RTX Spark After Launch: First Measurements, the Variant Trap, and the Linux Question

RTX Spark is NVIDIA’s umbrella name for laptops and compact desktops built on the N1X superchip, the laptop derivative of the GB10 in the DGX Spark. (Quelle: Simple Icons, CC0)

On October 16, 2026 the RTX Spark machines went on sale at ASUS, Dell, HP, Lenovo, Microsoft, and MSI, from 2,599 US dollars (press release of October 16, 2026). The core is the N1X superchip: Blackwell GPU and Grace CPU in one package, unified LPDDR5X, the same architecture family as the GB10 in the DGX Spark, covered here on October 3 after our own re-calculations. We had the IFA announcement in the ticker; our QN10 test of September 21 told readers to wait for this machine class. What was missing then: measurements and prices. Both arrived, from two independent labs.

What NVIDIA states officially, and what it does not

The spec table lists two laptop variants of the N1X. The upper one pairs 6144 graphics cores and a 20-core Grace CPU, adds 1 petaFLOP of FP4, up to 128 GB of LPDDR5X on a 256-bit interface, and a combined CPU-plus-GPU power budget of 45 to 80 watts. The second lists 5120 graphics cores, 18 CPU cores, and 827 teraFLOPS. Pulling these pages October 17, we noticed three inconsistencies inside NVIDIA alone:

For the smaller N1X the product page says "Up to 64 GB", the spec page of the same variant "up to 32GB LPDDR5X".
For that variant the spec table carries 827 teraFLOPS, the image alt text on the same page 784 teraFLOPS FP4.
The product page promises models "up to 120B parameters with 1M context", the chooser page puts the 128 GB tier "up to 200B parameters".

Both numbers stay standing at each location, nothing gets averaged. The chooser page sorts tiers by memory: 24 GB up to 12 billion parameters, 32 GB to 25B, 48 GB to 60B, 64 GB to 90B, 128 GB to 200B, all vendor claims. One figure is missing from every NVIDIA page we pulled: memory bandwidth in GB/s, just "256-bit" and "LPDDR5X". That was the open question in IFA week, and it stayed open through launch. The Microsoft product page for the Surface Laptop Ultra, pulled October 17, lists 24 to 128 GB, from 2,899 euros in Germany; the Austrian page showed no price at fetch time. So "up to 128 GB" at NVIDIA again means the largest configuration; the orderable base is considerably smaller. On the Surface the 24 GB tier even ships with Windows 11 Home, Windows 11 Pro only starts at 32 GB.

What the labs measured

c’t 3003 tested the 32 GB Asus ProArt P14 with llama-bench, no extras: 14.5 tokens per second with the dense Qwen 3.8-27B in IQ4_XS quantization, just under 61 with the MoE model Qwen 3.6-35B-A3B (October 16, 2026). c’t found the denser model better for programming, names multi-token prediction as reserve with at least a 25 percent uplift, more when coding. StorageReview tested two HP OmniBook Ultra 16 units with 32 and 128 GB plus the Surface Laptop Ultra and reports 56.5 tokens per second with Llama 3.1 8B, 58.0 with gpt-oss 20b, and 89.5 with Qwen 3.6-35B-A3B on the 128 GB device; its coding test ran three consecutive times, fastest in 72.0 seconds, best in the group (October 16, 2026). Across all four N1X machines: prompt processing 5,837 to 7,142 tokens per second, the RTX 5090 laptop in the same lab 11,560.

Four models, two labs, as of October 16, 2026, each without multi-token prediction. The bars tell more about model choice than the chip: on the same Asus the MoE model ran just under 61, over four times the dense variant. (Quelle: c’t 3003 / heise online (32 GB Asus) and StorageReview (HP OmniBook Ultra 16, 32 and 128 GB))

Against hardware we know: our 5090 lane delivers close to 60 tokens per second single-stream on Qwen 3.8-27B in 4 bit (measured here), the Mac Studio M5 Ultra reached 46 to 54 in the same class (our measurements), and our AI PRO R9700 runs 41 to 45 with llama.cpp and an MTP draft in the house voice box (own wall-clock run of October 16, 2026, 262k context, with draft model). The 14.5 from the c’t test is the low outlier, and it applies to the laptop model many want: a full dense 27B.

