What RTX Spark actually is
NVIDIA's launch release describes a single superchip combining an NVIDIA Blackwell RTX GPU and an NVIDIA Grace CPU with up to 128GB of unified memory and up to 1 petaflop of FP4 AI performance. The release says MediaTek collaborated on the custom CPU design.
NVIDIA's COMPUTEX 2026 announcement filled in the top-end silicon. The GPU has 6,144 CUDA cores and fifth-generation Tensor Cores with FP4 precision, linked over NVLink-C2C to a 20-core Grace CPU. The same announcement promised laptops as slim as 14 millimeters and as light as 3 pounds in 14-inch to 16-inch sizes.
Treat the "up to" figures as ceilings. Retail coverage shows entry models sit well below them. VideoCardz reported that Microsoft's base Surface Laptop Ultra uses an 18-core CPU, 5,120-core GPU and 24GB memory.
Launch lineup and the price confusion
NVIDIA's newsroom lists these launch models:
- ASUS ProArt P14 and P16
- Dell XPS 16
- HP OmniBook Ultra and OmniBook X 14
- Lenovo Yoga Pro 9n
- Microsoft Surface Laptop Ultra
- MSI Prestige N16 Flip AI+ and Prestige N16
Pricing needs care. NVIDIA's launch release says laptops start at $2,599, and that pricing and configurations vary by region. NVIDIA's own launch-celebration blog says $2,399. One outlet reported that the lower figure is not confirmed by NVIDIA or retail listings. I can't settle that conflict from the evidence in hand, so $2,599 is the safer figure to plan around.
Retail listings point the same way. Digital Citizen reported that at Best Buy, prices start at $2,599.99 and rise to $3,999.99 depending on the processor configuration, memory, storage and display. HP briefly listed its OmniBook X 14 and Ultra 16 starting at $2,999, with higher-end configurations reaching $4,999. In short, the cheapest RTX Spark laptop is not the one with 128GB of memory.
The Windows angle: local agents and Copilot+ caveats
NVIDIA's pitch is "tool to teammate," meaning local AI agents. It says guided setup and installers cover agents including GitHub Copilot, Hermes, OpenClaw, Perplexity, Pi and WorkBuddy, without requiring a command line. For control, it points to the NVIDIA OpenShell secure runtime and Microsoft Execution Containers, which it says help users manage agent permissions.
Developers get a specific list of tooling:
- The first PyTorch-CUDA build for Windows on Arm.
- A Python build that is native by default.
- A native Conda installer.
- Support for 97% of the top 15,000 Python packages, per NVIDIA.
- Windows Subsystem for Linux and Windows Developer Config.
Windows Central's hands-on reporting adds that Microsoft worked with NVIDIA on Windows 11 scheduling, unified-memory handling, Windows ML support for TensorRT, DirectX 12 neural rendering and Prism emulation tuning. The chip is Copilot+ PC-certified. Recall, Click To Do and on-device Semantic Search were not available at launch, and Microsoft said they would arrive in coming months. Certified is not the same as shipping those features.
On privacy, running a model locally can keep prompts and files on the machine. That depends on the specific software and settings. It does not mean every agent, plugin or connected service stays on-device.
What independent testing found
Windows Central tested the top 6,144-core configuration in a Surface Laptop Ultra for a week. The results are mixed in an interesting way.
- CPU: Broadly comparable to Apple's M4 Pro and Qualcomm's Snapdragon X2 Elite, according to the reviewer. Windows Central's earlier headline summed it up as a strong GPU and an "okay" CPU.
- GPU: In some workflows it landed near an RTX 4070 or RTX 5060. That is the reviewer's comparison from particular tests. It is not a universal equivalence, and results depend on chassis, cooling and power limits.
- Video export: In Premiere Pro Beta, the Spark Surface took 42.002 seconds for an H.264 export and 81.419 seconds for H.264 UHD. The RTX 5060 laptop tested took 32.111 and 64.140 seconds. A discrete GPU still wins some creator tasks.
