NVIDIA RTX Spark laptop benchmarks vary by workload

The first NVIDIA RTX Spark N1X laptop benchmarks show a workload-specific picture: HotHardware found strong results in selected local-AI and graphics tests, but mixed CPU and gaming outcomes. Tom’s Hardware’s tests recorded shorter completion times for the tested Surface Laptop Ultra than two comparison systems on the Qwen workloads, alongside substantial latency in its two game tests.

For earlier coverage of the Lenovo systems, see NeoTeo’s report on the Yoga 9n RTX Spark laptops.

What the tests covered

RTX Spark N1X is a laptop platform, not a single laptop model. HotHardware tested a top-end Microsoft Surface Laptop Ultra with a 20-core N1X, 6,144 Blackwell GPU cores and 128GB of unified memory. Tom’s Hardware tested a 15-inch Surface Laptop Ultra with 128GB of LPDDR5X-9400 memory and 1TB of storage. Results on these configurations describe those tests, not every N1X laptop.

CPU, AI and graphics results split by workload

HotHardware’s Speedometer 3.1 browser test placed the N1X system in the bottom third of its comparison; PCMark 10 Applications ranked it near the bottom. Cinebench 2026 told a different story: single-threaded performance trailed current competitors, while the 20-core system’s multi-threaded result ranked near the top. Geekbench also produced competitive single-thread results and among the strongest multi-thread results. Each test measures a different kind of work, so there is no single CPU result that captures them all.

The AI benchmarks were a stronger showing. HotHardware reported that the system led its tested field in UL Procyon AI Text Generation and Computer Vision 2.0; in Computer Vision, HotHardware said the second-place Snapdragon X2 Elite Extreme finished 30% behind. The N1X also led HotHardware’s MLPerf Client tests for Flux 2 Klein 4B image generation and Qwen 8B agentic tasks. In graphics, it led the review’s 3DMark Speedway and Steel Nomad Lite tests.

Local AI: times for two Qwen workloads

Tom’s Hardware reported these AgentPerf-local completion times for its three tested systems. The comparisons are specific to the named models and test methods:

For the Qwen 3.8-27B run, the Surface and MacBook used DFlash2 speculative decoding, while Ryzen AI Halo used the model’s built-in MTP. The second workload used MTP on all three systems. Those differences matter when comparing elapsed times; neither result predicts performance on every model or AI task.

Gaming depends on the title and software path

HotHardware’s 1080p tests put the N1X ahead of AMD’s Radeon 8060S in Cyberpunk 2077 with Ray Tracing Low, but behind it in The Talos Principle II at Medium. The latter ran through Microsoft’s PRISM emulator, rather than natively on Arm.

Tom’s Hardware measured 77.7 average FPS, a 39.9 FPS 1% low and 83.5ms average PC latency in Gears of War: E-Day at 1908×1272, High, with Mega Geometry and MFG 2X enabled. In Control Resonant at 2560×1440, High, ray tracing off and MFG 3X, it recorded 109.2 average FPS, a 51.6 FPS 1% low and 78.3ms average PC latency. The reviewer reported a major hang during a Gears map transition and said ray tracing reduced frame rates and increased latency. Tom’s Hardware recommends a dedicated gaming laptop for players who prioritize maximum performance and low latency. HotHardware also noted in its October 15 review that some anti-cheat systems, including EA’s, were not yet compatible with Arm.

Battery life and power measure different things

HotHardware measured 14 hours and 35 minutes in its battery-life test on a 128GB, 20-core Surface Laptop Ultra with a 3K OLED display. On battery, Speedometer performance fell about 18%, while multi-threaded Cinebench performance remained virtually unchanged. Those are separate results from one test system, not a single battery-performance score.

HotHardware also measured a peak of about 138W for the entire system, with display brightness turned way down. The review gives the N1X SoC a 45–80W TDP range; the 138W peak is a whole-laptop measurement, not the chip’s power draw.