Tests of RTX Spark, conducted on a Surface Laptop Ultra before the first devices based on it began shipping, reveal an unusual Windows on Arm platform: graphics performance close to the RTX 4070 class, strong capabilities for running AI models locally, and a good but not class-leading CPU. NVIDIA is targeting this platform at developers, content creators, and users who run AI models on their devices instead of sending their data to the cloud.
RTX Spark is based on the N1X chip designed by NVIDIA in cooperation with MediaTek. It is available in two configurations with either 18 or 20 CPU cores, alongside either 5,120 or 6,144 graphics cores based on the Blackwell RTX architecture. Unified memory reaches 128 GB, with bandwidth of 300 GB per second, while configurable power consumption ranges from 45 to 140 watts depending on the configuration and device design.
The CPU Is Good, but It Is Not the Main Strength
The 20-core N1X configuration scored 558 points in the single-core test and 6,885 points in the multi-core test in Cinebench 2026. In Geekbench 7, it scored 2,718 points in the single-core test and 23,571 points in the multi-core test. These results place the chip in the range of the Apple M4 Pro and Snapdragon X2 Elite in some workloads, but leave it behind the Apple M4 Max and Snapdragon X2 Elite Extreme, particularly in single-core performance.
In practice, the test described everyday performance as smooth, including applications running through the Prism emulation layer. The published tests also showed no significant decline between battery operation and operation from a power source, with Best Performance mode enabled.
The GPU Is the Main Strength
In graphics tests, the 6,144-core configuration approached the performance of an RTX 4070. It scored 12,771 points in 3DMark Time Spy at 1440p, with the ability to run games such as Cyberpunk 2077, Shadow of the Tomb Raider, and The Witcher 3 Remastered at high frame rates at this resolution. The platform supports technologies such as ray tracing, DLSS 5, Multi Frame Generation, Reflex, and G-SYNC.
However, some games still require adjustments because they rely on x86 emulation on the Arm64 platform. For example, Red Dead Redemption 2 crashed when Vulkan was used in the test, but worked after switching to DirectX 12. Video export performance also did not surpass that of modern discrete cards: exporting H.264 at Full HD resolution took 42.002 seconds, compared with 32.111 seconds on a system equipped with an RTX 5060.
Local AI Is the Clearest Use Case
RTX Spark stood out more clearly in quantized AI workloads. In the Stable Diffusion 1.5 test at INT8 precision, it scored 11,294 points and generated the image in 2.77 seconds, outperforming systems equipped with an RTX 5060, RTX 5070 Ti, and RTX 5080 in the same test. It also showed low time-to-first-token figures in language models such as Phi and Llama 3.1 8B, although sustained decoding rates remained behind higher-power discrete cards.
Unified memory enables large models to run locally. The tests used models such as Qwen 3.8 27B and Gemma 4 31B, while noting that models of up to 120 billion parameters can be run depending on memory allocation. The practical benefit is keeping documents, audio files, and intellectual property on the device during summarization or automation instead of sending them to a cloud service.
What Changes for Windows Users?
Microsoft worked with NVIDIA to improve Windows 11 scheduling, unified memory management, and the Prism layer. TensorRT support was also added to Windows ML, along with neural rendering support in DirectX 12. Microsoft plans to make Copilot+ PC features such as Windows Recall, Click To Do, and Semantic Search available later, although RTX Spark is classified as a Copilot+ PC and does not provide these features at launch.
Ease of use remains an important limitation. The Hermes Agent test successfully ran local tasks such as organizing the downloads folder and building a webpage, while OpenClaw could not run because of a gateway connection failure. The test suggests that local agent tools may confuse beginners, even with Windows’ efforts to simplify setup.
Editorial conclusion: RTX Spark is not a universal replacement for every laptop. It is a specialized platform that gives thin devices substantial memory and powerful acceleration for local AI, along with acceptable gaming and content-creation performance. However, the CPU is mid-range, power efficiency remains tied to the manufacturer’s decisions, and Intel- or AMD-based systems with discrete NVIDIA cards may cost less. Its value is therefore clearest for those who make running models locally, software development, or data processing the focus of their daily work.