- Local AI: Procyon time-to-first-token results were 0.27 seconds for Phi-3.5-mini, 0.47 for Mistral 7B, 0.40 for Llama 3.1 8B and 0.84 for Llama 2 13B. Sustained decoding lagged higher-wattage discrete GPUs.
- Image generation: In a Stable Diffusion 1.5 INT8 test, Spark scored 11,294 and generated an image in 2.77 seconds. It beat the RTX 5060, 5070 Ti and 5080 laptops in that test.
The takeaway is that Spark's advantage is memory capacity and efficiency for AI workloads. It does not win every benchmark. Large models that overflow a typical discrete laptop GPU's VRAM are where unified memory should matter most. That is my analysis, not a measured result.
Arm compatibility is still the tax
RTX Spark is an Arm64 platform. x86 games and apps rely on Prism emulation. Windows Central said most tested games ran, but not all. Red Dead Redemption 2 crashed with the default Vulkan setting and worked after switching to DirectX 12.
Agent setup was also uneven. The reviewer got Hermes Agent running locally, where it organized files and built a basic webpage. OpenClaw's gateway would not connect to the chat interface, so that agent never worked.
Anti-cheat is the other gap. NVIDIA says launch support covers titles protected by various anti-cheat systems. It is still working with vendors to expand support for BattleEye, EA Javelin, Easy Anti-Cheat, RICOCHET, Vanguard and Tencent Anti-Cheat Expert. If your main games use those, check before buying.
Who should buy, and who should wait
A good fit:
- Developers and AI tinkerers who want large local models on a portable Windows machine.
- Creators using apps with native RTX Spark support. Photoshop and Premiere have beta native support at launch, with Substance coming soon.
- Buyers who value Arm efficiency and can live with emulation.
Think twice:
- Anyone whose workflow depends on kernel-level anti-cheat, legacy x86 drivers or niche x86 tools.
- Buyers who just want raw creative speed per dollar. A conventional RTX laptop can be faster in some exports and cheaper.
- Anyone counting on Copilot+ features like Recall or Semantic Search today.
Battery life and thermals under load were outside the scope of Windows Central's benchmark article, so these remain open questions until more reviews land.
Bottom line
RTX Spark is a real new hardware tier for Windows, with credible local-AI strengths and a polished developer story. It is also expensive, Arm-based and still maturing in software. The $2,399 versus $2,599 discrepancy is a reminder to check the actual listing for your configuration. Wait for more reviews if battery life, sustained thermals or anti-cheat support decide your purchase.
Update: RTX Spark runs all 22 games in a broader compatibility test (October 23, 2026)
RTX Spark reportedly ran all 22 games in a new back-catalog test published by XDA Developers on October 23, spanning releases from 2006 through 2025. According to XDA, the selection covered older 32-bit titles, multiple graphics APIs and demanding modern games, offering broader compatibility evidence than the earlier launch testing—but not proof that every Windows game works.
The anti-cheat findings are particularly relevant for buyers: XDA reports that Elden Ring worked online with Easy Anti-Cheat, and that Destiny 2, Fortnite, CS2 and Marvel Rivals also worked. That narrows the earlier compatibility caveat for those specific games; it does not establish blanket support for their anti-cheat systems. According to XDA, support for Call of Duty, Valorant, Delta Force and Battlefield is planned for 2027, with timing still subject to change.
Older games needed some troubleshooting. XDA reports that Fallout 3 and Far Cry 3 displayed .NET 3.5 errors, resolved after downloading the component from Microsoft. Other issues involved launcher sign-ins and games selecting lower display resolutions.
Performance conclusions remain limited: according to XDA, only Cyberpunk 2077 and Shadow of the Tomb Raider received benchmark testing, while individual games were generally tried for roughly half an hour. For buyers, this is encouraging evidence for back-catalog compatibility—not a substitute for checking essential multiplayer titles or sustained performance.
