NVIDIA RTX Spark Laptops: N1X Specs, CUDA on Arm and Linux Support

NVIDIA RTX Spark is the DGX Spark chip squeezed into a Windows laptop. It pairs a 20-core Arm CPU with a Blackwell GPU and up to 128GB of shared memory. NVIDIA says the first laptops go on sale this month. Linux is the open question, and NVIDIA hasn’t answered it.

This explainer covers what the chip is, which laptops carry it, how CUDA works on Windows on Arm, and where Linux support stands. It also covers which local models fit in each memory tier. We have not tested an RTX Spark laptop. None were on sale when this was written, and no independent benchmark exists yet. Specs, driver releases and listings were verified against NVIDIA, Microsoft and OEM pages on 3 October 2026.

What RTX Spark is

NVIDIA announced RTX Spark at GTC Taipei during Computex, on 1 June 2026. MediaTek co-designed the CPU. The NVIDIA spec page lists three configurations and labels all of them “RTX Spark N1X”.

The NVIDIA porting guide fills in the rest. The CPU has two clusters of ten cores, each with five Cortex-X925 and five Cortex-A725. The GPU reports compute capability 12.1, the same as GB10. CPU and GPU share one coherent 256-bit memory bus. NVIDIA quotes up to 1 petaflop of FP4 AI compute. That is a vendor figure, and it does not say whether sparsity is counted.

Three numbers are softer than the rest. NVIDIA quotes 300 GB/s of memory bandwidth, but XDA reports that GB10 ships at 273 GB/s with LPDDR5X-8533. TSMC 3nm appears only in press reports, not on any NVIDIA page. Microsoft says RTX Spark PCs qualify as Copilot+ PCs, yet no NPU TOPS figure has been published.

RTX Spark and DGX Spark are the same silicon

Jensen Huang confirmed in September 2025 that the N1 silicon is the chip inside DGX Spark. TechPowerUp and Phoronix report the same lineage. The top laptop part matches the GB10 configuration: 6,144 CUDA cores, 20 CPU cores, 128GB. The laptop version drops the ConnectX-7 networking and runs at 45 to 80W instead of 140W.

That lineage matters for two reasons. First, software built for DGX Spark targets the same GPU architecture, so CUDA kernels compiled for sm_121 should carry over. Second, DGX Spark already runs Linux, which shapes the Linux question below. Our DGX Spark vs Strix Halo vs Mac Studio comparison covers the desktop box in detail.

Laptops announced so far

Six OEMs are in the first wave. Acer and GIGABYTE are listed to follow. NVIDIA said at IFA that the first systems go on sale starting in October.

No maker has published a price. Figures such as “$2,899” circulate online, but they trace back to low-quality sites, not OEMs. Microsoft holds a Windows and Surface event in San Francisco on 7 October. Satya Nadella and Jensen Huang are both expected on stage. Surface pricing and ship dates are expected there, not confirmed. Smaller memory configs exist too. A leaked Surface prototype had 24GB, so not every RTX Spark laptop is a 128GB machine.

No RTX Spark device is listed on Amazon US yet. This page will add prices and links once listings go live.

CUDA on Windows on Arm

Every RTX Spark laptop ships with Windows 11 on Arm. CUDA Toolkit 13.4 arrived on 9 September with native Windows on Arm support. It builds both ARM64 and ARM64EC binaries. Before 13.4, CUDA on Arm meant Linux only.

Microsoft tuned Windows for the chip. The scheduler knows about the 20-core layout, Prism emulation was tuned for the CPU, and the GPU can address more of the shared memory. Easy Anti-Cheat and BattlEye now support Arm, and EA is bringing its Javelin anti-cheat natively. Game demos at IFA ran through Prism, but frame counters were not allowed.

One detail matters for AI work. The NVIDIA porting guide says Windows splits GPU memory into a dedicated carveout and a shared segment. NVML-based tools such as nvidia-smi report only the carveout. The guide also tells developers to avoid cudaMallocManaged on Windows. So the memory a model can use will be less than the number on the box.

Drivers are the risk. A Surface prototype tested in July ran a preview driver, and TechTimes reports that its CUDA AI tests failed. That was prototype hardware on preview software. Retail units should ship with a newer driver, but nobody has confirmed CUDA AI workloads run cleanly on Windows yet.

The Linux question

NVIDIA hasn’t announced Linux support for RTX Spark laptops. The spec sheet lists Windows 11 for every part. XDA asked, and NVIDIA would not comment. Phoronix wrote that RTX Spark “will still work great on Linux”. That is the view of Phoronix founder Michael Larabel, not an NVIDIA statement, though some sites quote it as one.

