My First 15 Minutes with RTX Spark

Summary

Linus gives an initial hands-on look at NVIDIA’s RTX Spark/N1X platform, focusing on its local AI capabilities, Windows-on-ARM gaming potential, and software features intended to distinguish it from competing chips. This is a preview rather than a full performance review; Linus says the review and comparisons are due the following day.

Hardware and positioning

• NVIDIA presents the chip as an efficient ARM-based Windows platform combining CPU and GPU performance, with configurations reaching up to 20 CPU cores and 128 GB of LPDDR5X memory.
• The entry-level configuration is described as having 18 CPU cores, 5,120 CUDA cores, and up to 32 GB of RAM. Linus considers it adequate for gaming and moderate creative work, but less suited to running large local AI models.
• Linus highlights the MSI Prestige 16 Flip AI Plus as an example of a thin, feature-rich laptop design. NVIDIA claims the system can run at full speed on battery, though Linus notes that a demanding Minecraft demo made the laptop’s cooling audible.
• NVIDIA frames the high-memory model as potentially worthwhile for users who could replace recurring cloud AI costs with local workloads. Linus notes that the hardware is expensive and that AMD’s higher-memory alternatives have been available longer.

Local AI setup and security

Linus demonstrates setting up the Hermes AI agent and switching it from a cloud backend to a local model.

In an unsupervised test, he asked Hermes to build a monitoring dashboard showing silicon utilization, thermals, token throughput, tool calls, and memory usage, and to install NVIDIA App and Broadcast plugins. The process took just over an hour and produced a working dashboard without Linus writing code.

Linus emphasizes that AI agents raise security concerns, from unrestricted system access to constant permission prompts. He describes Microsoft’s operating-system sandboxing and containerization as ways to manage agent permissions through a graphical interface. An upgraded Task Manager is also intended to show an agent’s user ID and its use of system resources.

NVIDIA’s Nemo Switchyard and Microsoft’s Hydra Fusion are described as tools for routing tasks: simpler requests can run on local models, while more demanding ones can be sent to cloud models.

Gaming and creative work

• NVIDIA compares the platform’s gaming performance to an RTX 5070 laptop GPU. A demo shows NBA 2K27 at 1440p with DLSS 5. NVIDIA says the DLSS model is twice as fast and that Frame Generation is gaining a faster mode, with a modest image-quality trade-off.
• Marvel Rivals is shown running with anti-cheat support, including a native ARM build. Linus sees this as evidence of progress for Windows on ARM, but says the demos do not resolve all compatibility concerns: NVIDIA’s materials refer to “hundreds” of games, and the breadth of support remains unclear.
• Other demonstrations include ray-traced lighting in Minecraft and Blender’s DLSS Ray Reconstruction, which Linus says can make viewport manipulation feel close to real time.
• In Photoshop, users can select a local AI model for features such as object removal or create adjustable image filters using plain-language instructions.
• NVIDIA Broadcast adds frame generation and super resolution for webcams, and can apply effects such as audio cleanup to prerecorded content.
• ComfyUI gains RTX video frame generation and HDR features, which Linus says he wants to test further.

Other AI demonstrations

NVIDIA Pair can pool AI-capable computers on a home network and distribute workloads among them—for example, by borrowing compute from another system.

A Home Assistant demonstration uses a local agent to interpret a request and control smart-home devices without requiring the user to specify each device or setting. The broader example involved giving the agent household information and manuals so it could monitor water use, identify a possible sprinkler problem, and draft an email to a landscaping company. Linus notes that setting up this system appears to require substantial work.

Guides, reviews, and demonstrations

• Hands-on setup: Hermes onboarding and switching to a local model.
• Practical agent test: Building a custom system-monitoring dashboard through a natural-language prompt.
• Demos reviewed: Gaming and anti-cheat, Blender, Photoshop, ComfyUI, NVIDIA Broadcast, Pair, and Home Assistant.
• Full performance review: Promised for the next day, including comparisons with other integrated-chip platforms and a conventional laptop with a discrete CPU and GPU.

Main speaker/source: Linus Sebastian of Linus Tech Tips. NVIDIA’s presentations and product demonstrations provide many of the platform claims; a brief demo presenter also speaks during the Home Assistant segment.