Update: Testing finds RTX Spark’s little-publicized NPU is real, but software support remains limited (October 24, 2026)
According to XDA Developers’ October 24 investigation, RTX Spark laptops include a separate NVIDIA NPU/DLA accelerator rather than using the GPU to satisfy Copilot+ PC requirements. XDA reports the component has its own clock, power rail and compute hardware, and ran neural-network workloads while the GPU remained idle.
The outlet found the NPU can execute models through NVIDIA’s currently beta Windows-on-Arm TensorRT and cuDLA tooling, but not through normal Windows AI paths. DirectML reportedly cannot access it, and a usable Windows ML execution provider is not yet publicly available. That means developers cannot simply target the NPU with typical Windows ML tooling today.
XDA also reports that the NPU delivered roughly three to three-and-a-half times the tested CPU performance in modified MobileNetV2 and ResNet-50 runs. But its current hybrid submission model keeps part of the GPU active, weakening whole-system efficiency, while the outlet could not reproduce the 50–55 TOPS figures associated with the hardware.
For buyers, this reinforces that RTX Spark’s near-term AI value remains centered on its GPU and large unified-memory pool. The dedicated NPU may matter once Copilot+ features and broader framework support arrive, but it is not yet a generally accessible accelerator for everyday Windows AI software.
Update: RTX Spark beats Strix Halo in XDA local-AI comparison, but 64GB limits large models (October 25, 2026)
According to XDA Developers’ October 25 comparison, a 64GB Lenovo Yoga 9n with RTX Spark N1X outperformed a 128GB Asus ROG Flow Z13 using AMD’s Ryzen AI Max+ 395 across every local-AI test that fit on both machines. XDA measured document-processing gains of 2.1× to 3.5×, token-generation gains of 1.4× to 1.6×, and a 2.6× faster FLUX.2 image-generation result.
The outlet also found Spark maintained its tested generation speed on battery, while the Z13 slowed when unplugged. In one wall-power measurement, XDA reported Spark used roughly half the energy per completed summary despite drawing slightly more instantaneous power, because it finished sooner.
The comparison adds an important capacity caveat. XDA says the 64GB Spark system exposed about 46GB to CUDA and could not load a 120B model that ran on the Z13’s 128GB configuration. The AMD machine completed that workload at a reported 31.7 tokens per second. Buyers choosing Spark for large local models should therefore prioritize 128GB configurations rather than assuming the faster GPU compensates for less available memory.
XDA attributes part of Spark’s advantage to CUDA software maturity rather than its modest paper bandwidth lead. That is a meaningful advantage for supported tools today, but the test also used Windows-on-Arm beta or nightly software on the NVIDIA side, so it does not remove the earlier compatibility and application-support caveats.
Update: XDA testing qualifies NVIDIA’s RTX 5070-class gaming claim (October 26, 2026)
According to XDA Developers’ October 26 testing, a 64GB RTX Spark N1X laptop delivered roughly 51% to 62% of a desktop RTX 5070’s frame rate in four of five GPU-heavy game tests at 1440p maximum settings with ray tracing, while Horizon Zero Dawn Remastered reached only 26%. The comparison is not a direct laptop-versus-laptop match, but it shows the large effect of the desktop card’s far higher power budget and memory bandwidth.
XDA says Spark was closer to RTX 5070 laptop results in Cyberpunk 2077, reaching roughly 80% to 84% of previously tested mobile-5070 performance in a comparable ray-tracing scenario. However, CPU-limited 1080p testing varied widely, with Spark delivering 19% to 53% of the desktop system’s results depending on the game engine. That reinforces the earlier Arm and Prism caveat: lower graphics settings will not necessarily fix a CPU- or translation-bound game.
NVIDIA’s advertised 100-plus fps 1440p ray-traced result was achievable in XDA’s Cyberpunk test only after enabling 4x frame generation. XDA reports this produced 110 fps, but with 35–40ms latency and fewer underlying rendered frames than the desktop RTX 5070. For buyers, Spark appears credible for GPU-bound single-player games using DLSS and frame generation, but its RTX 5070 comparison should not be read as equivalent native performance—particularly in CPU-heavy titles.