NVIDIA Tried to Tell me How to Benchmark RTX Spark

Summary

The video reviews NVIDIA’s RTX Spark laptops, powered by the N1X system-on-chip, and questions whether they justify their high prices. NVIDIA reportedly barred independent benchmarks on launch day and warned about misleading tests, particularly for local AI. The reviewers agree that benchmarks should reflect real workloads, but argue that independent testing is still essential.

Hardware and Comparisons

The N1X combines an ARM CPU, NVIDIA graphics, and up to 128 GB of unified LPDDR5X memory. The tested systems were the Microsoft Surface Laptop Ultra and ASUS ProArt P16. The reviewers say the CPU cores are based on older ARM designs and suggest the chip reuses much of the design used in NVIDIA’s DGX Spark, adapted for laptop power limits.

The laptops were compared with systems from Qualcomm, Apple, AMD, and Intel, including Ryzen AI Max devices.

Gaming

The team tested Cyberpunk 2077, Grid Legends, Shadow of the Tomb Raider, and Total War: Pharaoh. These x86 games ran on ARM systems through Microsoft Prism.

• The N1X’s integrated graphics beat Apple’s and Qualcomm’s integrated offerings in some tests, but results varied substantially by game.
• Cyberpunk 2077 ran at 1200p with medium settings, averaging in the mid-90 FPS range, with lows around 80 FPS.
• The N1X was competitive with Apple’s M5 in Shadow of the Tomb Raider, but weaker in Grid Legends. It averaged under 60 FPS in Total War: Pharaoh at ultra settings.
• The tests did not include DLSS. The reviewers say NVIDIA’s higher frame-rate claims rely on DLSS, but the relevant update was unavailable during testing.

Productivity and Creator Workloads

Native Windows ARM versions were used where available, reducing concerns about emulation overhead.

• Cinebench showed weak multi-core performance and single-core results roughly comparable to an Intel system, but behind newer Snapdragon and Apple chips.
• 7-Zip results were fairly close across devices, with Ryzen AI Max leading.
• CPU rendering in Blender was slow, while GPU rendering with OptiX was more in line with expectations.
• In compilation and other tests, the N1X was generally average or behind Qualcomm and Apple.
• PugetBench creator results were around average. NVIDIA-specific Adobe features and future GPU-optimized software may benefit some users, but the reviewers did not find a clear creator-focused recommendation.

Local AI

Tests used Gemma 4 E4BQ8 across all systems and Qwen 3.8 27B Q4KM on devices with enough memory.

In Llama.cpp’s LlamaBench, the N1X was especially strong at prompt processing for Gemma—roughly 2,000 more tokens per second than Ryzen AI Max—thanks to CUDA. It did not approach a discrete GPU, and its advantage disappeared during text generation, where it landed around the middle of the group. The team cautions that inference speed varies with the model, quantization, and software engine.

NVIDIA provides setup resources, and Windows includes an initial-setup wizard for tools such as Hermes. The wizard reportedly does not appear if Windows is configured with an offline account.

Battery Life and Thermals

In a 1080p video playback test, the Surface lasted about nine hours, roughly matching the MacBook Pro 14, while the ProArt lasted around twelve hours. Neither beat the ASUS Zenbook’s nearly thirteen hours, despite the Spark laptops having batteries 30–40% larger. The ProArt’s performance also varied on battery, contrary to NVIDIA’s claims of consistent plugged-in and unplugged performance.

Thermal testing was incomplete. The Surface hovered around 95°C without throttling, but the team could not get reliable results from the ProArt. They also lacked the sensor support needed for full monitoring.

Verdict and Resources

The reviewers find the N1X laptops difficult to recommend broadly. They offer capable graphics and substantial memory, but performance is inconsistent across workloads, battery efficiency is not class-leading, and prices are extremely high: nearly $6,000 for the tested Surface configuration and about $1,000 more for the 128 GB ProArt.

Their clearest use case is running AI locally when internet access is unreliable. Otherwise, they see few compelling advantages over existing alternatives.

Review Resources

• The full LTT Labs write-up provides more detailed test results.
• A previous video covered the devices at NVIDIA’s event and compares the launch promises with the review findings.
• NVIDIA’s setup guidance and Windows’ AI setup wizard are noted as practical resources for users configuring local AI.

Main Speakers and Sources

Linus presented the review, with testing support from the LTT Labs team. Other sources and tools mentioned include NVIDIA’s benchmark guidance and product claims, Microsoft Prism and Windows setup tools, and Proton, the video’s sponsor.