NVIDIA has officially unveiled RTX Spark, a new family of chips for Windows, based on the Grace and Blackwell chips. This announcement marks NVIDIA’s entry into the desktop processor market, with an all-in-one system solution rather than just a dedicated GPU. Of course, as with Qualcomm’s SnapDragon chips, it will run exclusively on the Windows ARM version. Although the rumor has now become official information, we’ll have to wait until autumn to see the first machines arrive. The Greens point out that the RTX Spark family is set to expand, and is designed for AI agents, content creation and gaming.
Details of the two RTX Sparks announced :
NVIDIA N1x: the most muscular chip in the family
The N1x will be offered in two variants.
The first features 20 CPU cores, divided into 10 + 10, with 48 SM on the GPU side.
This configuration would be equivalent to 6144 CUDA cores.
The second version would have 18 CPU cores, in 9 + 9, with 40 SM, equivalent to 5,120 CUDA cores.
Both N1x models would target an envelope of 45 to 80 W, but it should be noted that this figure refers to the complete CPU/GPU package, not just the CPU part.
In terms of connectivity, the N1x is said to integrate 12 PCIe 5.0 lines and 5 PCIe 4.0 lines.
Memory will be 16-channel LPDDR5X, with capacities ranging from 16 to 128 GB, depending on configuration.
NVIDIA N1: a more affordable version
The N1 non-X would also be available in two versions.
The first would offer 12 CPU cores, in 8 + 4, combined with 20 SMs, for a total of 2,560 CUDA cores.
The second version would feature 10 CPU cores, in 7 + 3, with 16 SM, i.e. 2048 CUDA cores.
This chip would aim for a more contained envelope, between 18 and 45 W, placing it more within the logic of slimmer or more affordable machines.
The PCIe part would be reduced to 8 PCIe 5.0 lines and 3 PCIe 4.0 lines.
For storage, NVIDIA is said to offer up to three M.2 SSDs on the N1x, compared with two M.2 SSDs on the N1.
Finally, N1 memory is said to be based on 8 LPDDR5X channels, with capacities ranging from 8 to 64 GB.
Nvidia looks to the future with RTX Spark roadmap
Nvidia confirms its ambitions for this family of SoCs, as the announcement of chips to be launched this autumn is complemented by a roadmap extending to 2029.
You now know the first stage, which will start in autumn 2026, with a generation based on Blackwell.
NVIDIA is placing the RTX Spark compact laptops and desktops alongside DGX Station for Windows.
RTX Spark is based on Grace Blackwell Spark silicon, accompanied by LPDDR5X memory.
DGX Station, meanwhile, moves up a notch with Blackwell Ultra hardware, HBM3e memory and a ConnectX-8 800G network.
Next up is Rubin, scheduled for the 2027-2028 window.
NVIDIA mentions Vera Rubin Spark, this time with LPDDR6 memory.
The same generation also includes Rubin with HBM4, a Vera CPU and CX9 1600G connectivity.
The idea is clear: to develop the more compact Spark machines in parallel with the more AI-heavy DGX Station-class systems.
But the roadmap doesn’t stop there…
NVIDIA has already announced the Feynman generation, with HBM Next, Rosa CPU, CX10 network, and two Rosa Feynman Spark entries scheduled for 2029 and 2030.
Ambition not only on laptops, but also on desktops
On desktop PCs, don’t expect a solution to replace your gaming PC, the target is quite different. Nvidia is targeting a machine designed for local professional AI, not gaming.
The system is based on the GB300 Grace Blackwell Ultra Desktop Superchip, which combines a Blackwell Ultra GPU with a 72-core Grace CPU. Communication between the two chips is via NVLink-C2C, to ensure high consistency between CPU and GPU computation.
NVIDIA has announced up to 748 GB of coherent memory, which is ideal for massive models and intensive workloads. In terms of performance, DGX Station for Windows can achieve up to 20 petaFLOPS in FP4.
NVIDIA reports that it can locally run AI models with up to 1 trillion parameters.
Availability is scheduled for Q4, via ASUS, Dell Technologies, GIGABYTE, HP, MSI and Supermicro.
