At Computex 2026 in Taipei, NVIDIA invited the press to a hands-on walkthrough of the RTX Spark platform, spread across multiple rooms covering gaming, creator, and AI workloads. Everything ran on Microsoft Surface Laptops, and Microsoft was also present—a clear signal that Redmond is fully committed to making Windows on Arm succeed this time around. Microsoft revealed that it has made several kernel-level optimizations to Windows 11 specifically for RTX Spark, changes that notably were never made for Qualcomm's Snapdragon platforms.

At the heart of RTX Spark is the NVIDIA N1X chip, which pairs a 20-core NVIDIA Grace Arm-based CPU with a Blackwell RTX GPU packing 6,144 CUDA cores and 5th Generation Tensor Cores with FP4 math precision, all connected via NVLink-C2C. The chip supports up to 128 GB of LPDDR5X unified memory in a 45-80 W power envelope (H-segment), and NVIDIA claims up to 1 petaFLOP/s of AI compute. A lower-end N1 variant with 5,120 CUDA cores is also planned. RTX Spark is not strictly a laptop platform either—NVIDIA also showed compact desktop mini PCs and the DGX Spark, targeting creators and developers who want the same silicon in a stationary form factor. And, looking at that power envelope, I wouldn't be surprised if N1X (or a variant) eventually ends up in a handheld gaming console—it just makes a lot of sense. A chip like this could also compete with semi-custom AMD chips for the next Xbox or PlayStation.

The N1X Chip

We got a close look at two N1X samples. The first shows the chip on a PCB, marked "GSE1-650-A1" with a date code of 2542 (week 42 of 2025), surrounded by eight Micron LPDDR5X memory packages. The memory is Micron "MT62F4G32D8DV-020 WT:F"—128 Gb (16 GB) per package running at 9600 MT/s, for a total of 128 GB unified memory, and a combined memory bandwidth of 307 GB/s. The orange border visible around the die is a protective film covering the substrate—not part of the silicon itself. The second sample clearly reveals the two-die design: the Grace CPU (E7A227) and Blackwell GPU (E7A214) sit side by side, connected via NVLink-C2C interconnect. This chiplet approach is what enables the unified memory architecture—similar in concept to Apple's M-series—eliminating the traditional bottleneck of shuttling data between separate CPU and GPU memory chips. Both chips are built on the TSMC 3 nm node.

Gaming on RTX Spark

The gaming rooms were a highlight of the press tour, and the company clearly wanted to make a statement: Arm-based Windows PCs can game. The flagship title was Remedy's "Alan Wake 2," running as a fully native Arm build at 2560 x 1600 resolution with DLSS 4.5 Ray Reconstruction, Multi Frame Generation, and NVIDIA Reflex, all enabled. The result was smooth, visually impressive gameplay. No issues, no lag, no stuttering. Capcom's "Pragmata" was also on display, though this one ran through Microsoft's Prism emulation layer since a native Arm port does not exist yet. Even so, performance was solid—reflections and lighting looked great, and there were no noticeable frame drops during combat sequences.

What we saw delivered a good gaming experience, but keep in mind that DLSS upscaling and Multi Frame Generation are doing a lot of the heavy lifting here—not raw rendering performance. This is not a hardcore gaming GPU. The N1X may share its CUDA core count with the desktop GeForce RTX 5070 graphics card, but the die is much smaller and operates within a 45-80 W power envelope that it shares with the CPU. Raw rasterization performance will be well below a discrete desktop card. For a thin-and-light laptop, though, what NVIDIA showed was impressive—and it is a massive step up from anything Qualcomm has offered for gaming on Windows on Arm.

The big question remains how well the broader gaming library will run on Arm. Titles with native Arm builds like Alan Wake 2 are the best-case scenario, but a large portion of the PC gaming catalog will rely on Prism emulation for the foreseeable future. The good news is that Prism handled "Pragmata" well, and NVIDIA says it is working closely with developers to bring more Arm-native ports to the platform.

One area where NVIDIA should have a significant advantage over Qualcomm is GPU drivers. Qualcomm's Adreno GPU drivers on Windows have been a persistent pain-point for Snapdragon X Elite laptops—buggy, incomplete, and lacking support for many games and professional applications. NVIDIA, by contrast, has decades of experience shipping GPU drivers on Windows with clockwork cadence, and its existing GeForce driver team and infrastructure should extend directly to RTX Spark. This alone could make or break the platform's gaming credibility, and it is encouraging that NVIDIA is treating this as a first-class effort rather than an afterthought.

DLSS 4.5 Ray Reconstruction

NVIDIA also showed the upcoming DLSS 4.5 Ray Reconstruction update, scheduled for August 2026. This is a second-generation transformer model that replaces the previous CNN-based approach, delivering 35% more compute capability and processing 20% more parameters while maintaining similar performance overhead. The result is improved lighting accuracy, better temporal stability, and cleaner motion in ray-traced and path-traced scenes.

