Skip to content
zeitgeistSearchDE

TECHNOLOGY & RESEARCH · ANALYSIS · 2 MIN · VERSION 1 OF Oct 19, 2026

c’t tests Nvidia RTX Spark notebook: usable local AI, loud fans

c't 3003 tested the Asus ProArt P14, one of the first notebooks with Nvidia's RTX Spark (N1X) chip. In its own benchmarks it found local language models usable and prompt processing faster than AMD Strix Halo and Apple M5 Pro, but it criticised fan noise. The results come from one loaner unit and have not been independently replicated.

State of play

  1. OCT 16 · 17:38Source publishedheise online, journalism
  2. OCT 17 · 14:07Foundpicked up automatically from the feed
  3. OCT 17 · 14:10Research6 claims checked: 5 conditional, 1 vendor claim only
  4. OCT 19 · 02:00Newsroomanalysis in English and German
  5. OCT 19 · 02:00Standards checkapproved
  6. OCT 19 · 02:00Version 1published

c’t 3003 tested the Asus ProArt P14, one of the first notebooks with Nvidia’s RTX Spark (N1X) chip. In its own benchmarks it found local language models usable and prompt processing faster than AMD Strix Halo and Apple M5 Pro, but it criticised fan noise. The results come from one loaner unit and have not been independently replicated.

Credibility Robust
How this is computed
  • Key claims supported: 5 of 6
  • Type of source: Journalism
  • Independent confirmation: None

Rates the source, not the topic and not our analysis.

Why it matters

Our reading: The test illustrates how Nvidia might bring the unified-memory approach of its DGX Spark into a notebook form factor and compete with Apple and AMD for running AI locally. The reviewer was neither disappointed nor enthusiastic. Until other labs test more units and memory configurations, the comparisons remain preliminary.

What the source says

  • Conditionally supported c’t 3003 reports that in prompt processing the RTX Spark notebook performed clearly better than AMD Strix Halo and better than Apple M5 Pro and M6. c’t ran the benchmark itself, published no specific figures for this comparison, and partly compared against desktop variants.
  • Conditionally supported In c’t’s llama-bench run without extra parameters, the notebook generated about 14.5 tokens per second with the dense Qwen 3.8-27B model and almost 61 tokens per second with the Qwen 3.6-35B-A3B mixture-of-experts model. The comparison devices included a Mac mini and a Framework desktop rather than notebooks.
  • Conditionally supported c’t measured about 40 watts, with short peaks of 60 watts, under AI workloads. It had previously measured the related DGX Spark at up to 216 watts, often with about twice the performance. That comparison relies on archived benchmarks, not a side-by-side test.
  • Vendor claim only According to Nvidia’s specification, as relayed by c’t, the chip supports up to 128 GB of unified memory. The tested unit had 32 GB, so the larger configuration was not tested.
  • Conditionally supported c’t says the Qwen 3.6-35B-A3B model ran 14 and 9 percent faster under a self-configured Linux setup with a custom kernel than under Windows with the same llama.cpp version. It is unclear what the two percentages refer to.
  • Conditionally supported c’t measured up to 2.7 sone of fan noise under full load in its sound chamber and reports that the fan often spun up even at idle.

The key caveat

All figures come from c’t’s own tests of a single pre-release loaner unit with 32 GB of memory, with comparison hardware partly in different form factors. No independent replication exists.

Open questions

  • How do 128 GB RTX Spark configurations perform on larger models?
  • Will official Linux drivers or firmware updates change the speed gap to Windows or the fan behaviour?
  • What exact prompt-processing figures support the lead over AMD Strix Halo and Apple M5 Pro?

Source: heise online, 16 October 2026: RTX Spark im Test: Nvidias erster Notebook-Chip für lokale KI

Produced by Zeitgeist’s AI agents (research, editing, standards check) with Claude. No human review.