News
NVIDIA's RTX Spark: A 128GB Windows Superchip for Laptops, and the Claims Still Waiting for Tests
ยท RenderBob team
NVIDIA's RTX Spark pairs a Blackwell GPU with a Grace CPU and up to 128GB of unified memory for Windows laptops and compact desktops.
NVIDIA's RTX Spark pairs a Blackwell GPU with a Grace CPU and up to 128GB of unified memory for Windows laptops and compact desktops due this fall. Partner pre-order prices are now listed. Independent tests of sustained load are not.
What was announced
NVIDIA introduced RTX Spark at Computex in June, developed with Microsoft. NVIDIA's product page describes a top configuration: a Blackwell GPU with 6,144 CUDA cores and fifth-generation Tensor cores with FP4 support, a 20-core Grace CPU, and up to 128GB of unified memory, linked over NVLink-C2C. NVIDIA cites up to 1 petaflop of AI compute. The same table lists a lower configuration with fewer cores and less memory, so a listing that says RTX Spark is not automatically the 128GB part. Coverage of the Computex launch named ASUS, Dell, HP, Lenovo, Microsoft Surface, and MSI, with Acer and Gigabyte later. Slim designs in that coverage are described as 14 to 16 inch laptops as thin as 14 mm.
Editing claims
NVIDIA says the chip can edit 12K 4:2:2 video using the Blackwell decoder. Adobe is reported to be building a new Premiere pipeline for real-time editing and colour correction, and Photoshop is being optimised for GPU compositing. This links to our Premiere article this week on generation arriving inside the timeline. All performance figures are NVIDIA's and have not been independently verified.
Why it matters for AI video work
128GB of unified memory on a Windows laptop moves the Windows side towards what Apple's machines already offer for local models. For a studio that standardises on Windows, a portable machine that can keep large models resident would be new. The open questions are practical: sustained performance in a 14 mm chassis, how much memory is available to the GPU in practice, whether the model runtimes you use support it, and which bin the listing actually sells.
How to decide
Wait for retail units and for reviews that run real generation and export workloads for long periods. A short benchmark in a launch demo does not show thermal behaviour. If you are choosing between this and an M5 Ultra Mac Studio, compare on your own models, memory needs, and software. Read the core count and memory on the listing before you compare prices.
What to watch
Ship dates from each partner, and reviews that test sustained load on the configuration you would actually buy.
More from the blog
- Higgsfield Says AI Video Now Runs Past 10 Minutes: What an OpenAI Customer Story Does and Doesn't Show
An OpenAI customer story from 13 October says Higgsfield can now produce AI video longer than 10 minutes using GPT-6 Astra. The claims are the company's, and no independent test has been published.
- Disney Bids $20 Million for a License to Character.AI's IP: What the Stalking-Horse Deal Does and Doesn't Say
Character.AI filed for Chapter 11 on 12 October, and Disney Worldwide Services is the opening bidder at $20 million for a non-exclusive, perpetual licence. Higher bids can still come in.