One room was set up with side-by-side comparisons in "Alan Wake 2," "Pragmata," and "Indiana Jones and the Great Circle." The improvements were visible—CRT television screens in Alan Wake 2 now properly render individual lines of static noise, and "Pragmata" showed fewer residual artifacts with improved lighting responsiveness. DLSS 4.5 Ray Reconstruction will launch with support for 27 games, including "Cyberpunk 2077," "Hogwarts Legacy," "Star Wars Outlaws," "DOOM: The Dark Ages," "Half-Life 2 RTX," "Resident Evil Requiem," and "Call of Duty: Black Ops 7." An additional 11 titles were announced at Computex, including "Phantom Blade Zero,""Marvel Rivals," "Hell Let Loose: Vietnam," and "Gothic 1 Remake."

Creative and AI

Several rooms were dedicated to creative and AI workloads. Adobe Photoshop already runs natively on Arm64 Windows, but for RTX Spark, Adobe has added GPU-accelerated compositing at its core—live filters, HDR editing, and effects all offloaded to the Blackwell GPU, with NVIDIA claiming 2x faster AI and graphics performance. The Photoshop station had an AI agent driving the application through its existing public automation API—not natively integrated, but a clever approach that lets a local LLM control Photoshop without Adobe having to build AI features into the app itself. Starting from a hand-drawn creature sketch with texture style assignments, the agent used Stable Diffusion to generate a photorealistic 2D image in Photoshop, and then went a step further by producing a 3D scene video from the result—not an actual 3D model, but a convincing generated video with depth and camera movement, all running locally on RTX Spark. Adobe Premiere Pro is receiving similar GPU acceleration for RTX Spark, including its ability to leverage the latest NVENC and NVDEC accelerators on the N1 silicon.

DLSS 4.5 Ray Reconstruction is also coming to Blender 5.3 as a new denoiser this fall, replacing the current OptiX denoiser with much better quality. The difference is clear—fine details like window panes, ivy, and stonework that are smeared or lost with OptiX come through cleanly with DLSS Ray Reconstruction.

NVIDIA also loaded Epic's Unreal Engine 5 City Sample on an RTX Spark laptop—a notoriously memory-hungry scene that typically requires a high-end workstation to edit smoothly. With 128 GB of unified memory available, the UE5 editor navigated the massive cityscape without the out-of-memory issues or constant asset streaming that would plague a traditional laptop. It was a good example of why unified memory matters for professional creative workflows, not just AI inference.

Blackmagic Design's DaVinci Resolve was also on the show floor with RTX Video integration, and NVIDIA mentioned the platform can handle 12K video editing. ComfyUI was shown with RTX Video Super Resolution as a node for AI video workflows, along with RTX Video Frame Generation that can double or quadruple video framerate in real time. NVIDIA says the 4K upscaling runs 30x faster than popular alternatives at a fraction of the VRAM cost. Local AI agents were also part of the pitch, with NVIDIA's OpenShell secure runtime providing 2x inference performance for local models on RTX Spark.

Partner Laptops and Pricing

Eight RTX Spark laptops are confirmed for fall 2026, and NVIDIA had most of them on display. The Microsoft Surface Laptop Ultra was the main machine used throughout—a 15-inch device with a mini-LED PixelSense Ultra display and up to 128 GB of unified memory. ASUS brought the ProArt P16 and P14, both slim creator laptops designed for the platform. MSI showed the Prestige N16 Flip AI+, a 2-in-1 creator device. HP had the OmniBook X 14 and OmniBook Ultra 16, Dell brought the XPS 16 Creator Edition, and Lenovo showed the Yoga Pro 9N. Acer and GIGABYTE are expected to follow with their own models later. The Surface Laptop Ultra can be considered a de facto reference design laptop for this silicon.

RTX Spark is not limited to laptops—several partners also showed compact desktop mini PCs based on the platform. ASUS had the ProArt Mini PC with 140 W thermal design, 10 GbE networking, and PCIe Gen 5 in a 150 x 150 x 51 mm chassis, while MSI brought the EdgeMesa N AI+ and HP showed a sleek mini desktop of its own.

NVIDIA did not provide official pricing at the event, but we heard numbers floating around the show floor from various sources. The consensus points to N1-based laptops starting around $1,799, while the higher-end N1X models are expected to come in at $2,899 or above—figures that a Morgan Stanley report has since corroborated. If accurate, these prices put RTX Spark squarely in MacBook Pro territory, which makes sense given the target audience of creators and power users. Whether the broader market will accept those price points for a first-generation Windows on Arm gaming and creator platform remains to be seen. NVIDIA's play appears to be to compete with some of the higher-end variants of MacBook Air or midsegment MacBook Pros powered by M5 Pro.

NVIDIA also laid out a multi-generation roadmap for the platform, and during the press Q&A, Jensen Huang confirmed that N2X and N3X are already in development. This is not a one-off experiment—NVIDIA is clearly in this for the long haul.