The public driver does not cover it yet. NVIDIA open GPU kernel modules 615.71.09 were tagged on 9 September. The README lists GB10 as a supported device, but no RTX Spark N1X device IDs. The Windows porting guide gives one N1X PCI ID as 10de:2e03, and that ID is absent from the Linux list.

NVIDIA itself runs Linux on these laptops, though. Its public NemoClaw GitHub tracker holds 18 issues with an “[N1x Linux]” title prefix from August and September. They describe an internal developer image called N1x FastOS. It is based on Ubuntu 24.04 for aarch64, with an NVIDIA 7.0 kernel, the 615 open driver and CUDA 13.4. In those reports, newer units name the GPU “NVIDIA RTX Spark N1X” and earlier ones a generic placeholder. Docker GPU passthrough works. All of this comes from one NVIDIA repository. It shows the stack exists. It does not show NVIDIA will release it.

The same issues show rough edges. On N1X, nvidia-smi returns N/A for memory usage, as it does on unified-memory parts. One report has the vLLM container OOM-killed at --gpu-memory-utilization 0.4 while loading weights, which left the host unresponsive. That unit reported 61 GiB of shared memory.

DGX Spark is the best guide to what might happen. It ships DGX OS, an Ubuntu base with an NVIDIA kernel, and NVIDIA validates only that image. The NVIDIA Ubuntu kernel still carries Grace Blackwell patches that are not in mainline. Others have ported it anyway. There is an Arch kernel project, a Rocky Linux 10 build, and a Red Hat guide to building a RHEL 10 kernel for DGX Spark. David Airlie has Nouveau and NVK running on GB10 with experimental patches. CUDA on Linux aarch64 is mature, since server Arm has used it for years. The driver and toolkit flow mirrors the x86 one in our guide to installing NVIDIA drivers and CUDA on Ubuntu.

A laptop adds problems a desktop box does not have. Firmware must expose hardware in a way Linux understands, through ACPI or a device tree. Snapdragon laptops have needed a device tree written for each model. No OEM has said how RTX Spark firmware boots, no distro has published an image, and Canonical has said nothing. Fedora has only a community thread.

The honest answer today: a Linux RTX Spark laptop is likely, but it does not exist yet. If you need Linux on day one, don’t buy one on the hope of it. WSL2 is the nearer option. NemoClaw already has an Ubuntu 24.04 aarch64 WSL path on N1X, though its own QA issues show it still breaks.

What fits in RTX Spark memory for local AI

NVIDIA claims the 128GB laptops run 120B-parameter models with up to a million tokens of context. That is a vendor claim. The table below is our estimate from published weight sizes. Windows reserves part of the shared memory, so plan for less than the full figure.

Our LLM VRAM calculator works out the exact figure for a given model, quantization and context length.

Speed is unknown. No one has measured tokens per second on an RTX Spark laptop. The closest reference is the official benchmark from llama.cpp on the 140W DGX Spark. It ran gpt-oss-120b at 58.7 tokens per second for generation and 2,444 for prompt processing. Generation speed depends mostly on memory bandwidth, which should be close to GB10 if laptops use the same LPDDR5X speed. Prompt processing depends on GPU compute, which the 45 to 80W limit will cut. Expect generation numbers nearer the desktop than prefill numbers. Treat that as a reasoned guess until reviews land.

Bandwidth puts RTX Spark in the same class as AMD Strix Halo, well below the Apple M5 Max.

Prices are Amazon US on 3 October 2026. On bandwidth alone, the M5 Max should generate tokens about twice as fast as RTX Spark or Strix Halo. RTX Spark brings CUDA instead. That means vLLM, TensorRT-LLM and PyTorch builds without a translation layer. For the wider trade-off, see unified memory vs VRAM for AI.

Should you wait for one

Wait if you want CUDA in a laptop with more than 24GB for models, and Windows is fine. No other laptop offers CUDA with more than 24GB for models. Prices and ship dates should firm up on 7 October.

Do not wait if you need Linux now. A Strix Halo laptop runs Linux today, and ROCm and Vulkan both run llama.cpp on it. Do not wait if raw generation speed matters most either. The M5 Max has more than twice the bandwidth. If you only need CUDA and a 24GB or 32GB model, a desktop with a discrete GPU for LLMs is usually faster per dollar.

We will update this page with prices after the October 7 event, and with measured numbers once retail units ship.