RTX Spark laptops render 90 GB scenes on battery
On October 16 NVIDIA opened sales of RTX Spark laptops — Windows machines built on the company's own superchip. In the N1X chip, the Grace processor and Blackwell graphics are joined by the NVLink-C2C interconnect and share a single memory pool of up to 128 GB. For 3D artists, compositors and editors this removes the main limitation of a mobile workstation: scene size no longer runs up against a separate pool of video memory.
NVIDIA states that RTX Spark renders scenes of up to 90 GB through OptiX and edits 8K video in 4:2:2 — including without a power outlet. The compute side delivers up to 1 petaflops in FP4 on fifth-generation tensor cores. Prices start at $2,599; configurations and cost depend on the region.
Shared memory removes the familiar scene ceiling
Previously an artist who hit the video memory limit of a laptop RTX had to split the scene, cut textures or move to a desktop and a render farm. In RTX Spark the processor and the graphics address the same array of LPDDR5X, so heavy creative projects and local models run at the same time, without separate memory pools. NVIDIA cites as an example running language models with 120 billion parameters and a context of a million tokens right on the device.
There are two configurations. The higher one — a 20-core Grace, 6,144 Blackwell RTX cores and up to 128 GB of memory; the lower one — 18 CPU cores, 5,120 GPU cores and up to 64 GB. The chip's thermal budget in laptops is 45–80 W, and 140 W in the desktop version. The chassis have remained laptop-like: diagonals from 14 to 16 inches, thickness from 14 mm, weight from three pounds (about 1.4 kg), milled aluminium, tandem OLED panels with G-SYNC.
The first wave — ASUS ProArt P14 and P16, Dell XPS 16 Creator Edition, HP OmniBook Ultra and OmniBook X 14, Lenovo Yoga Pro 9n, Microsoft Surface Laptop Ultra, MSI Prestige N16 and Prestige N16 Flip AI+; sales run through the manufacturers' websites and major retail. Compact desktop systems on the same chip are announced for the autumn, with Acer and GIGABYTE joining later.
Adobe is rebuilding Photoshop and Premiere for the new architecture
The platform has been backed by more than a hundred Windows software developers, including Adobe, Blackmagic Design, Blender, CapCut, ComfyUI and OTOY. Adobe is rewriting Photoshop and Premiere for RTX Spark with acceleration through TensorRT and promises up to a twofold gain on AI features, editing, colour grading and effects; at launch native support in both applications is available in beta status, while Substance 3D Painter and Stager will get it later.
A separate point for those who work with path tracing: Blender 5.3 will bring DLSS 4.5 Ray Reconstruction with the second generation of the transformer model — the same technology is announced for dozens of games. The Blackwell decoder adds hardware handling of AV1 and 4:2:2, and RTX Video with fourfold frame generation is coming to ComfyUI: new nodes turn AI video at 15 frames per second into 60 fps and convert it to HDR without cloud charges for every request.
The move to Arm will cost you in compatibility
The processor was designed jointly with MediaTek, one of the leaders of the Arm system-on-chip market, so applications without native Arm builds run on the platform through x86 emulation in Windows. This is the main risk for a studio pipeline: plug-ins, codecs, network licence managers and utilities without Arm versions will have to be checked separately.
For developers NVIDIA states support for 97% of the 15 thousand most popular Python packages, the first PyTorch build with CUDA for Windows on Arm, native Python by default and a native Conda installer; setup is simplified by Windows Developer Config, and the familiar Linux tools are available through Windows Subsystem for Linux.
On games NVIDIA promises hundreds of titles at launch and 1440p above 100 frames per second with DLSS 5, Reflex and G-SYNC — among the titles launching are Cyberpunk 2077, Fortnite, Marvel Rivals, World of Warcraft, Gears of War: E-Day and NBA 2K27. NBA 2K27 runs DLSS 5 with the 3D-Guided Neural Rendering feature, and a twice-as-fast AI model, available from today, raises the frame rate in DLSS 5 by up to 1.4 times.
