NVIDIA RTX Spark PCs Arrive in US Market; The New Stack Reviews Surface Laptop Ultra's Local LLM Performance

Key summary

NVIDIA has launched its RTX Spark PC lineup in the United States, marking a new entry into the consumer PC market following its earlier Tegra and Surface RT era. Several media outlets have begun reviewing these devices from various manufacturers, including Microsoft's own Surface Laptop Ultra and hardware from other partners.

The New Stack conducted a specific test of the Surface Laptop Ultra, focusing on running Large Language Models (LLMs) locally, which is a key selling point of the RTX Spark platform. The review utilized a unit with 64GB of RAM provided by Microsoft, noting that this is not the highest specification available, as a 128GB version exists.

Performance results for various models were recorded while the device was plugged in:

The review highlighted that unplugging the device significantly degrades performance. For instance, Qwen 3.8 in 4-bit quantization dropped to 18 tokens per second, and Nemotron 3.5 Bolt fell to 42-45 tokens per second, representing an overall reduction of approximately 20%.

Thermal management was also a point of discussion; the Surface Laptop Ultra runs hot with fans spinning at full speed and generating noticeable noise when running LLMs, whereas it remains quiet during general use. Consequently, The New Stack suggests that for users seeking maximum RTX performance for model running, the Surface RTX Spark Dev Box desktop might be a better alternative due to its superior cooling design.

Finally, the article addressed RAM limitations, noting that few models are runnable with only 64GB of RAM, as most available models range between 27B and 30B parameters. While users can opt for the 128GB RAM configuration, it comes at a premium price tag of $5,900 (approximately 200,000 Thai Baht).

Key points

• Nemotron 3.5 Bolt (9B): A new model from NVIDIA released to welcome RTX Spark for local agent tasks, achieving 56-60 tokens per second.
• Qwen 3.8 (27B): Achieved 8-9 tokens per second with 8-bit quantization and 25-28 tokens per second with 4-bit quantization.
• Gemma 4 26B-A4B: Achieved 40 tokens per second with 8-bit quantization. The New Stack attributes this speed to the model's mixture-of-experts (MoE) architecture.