Power readings show why the laptop chip trails the box chip: with no DGX Spark in the lab, c’t compared against archived runs that measured up to 216 watts on the DGX Spark against a 60-watt short peak on the Asus notebook, roughly 40 average. On those older models the DGX box often delivered double. Same picture on memory: StorageReview measured 112 to 122 GB/s in STREAM Triad on all four tested N1X devices, against 48.3 GB/s of a Razer Blade 16 with DDR5. From LPDDR5-9400 and 256-bit width StorageReview calculates around 300 GB/s in theory; a good third of that arrives in the CPU-side Triad. For comparison: the DGX Spark GB10 carries 273 GB/s in its datasheet, our DGX run reached 85 percent of that, and the Mac Studio M5 Ultra a measured 1,039 of 1,200 promised. No test in this wave carries a GPU-side bandwidth measurement, the open number of launch week.

StorageReview also filed a defect finding on the Surface: in the OpenBLAS full-load test two identical units shut down completely several times on October 14, last at 95 degrees, Kernel-Power event 41, no dump; the author reads the logs as pointing to power delivery or firmware protection, ruling out heat. The HP machines held the same loads at 98 degrees.

Euro prices: Asus and MacBook Pro from the c’t test of October 16, 2026, Surface base from the Microsoft product page of October 17, 2026, DGX Spark from our own test of October 3, 2026. The desktop box with full memory stays cheaper than every 128 GB laptop here, and in the c’t comparison runs roughly twice as fast. (Quelle: c’t test October 16, 2026, Microsoft product page October 17, 2026, own DGX Spark test October 3, 2026)

Memory and price: where the decision is made

According to StorageReview the 32 GB HP OmniBook kept up on speed but lost every model-loading and agent test: after Windows and reservations only 12 GB remain for the GPU, and gpt-oss 20b alone needs 12.8 GB. The 128 GB device exposed 62.5 GB of GPU memory and passed the full ladder; StorageReview names roughly 40 GB of GPU share as the floor for the large MoE model. Its buying advice verbatim: buyers who want the N1X for local AI should look at 64 or 128 GB. The memory is soldered, the chosen size stays final. The fixed share is a Windows problem: Windows does not yet share unified memory as freely as the Linux stack on the DGX Spark; work on that is announced. c’t ran the math on the 32 GB Asus: Q4_K_M worked, but it got tight once a second program opened, and Photoshop ran alongside the model only with the smaller IQ4_XS variant. After the test c’t wanted at least 48 GB. Translated into prices: the 32-to-128 GB jump adds 3,400 euros at Asus and 3,030 US dollars at HP, and the tested 128 GB Surface stands at 5,899.99 US dollars (price as of October 14, 2026). The business is in the extra gigabyte, the compute stays the same at every tier.

Windows, WSL, and the DIY Linux route

NVIDIA’s spec table names Windows 11 as the only operating system. Linux appears in the developer blog only via the Windows Subsystem for Linux, and on the network only as a recipient: NVIDIA’s open router PAIR should spread inference to Windows, Linux, and macOS boxes at home. The developer blog of October 16 otherwise celebrates the Windows-on-Arm buildout: first PyTorch CUDA release, native Python 3.15, over 2,000 Conda packages, 97 percent of the 15,000 most important Python packages ARM-capable, plus factors against Intel’s Panther Lake: 4.6 times on agents, 2.7 times on fine-tuning, 1.8 times on compiling, all vendor claims.

The reviewer at c’t 3003 tried native Linux anyway: standard Ubuntu for ARM boots from a USB stick, then runs without NVIDIA, touchpad, or keyboard drivers on a very rudimentary UEFI. With Claude Code as help and a self-compiled kernel he got all three working; both models ran, the MoE model 9 to 14 percent faster than under Windows. Curious, for a platform launched as a reinvention of the Windows PC. Standing question: when Linux arrives, with whose driver support.

Jan-Keno Janssen (c’t 3003, October 16, 2026) with the Asus ProArt P14 in the lab: measurements, Linux experiment, loudness readings. Transcript with the linked heise article.

What this means for procurement

This machine class answers a question that does not come up for us in this form, but does for two types in our client base. The consultant or executive traveling with client data, wanting models without a cloud path, gets a usable machine with 64 GB and a MoE model of this class, the model stays on the disk. The fan belongs in the decision: c’t measured 2.7 sone at full load and noted the Asus often spins up already in desktop use, against 4.8 sone on the large MacBook Pro M5, there only under load. That fits over 19 hours of video playback at 200 nits and an OLED with a measured 599 nits. Whoever buys compute for the home office without needing portability buys the wrong category by these numbers: the DGX box with full memory stays cheaper than any 128 GB laptop here, delivers double the speed in the archived comparison, and runs in the Linux world we live in anyway. For mobile use: 48 GB is the minimum, 64 the sensible middle, and anyone running inference under Linux buys a hobby track with the N1X for now. NVIDIA has the desktop version of the N1X slated for November, without a price.