In parallel NVIDIA is promoting local agents: GitHub Copilot, Hermes, OpenClaw, Perplexity, Pi and WorkBuddy are announced for RTX Spark with installers that need no command line, agent permissions are limited by the NVIDIA OpenShell runtime and Windows containers, and NVIDIA PAIR makes it possible to distribute inference to other computers on the home network. Full specifications and configurations are published on the RTX Spark page.
Why this is happening right now
RTX Spark was unveiled on May 31 at GTC Taipei — NVIDIA spent about four and a half months preparing the ecosystem, and that is probably no accident. Releasing the chip turned out to be easier than agreeing with Adobe, Blackmagic Design, the Blender community and game publishers on native builds: without them an Arm laptop is of little use to a professional, however fast it may be. Judging by the fact that the software partners are named individually, NVIDIA has taken into account the experience of previous attempts to move Windows to Arm, where the hardware came out noticeably earlier than the programs.
Behind the architecture itself lies a larger shift: computation for generative models is gradually returning from rented servers to local machines. A unified memory pool is equally convenient both for a heavy scene and for a language model that a studio does not want to hand to the cloud because of a project NDA. If the trend continues, the boundary between a graphics workstation and an AI machine in the laptop segment will be erased for good, and specialists will start to be expected to know how to deploy models locally.
For makers of classic mobile processors this is a direct threat: NVIDIA is entering the market of central processors for PCs with its own chip, and straight into the premium segment. It is telling that the company is hedging its bet on both sides — earlier it announced a partnership with Intel, under which Intel will release x86 systems-on-chip with built-in RTX graphics. NVIDIA is leaving the architectural choice to the market, and with a price tag from $2,599 studios would do better to wait for tests on their own projects — especially where the pipeline rests on old plug-ins without Arm versions.
Community reaction
On the day the RTX Spark laptops went on sale, the tone split between excitement and restraint. People from graphics and development are drawn to the combination of Blackwell and up to 128 GB of shared memory in a thin Windows computer: an author of tools for computer graphics writes outright that he has never seen anything like it in a laptop, and developers note that the memory is enough at once both for local models and for coding agents on the same CUDA stack. By The Verge's measurements, the Blender Classroom scene on the N1X finishes in the same 15 seconds as on a MacBook Pro with M5 Max, meaning that on graphics the chip is placed on a par with Apple's best solution. The argument is not about whether the machine exists, but about whether it is worth its money already now. With the same 128 GB and the same price, the Mac responds faster, edits video faster and lasts about twice as long on battery, while the processor falls behind in multithreading. In game reviews people are annoyed by frame rates only with frame generation and DLSS, without an honest native mode. Separately, people write that the day-one driver looks raw, and that Windows on Arm software has not yet caught up with the hardware. Interest is strong among those who need local generation and heavy scenes in memory, but many are postponing the purchase.
«NVIDIA, what are you doing?! A one-petaflops Blackwell graphics chip and 128 GB of shared memory — inside a Windows laptop. This is RTX Spark: a Grace processor plus a Blackwell graphics chip.»
— @NuwanPanditha (X)
«The Verge compared the RTX Spark N1X with a MacBook Pro on M5 Max. In Blender Classroom both came in under 15 seconds. On graphics, parity with Apple's best chip; on the processor, no. Even a Mac mini on M6 for 1,299 dollars beats it in multithreading.»
— @rudraxx13 (X)
«I pitted NVIDIA's first laptop chip, RTX Spark, against a MacBook Pro on M5 Max. Same price, 128 GB of memory each. NVIDIA is faster on images and on long texts with a 120-billion-parameter model. The Mac responds faster, edits video faster and lasts twice as long on battery.»
— @mattmoniz (X)
What is behind RTX Spark's arrival in laptops: the pressure of local AI, a dead end for discrete cards, the fight for studios at $2,599?