Nvidia RTX Spark (N1X) tested: how fast do local LLMs run on the notebook chip?

ComputerBase published the first RTX Spark review on 15 October 2026, and c't (heise) followed a day later with local-AI measurements. The chip is called N1X internally and is Nvidia's first SoC for notebooks. CPU and GPU share up to 128 GB of fast memory, and anyone who runs language models locally asks about that number first.

All numbers come from the reviews by ComputerBase (15 Oct 2026) and c't/heise (16 Oct 2026). We do not own a unit and measured nothing ourselves. (Quelle: Simple Icons, file nvidia.svg (CC0))

What is inside

According to ComputerBase, the N1X is based on the GB10 from the DGX Spark, drops the ConnectX-7 networking and offers memory options that include 24, 32 and 48 GB. There are two variants:

	18 cores	20 cores
CPU	9× Cortex-X925 + 9× A725	10× X925 + 10× A725
GPU	5,120 shaders (40 SM)	6,144 shaders (48 SM)
Mini PC	no	yes, up to 140 W

Both have a Blackwell GPU, LPDDR5X-9400 on a 256-bit bus with up to 300 GB/s, and up to 128 GB of memory, of which up to 96 GB can be allocated to graphics. Notebook power is 45 to 80 W. The 55 TOPS NPU is reserved for Copilot+ features, says ComputerBase; third-party software cannot use it. c't adds that the NPU is irrelevant for local LLMs anyway, because the models run on the GPU cores. The system runs Windows 11 on Arm, and there is no x86 CPU.

c't measurements

c't tests the Asus ProArt P14 with 32 GB of shared memory. With llama-bench and no extra parameters, the dense Qwen 3.8-27B (as the article writes it) reaches 14.5 tokens per second and the MoE model Qwen 3.6-35B-A3B almost 61.

Both values without multi-token prediction. c't says MTP should add at least 25 percent, more for coding. The article gives the MoE figure as almost 61. (Quelle: c't / heise, 16 Oct 2026)

How c't ranks it: the dense model runs faster than on Apple's M6 and AMD's Strix Halo, but slower than on the M5 Pro. The Intel Arc Pro B70 (32 GB), an RTX 4090 and a Mac Studio with M5 Ultra are clearly faster. For prompt processing, the RTX Spark notebook beats Strix Halo, M5 Pro and M6. A DGX Spark often manages double but draws up to 216 W, while the notebook in the test took about 40 W sustained and 60 W in peaks. The Strix Halo figure comes from a Framework Desktop under a 120 W sustained load, which the author himself calls unfair. The 27B model needs 14 GB as an IQ4_XS quantization, and realistically under 20 GB with context.

A word on our own setup: four RTX 5090 (4 × 32 GB) run a model of the same Qwen family under vLLM. The numbers cannot be laid side by side. There a different model is spread over four cards, here CPU and GPU share one memory of up to 300 GB/s. We have no measurement of our own for the RTX Spark.

Linux, prices and daily use

Stock Ubuntu for Arm starts, c't reports, but at first without Nvidia, touchpad and keyboard drivers. After manual work, including a freshly compiled kernel, all three worked. According to c't, mainly the MoE model ran faster on Linux, by 14 and 9 percent respectively; the text does not make clear which value belongs to which model.

The two reviews differ on prices because the test units differ. ComputerBase gives 2,899 euros as the entry price for 24 GB and a 512 GB SSD, the P14 with 32 GB costs 3,299 to 3,300 euros, devices with 64 GB start at 4,300 euros and with 128 GB at 6,700 euros. ComputerBase's ProArt P16 with 64 GB and 1 TB is listed at 4,299 euros. c't prices the MacBook Pro with M5 Pro and 48 GB at 3,550 euros, but elsewhere in the same review says "roughly 3,800 euros", so those figures contradict each other.

As a value alternative, c't names older AMD Strix Halo notebooks that cost under 2,000 euros with 32 GB. Battery life is over 19 hours of video playback at 200 nits, and under full load the machine gets as loud as 2.7 sone.

Who should look at it

For a notebook that should run 27B models on the side, the memory size fits. As a permanent inference server, 14.5 tokens per second is slow, and cards with their own memory suit that job better. The MoE result of almost 61 tokens per second shows where the chip belongs: models with few active parameters that fit into the large shared memory. Windows on Arm is the baseline, and anyone who wants Linux still compiles their own kernel.