RTX Spark Superchip in the Microsoft Surface Laptop Ultra uses the GB10 Superchip with 20 CPU cores, 6,144 CUDA cores, and 128GB of RAM. Its key strength is running local AI on the device, while gaming is not yet a strong point of the platform.
However, the first round of testing did not cover battery life, display quality, build quality, or detailed device-specific results. This information will be included in upcoming reviews of different RTX Spark systems.
The System Tested
The system tested was the top-end Microsoft Surface Laptop Ultra model, equipped with the RTX Spark Superchip, 20 CPU cores, and 6,144 CUDA cores.
It features 128GB of LPDDR5X-9400 RAM. Details about its appearance, build quality, and display were not disclosed in this round of testing.
When Computers Have to Work for Us
After traveling to San Francisco to receive information about the RTX Spark platform, the author brought the Microsoft Surface Laptop Ultra back for a one-week testing period.
The idea behind this platform is to turn the computer into an automated assistant for work and creative tasks. The testing therefore focused on how far the RTX Spark Superchip could take this vision of the future in real-world use.
Where RTX Spark Fits into Nvidia’s PC Strategy
Jensen Huang introduced RTX Spark at Computex 2026 with the goal of “reinventing the PC”—turning the PC into an assistant capable of handling more work and creative tasks.
RTX Spark is a Windows on Arm platform that pairs the GB10 Superchip with a version of Windows designed to run local AI agents securely in a controlled environment. It is therefore positioned not merely as a powerful chip, but as the foundation for a new kind of PC.
Nvidia is pursuing this strategy together with Microsoft, Dell, HP, MSI, Lenovo, and Asus to bring the platform to a range of future system designs.
What Can Be Compared—and What Is Still Missing
This article does not provide information about previous models or a direct old-versus-new comparison of the Surface Laptop Ultra, so it is not yet possible to summarize improvements over the previous generation. The available information covers the RTX Spark Superchip in the test system, which has 20 CPU cores, 6144 CUDA cores, and 128GB of RAM.
The important point is the new measurement methodology. Traditionally, local AI testing has often focused separately on time-to-first-token and tokens per second. AgentPerf-local measures the time from the start of a task until every step is fully completed, making it better suited to evaluating how quickly an AI agent can actually finish a job.
From Impressive Specifications to Real-World Use
RTX Spark is well suited to running local AI on the device without sending workloads to the cloud for processing. It is suitable for tasks that require an AI agent to work through multiple cycles until a mission is complete, because AgentPerf-local measures the time from the beginning to the actual completion of the task.
Using Qwen 3.8-27B provides a local model known for being highly capable among models of a similar size. Qwen 3.6 35B A3B is a less intelligent option, but it runs quickly across a wide range of systems, making it suitable for workloads that prioritize speed and continuous responsiveness.
Compared with the MacBook Pro and Ryzen AI Halo
| System | Hardware specified by the source | Acceleration system | Role in testing |
|---|---|---|---|
| Surface Laptop Ultra | RTX Spark Superchip, 20 CPU cores, 6,144 CUDA cores, 128GB LPDDR5X-9400 RAM | CUDA | Primary system tested |
| 14-inch MacBook Pro | Apple M5 Max, 18 CPU cores, 40 GPU cores, 128GB LPDDR5X RAM | Metal | Comparison system |
| Ryzen AI Halo | AMD Ryzen AI Halo compact model | ROCm stack and llama.cpp | Comparison data from AMD |
The testing used AgentPerf-local with the Qwen 3.8-27B model. llama.cpp was built from source and optimized for CUDA, Metal, or ROCm on each system.
Strengths and Limitations from the Initial Data
The RTX Spark Superchip shows strong potential for local AI workloads, particularly in a system with 128GB of LPDDR5X-9400 RAM and CUDA acceleration through llama.cpp. This makes it suitable for tasks that require AI to run on the device and handle large models.
Its limitation is that gaming is not yet a strong point of the platform. Battery life, system quality, and the Surface Laptop Ultra’s display have not been disclosed in the initial round. Meanwhile, local AI benchmarking methods are still under development, and AgentPerf-local is a new approach that requires additional feedback.
Costs the Testing Has Not Yet Answered
This round of information does not include the price, battery life, build quality, or display quality of the Microsoft Surface Laptop Ultra, nor does it provide details specific to other models. As a result, the true cost of ownership cannot yet be evaluated.
Tom’s Hardware also has other RTX Spark laptops in its testing lab and will publish detailed reviews soon. Anyone considering a purchase should wait for that information first.
Who Is It For, and Who Should Wait for More Information?
The RTX Spark platform is suitable for people interested in running local AI and agentic workflows, since the testing uses AgentPerf-local to measure the time from the start of a task until it is actually completed. It is well suited to work that requires AI to perform multiple consecutive steps.
People who prioritize gaming, or who need to know the battery life, build quality, and display quality before making a decision, should wait for more information. The initial round does not disclose these details, and gaming is not a strong point of the platform.
The Final Answer Must Wait for Full Testing
RTX Spark should be evaluated based on how long AI takes to actually complete a task, rather than by token-generation speed alone, because AgentPerf-local measures end-to-end performance in workflows involving multiple processing cycles.
However, the final answer must wait for test results from each individual system, including battery-life data, because laptop design directly affects the real-world user experience.