review.

There’s a practical trade-off: the memory is soldered to the chip package and can’t be upgraded later. Buyers need to choose capacity with future workloads in mind. The 128GB option makes the most sense for people who already know they need large local models, complex scenes or substantial video projects. It also pushes the price into premium territory, and the source gives no price for the reviewed HP configuration.

The chip includes fifth-generation Tensor Cores with FP4 support, fourth-generation ray-tracing cores and up to 1 petaflop of FP4 AI compute. It supports NVIDIA’s CUDA, TensorRT and OptiX tools, giving developers and creators access to software built around NVIDIA GPUs. That’s a meaningful change for Arm laptops, where graphics and specialist software support have often been the harder parts of the proposition.

More graphics, with laptop power limits

The reviewed HP configuration can draw up to 70 watts, within the listed 45–80W laptop range. HP uses two fans and three heat pipes in the OmniBook Ultra 16. The source reports a 99Wh battery rated for around 17 hours, and describes the laptop as quiet under use. Those battery and noise observations are tied to the HP system, not a guarantee for every RTX Spark laptop; cooling, power settings and battery capacity will vary by model.

RTX Spark also brings hardware ray tracing, DLSS and a dedicated media engine to the platform. The listed engine includes a ninth-generation NVIDIA encoder and sixth-generation decoder, with hardware support for 4:2:2 video, HEVC, AV1 and H.264. In a video workflow, dedicated encoding and decoding can help keep demanding timelines responsive while reducing the load on the CPU.

The article reports native 1440p gaming above 100 frames per second on modern titles, but doesn’t identify the games or test settings behind that figure. Treat it as a claim about supported titles, rather than a result you can expect across every game. RTX Spark is described as supporting DLSS 5, Reflex 2, G-SYNC and DirectX 12 Ultimate, alongside Vulkan ray tracing and OpenGL 4.6.

Some games are already available as native Arm builds, including Fortnite, VALORANT, League of Legends and PUBG. The source also names Cyberpunk 2077 and Gears of War: E-Day as confirmed, though it doesn’t provide release timing for those versions. Other multiplayer games remain blocked by anti-cheat compatibility, and some older x86 utilities may not work properly.

Compatibility is still part of the buying decision

Windows 11’s Prism emulation can run legacy x86 software, but translated apps may lose some performance and encounter driver problems. Native Arm applications are the better fit. Before buying, creators should check that their essential applications and hardware drivers support Windows on Arm; gamers should check each multiplayer title’s anti-cheat support. The source says Easy Anti-Cheat and BattlEye are supported, but that doesn’t resolve every game’s compatibility.

For that reason, the review’s enthusiasm comes with a clear audience in mind. Developers running local models, video editors working away from a workstation and power users who want CUDA in a portable system stand to gain the most. Competitive players whose game libraries depend on unsupported anti-cheat systems, and buyers looking for the lowest-cost option, have stronger reasons to stick with an x86 PC for now.

RTX Spark is also not a substitute for every desktop-class GPU. Its efficiency suits compact laptops and small desktop systems, but the source says it falls behind the sustained rasterization performance of discrete mobile RTX 5090 GPUs drawing 150–175 watts. That’s the expected compromise when a slim system prioritizes portability, lower power use and shared memory over the highest sustained graphics output.

Configurations and availability

NVIDIA announced RTX Spark on May 31, 2026, with Microsoft. The chip is sold as part of a laptop or compact desktop, not as a standalone component. The listed desktop configuration has a 20-core CPU and 6,144-core GPU, with a stated 140W power figure; the laptop variants are listed at 45–80W. The source says other brands will offer RTX Spark systems but doesn’t give a wider release schedule or regional availability.

Prices depend on the full system configuration. The source says several manufacturers open at $2,600, while higher-memory 64GB and 128GB systems cost more, but it doesn’t identify those models or provide prices for each tier. Those figures shouldn’t be taken as the price of the reviewed HP laptop.

My read after a week is that RTX Spark makes Windows on Arm a serious option for work that benefits from CUDA, capable graphics and a large shared memory pool. The HP OmniBook Ultra 16 shows what the top configuration can do in a portable design, but the non-upgradable memory and uneven app and game support make the choice highly dependent on your workload. Buy for the software and compute you can use today; treat broader compatibility as something to check, not a promise.

Key specifications

• CPU: Up to 20-core NVIDIA Grace Arm CPU; an 18-core laptop option is also listed.
• GPU: Blackwell RTX, with up to 6,144 CUDA cores; the 18-core laptop option has 5,120 GPU cores.
• Memory: Up to 128GB LPDDR5X unified memory; up to 64GB on the 18-core variant. Memory is not upgradeable.
• AI: Fifth-generation Tensor Cores with FP4; up to 1 petaflop FP4 compute and listed support for models up to 120 billion parameters with a 1-million-token context.
• Graphics: Fourth-generation ray-tracing cores, DLSS 5, Reflex 2, G-SYNC and DirectX 12 Ultimate.
• Media: One ninth-generation NVIDIA encoder and one sixth-generation decoder.
• Connectivity and software: PCI Express Gen 5, HDMI, DisplayPort, Windows 11, CUDA, TensorRT, OptiX and RTX Studio tools.
• Power: 45–80W for laptops; 140W for desktops, according to the listed specifications.