Tom’s Hardware Finds an AI Lead for Surface Laptop Ultra, With Gaming Limits

Tom’s Hardware’s October 15, 2026, report found that the tested Surface Laptop Ultra with NVIDIA RTX Spark finished two AgentPerf-local AI tasks faster than a 14-inch MacBook Pro with M5 Max and AMD Ryzen AI Halo. Its game tests were less decisive: Gears of War: E-Day had a major map-transition hang, and both games recorded average PC latency above 78 ms. The measurements belong to Jeffrey Kampman and Tom’s Hardware, and apply to the specific hardware, models and settings in Kampman’s report.

Two local-AI tasks, three systems

The Surface had the shortest completion time in both AgentPerf-local runs. The benchmark measures end-to-end task completion using AI inference traces; a shorter time means the system finished that particular task sooner.

These are two model-specific results, not a general ranking of AI systems. The decoding paths differed in the Qwen 3.8 27B run, and the second workload used Qwen 3.6 35B A3B, which Tom’s Hardware describes as a lighter, less capable model that is easier for more systems to run quickly. The report says each system ran llama.cpp built from source.

The comparison also spans different machines: the tested Surface was a 15-inch laptop with a 140 W adapter; the MacBook Pro was a 14-inch model with a 96 W adapter. The third system, Ryzen AI Halo, is a compact desktop, not a laptop. Tom’s Hardware did not include a 16-inch MacBook Pro with M5 Max in its comparison.

Gaming results depend on the title and settings

In Gears of War: E-Day, Tom’s Hardware recorded an average of 77.7 frames per second, 39.9 FPS for 1% lows and 83.5 ms average PC latency at 1908×1272, High settings, Mega Geometry on and 2X frame generation. Kampman also reported a major hang during a map transition.

For Control Resonant, the report gives 109.2 average FPS, 51.6 FPS for 1% lows and 78.3 ms average PC latency at 2560×1440, High settings, ray tracing off and MFG 3X. Kampman says enabling ray tracing reduced frame rates and raised latency.

Kampman points to Prism, which emulates x86 apps on Windows on Arm, the Grace CPU and shared memory bandwidth as constraints in the tested games. The report puts the platform’s shared bandwidth at up to 300 GB/s, compared with 384 GB/s of dedicated bandwidth for the cited RTX 5060 Mobile. These findings describe the tested games and configurations; they do not predict performance in every game.

The report also says Microsoft and NVIDIA’s promised Windows 11 features for secure, manageable local agents were not ready for Kampman and Andrew Freedman to test when the article was published.

US pricing and what the report covers

Microsoft’s US product page lists the Surface Laptop Ultra starting at $2,599.99. Tom’s Hardware separately reports a $5,899 list price for its tested system, configured with 128 GB of memory and 1 TB of storage. The starting price and the tested configuration are distinct price points.

Tom’s Hardware’s article is a platform-focused test, not a full system review: it does not include its own battery-life, build-quality or display-quality results. Microsoft advertises up to 15 hours of local video playback and up to 12 hours of active web browsing, based on September 2026 testing with preproduction software and a preproduction 6,144-core unit with 64 GB of memory and 1 TB of storage. Microsoft’s video test used 1080p playback at 24 frames per second, a 150-nit screen and Wi-Fi; its browsing test also used a 150-nit screen and Wi-Fi. These are manufacturer test claims, not measurements from Tom’s Hardware. Microsoft’s US product page lists the starting price and battery claims.