RTX Spark N1X: 128GB of CUDA Wrapped Around a CPU That Feels Late

NVIDIA's first serious Windows-on-Arm laptop chip is not really a laptop chip. It is a way to put 128GB of CUDA-addressable memory in a backpack, and the N1X delivers that along with excellent prompt processing and a CPU design that was already being superseded when the silicon taped out. Whether the trade is worth $5,900 depends entirely on what you do with it.

Two Configurations, One Confusing Name

The N1X is half of the RTX Spark family, the other half being the lower-power N1 aimed at thin-and-lights in the 18W to 45W range. Both sit on the GB10 Grace Blackwell Superchip that debuted in the DGX Spark, which means the graphics side is known good and the CPU side is inherited. Two N1X configurations exist and neither has a model number to tell them apart. One runs 18 cores split evenly between Cortex-X925 and Cortex-A725, 5,120 CUDA cores, 160 tensor cores and 32GB of memory. The other runs 20 cores, 6,144 CUDA cores, 192 tensor cores and up to 128GB.

Both use a 256-bit memory interface with up to 300 GB/s of bandwidth, slightly ahead of the DGX Spark's 273 GB/s, and both are rated 45W to 80W split between CPU and graphics against the 140W the desktop-adjacent GB10 can draw. The CPU, I/O and memory tiles were built with MediaTek. The cores themselves are Arm's Cortex-X925 and A725, announced in May 2024 and superseded by the C-Series in September 2025, which is the quiet problem with this entire product.

Arm Windows Is Not New; This Is the First Time It Has NVIDIA's Money

NVIDIA has tried this before. Tegra 3 powered the original Surface RT in 2012, after which the company left Qualcomm to carry Windows on Arm more or less alone. Reports suggest an exclusivity arrangement kept other vendors out until 2024; whatever the truth of it, the field is open now and the money is real.

What NVIDIA inherits is Prism, Microsoft's x86-to-Arm translation layer, which has had years of tuning and now handles AVX and AVX2, instruction sets that demanding applications depend on. That is why every test in the suite had a native Arm build while the scores still landed within reach of x86 rivals. Apple is winding Rosetta 2 down because there are no Intel Macs left that it wants to support. Microsoft cannot do the same while AMD and Intel keep shipping x86, so Prism stays, and NVIDIA gets it free.

Gaming: Adequate, With an Asterisk the Size of DLSS

NVIDIA's claim of 100+ FPS at 1440p holds only with upscaling and frame generation enabled, which the fine print does admit. Without either, across Total War: Pharaoh, Grid Legends, Cyberpunk 2077 and Shadow of the Tomb Raider, all running through Prism, the N1X is simply adequate. The discrete RTX 5070 Ti in the comparison laptop wins comfortably on a much larger power budget, and AMD's Ryzen AI Max+ 395 beats both N1X machines in all four titles while running them natively instead of translated.

The silicon explains the outcome. The N1X carries CUDA core counts in desktop RTX 5070 territory, but its SMs receive a fraction of the power, so expecting 5070-class results was never reasonable. What is less reasonable is shipping a flagship laptop chip in 2026 whose CPU cores cannot keep pace with a Snapdragon or an M5 Pro. DLSS with frame generation is what closes the framerate gap, and NVIDIA's DLSS 5 push is the entire reason that 100+ FPS framing works. None of it touches the compile times.

AI Is Where the Money Actually Is

The pitch is local AI first, agentic workloads second, and the hardware backs it up. In llama.cpp prompt processing, the compute-bound half of inference, the N1X beats everything in the comparison set by a wide margin. Token generation, which is bandwidth-bound once the cache is warm, lands roughly level with the rest, exactly as a 300 GB/s ceiling predicts.

Two practical notes follow. Use CUDA rather than Vulkan, because the Vulkan path was consistently slower on identical hardware, with TensorRT the other supported option. And remember that a depth-zero benchmark is a best case: real agent sessions carry long contexts that change the picture. A small model like Gemma 4 E4B runs anywhere in this lineup, while a 27B model at Q4 is the test that separates the memory configurations, which is the actual argument for the Surface's price.

Battery, Heat, and a Price That Needs Justification

Battery endurance is where the Arm story collapses. Playing a local 1080p video at 200 nits and 60Hz, the 128GB Surface lasted nine hours on a 92Wh pack while the 99.9Wh ProArt managed just over twelve. Under sustained 1440p graphical load both RTX Spark machines ran for roughly an hour. That is x86 behavior wearing an Arm badge, and it undercuts the efficiency premise the whole category rests on.

Thermals are less alarming than expected: the chip runs hot under load but does not appear to throttle, though sensor support is immature enough that the readings deserve skepticism. Pricing is the harder problem. $5,899.99 for the 128GB Surface and $4,399.99 for the 64GB ProArt, against $3,716.69 for a 128GB Ryzen AI Max+ 395 machine. AMD is the cheaper route to local-AI memory, and Apple reaches 64GB at $3,699 before leaping to $6,699 at 128.

What I Think

If the CPU were current, this would be a landmark product. A laptop with a full CUDA stack and 128GB of unified memory solves a real problem for anyone running models locally, and nothing on the Windows side does that today. The prompt-processing lead is not a marketing artifact, it is the difference between an agent that answers and one that sits there thinking.

But NVIDIA shipped last-generation CPU cores in a device that costs more than a used car, and the battery numbers prove efficiency was never the point. My read is that this is a first draft: the memory architecture is right, the software stack is right, and the compute complex needs one more generation before the price makes sense. Buy it if 128GB of CUDA in a bag is worth $5,900 to you today. Otherwise wait, because the next one will be better and the one after that will be cheaper.