The desk
Die Brief

DIE.BRIEF

Hardware desk
Systems CONFIRMED

RTX Spark pairs a Blackwell GPU with a Grace CPU for Windows agents

The Die Brief Desk, after blogs.nvidia.com

The NVIDIA blog page says RTX Spark pairs a Blackwell GPU with up to 6,144 cores and a Grace CPU over a 600 GB/s link.

Picture from blogs.nvidia.com
Picture from blogs.nvidia.com

The NVIDIA blog page says RTX Spark combines a Blackwell RTX GPU with up to 6,144 cores. It pairs that GPU with an up to 20-core Grace CPU over a 600 GB/s link. Surface Laptop Ultra is named, with up to 128 GB of unified memory and up to a petaflop of AI compute. The blog also previews DGX Station for Windows, described as the first deskside AI supercomputer built for the Windows enterprise desktop.

ComputerBase, TechPowerUp and Wccftech added the launch to their feeds, but the vendor page does not record what they added or where they got wrong. Their headlines are not the document in the feed here. The vendor page names Acer, ASUS, Dell, HP, Lenovo, Microsoft, MSI and Gigabyte. The launch date, memory ceiling, and interconnect number come from that page.

The judgment is that this is a Windows platform bet, not a GPU spec sheet. The document ties the silicon to Microsoft Execution Containers, Agent 365, and a Windows enterprise desktop. That matters because the buyer is not choosing a card; the buyer is choosing an agent runtime that needs local memory and a secure container. The constraint is memory, not raw core count, and the filing does not prove that the 6,144-core part is in every system or that the petaflop figure is sustained.

What is still unmeasured is the price, the power draw, the thermal envelope, and the actual model size that fits in 128 GB. The page says preorders open today and gives availability dates, but it does not give a benchmark. No latency figure appears here yet. The compact desktops arrive in November, and the laptops on October 16, but the document leaves the real workload numbers out.

The buyer is choosing an agent runtime, not a card. The document does not prove that the 6,144-core part is in every system or that the petaflop figure is sustained.

The document leaves out price, power draw, thermal envelope, latency, and the actual model size that fits in 128 GB.

The constraint is local memory and container isolation, not raw core count. The vendor page does not prove that every system ships the 6,144-core part or sustains the petaflop figure.

up to 6,144 cores · up to 20-core NVIDIA Grace CPU · 600 GB/s · up to 128GB unified memory · One petaflop of FP4 AI performance · October 16

After blogs.nvidia.com. We did not report this. The pictures, if any, are theirs. Also filed by ComputerBase, TechPowerUp, Wccftech, High Yield.

October 07, 2026 · 3 min
Copy link Share to X