Skip to content
NewsProduct launch

Nvidia's RTX Spark laptops ship with 128GB of memory, but the GPU can only use 110GB of it

· by Pondero Newsdesk

The short version

Nvidia and Microsoft opened preorders for RTX Spark Windows laptops on October 7, and a Microsoft engineering post shows the 128GB unified-memory SKU only exposes about 110GB to the GPU once Windows and the OS take their share.

Nvidia's RTX Spark laptops ship with 128GB of memory, but the GPU can only use 110GB of it

Nvidia and Microsoft opened preorders October 7 for RTX Spark Windows laptops built to run large models offline, and the spec sheet number buyers will anchor on, 128GB of unified memory, is not what the GPU actually gets to use. A Microsoft and GitHub engineering post, surfaced in StorageReview's reporting on the launch, shows a 128GB Surface Laptop Ultra exposing only about 110GB to the GPU once Windows and system overhead take their share, per StorageReview.

What

Jensen Huang and Satya Nadella unveiled RTX Spark at a joint Windows AI and Surface event in San Francisco, where Microsoft also took its agent-sandboxing layer, Microsoft Execution Containers, to general availability on Windows 11, per Nvidia's blog post on the event. RTX Spark pairs a Blackwell RTX GPU, up to 6,144 cores, with a Grace CPU of up to 20 cores over a 600 GB/s link, rated at one petaflop of FP4 performance, which Microsoft's own footnotes describe as a theoretical, sparsity-dependent figure rather than sustained throughput. Laptops from Acer, Asus, Dell, HP, Lenovo, Microsoft, MSI and Gigabyte ship October 16; compact desktops follow in November. Nvidia also previewed DGX Station for Windows, built on a GB300 Grace Blackwell Ultra Desktop Superchip with 748GB of coherent memory and up to 20 petaFLOPS of FP4 compute, pitched as bringing trillion-parameter-scale inference to a Windows box without a separate Linux machine. Nvidia set no price or ship date beyond a signup page; StorageReview reports the company told press to expect a fourth-quarter 2026 release, per StorageReview's coverage.

The memory math decides what actually runs

That 110GB ceiling, not the 128GB nameplate figure, is the number that matters for anyone sizing a local deployment, because model weights, the KV cache that holds an agent's growing context, and runtime overhead all compete for it. Microsoft's own quantized build of its 137-billion-parameter MAI Code 1.1 Flash model fits in roughly 53GB at about 3.3 bits per weight, down from a 274GB footprint at full BF16 precision, and scored 70.8% on SWE-bench against 72.6% for the cloud version, a gap StorageReview calls close enough to read as near parity, per StorageReview's analysis of Microsoft's benchmark data. Buyers configuring below the top SKU should not expect the same headroom: Nvidia's preorder listings also include 24GB, 32GB and 48GB variants, and the large-model claims in the launch material all assume the 128GB laptop.

What to watch next

Independent, non-vendor benchmarks will start arriving once laptops actually reach buyers October 16. Also worth tracking is whether Nvidia attaches a firm price and date to DGX Station for Windows once the quarter it previewed for actually arrives.

Sources