NVIDIA Expands RTX Spark Roadmap Through 2030
NVIDIA and Microsoft unveiled RTX Spark on May 31, 2026, a new superchip built to power Windows PCs designed for on-device personal AI agents, then laid out a multi-generation roadmap through 2030 at Computex Taipei. The initial Grace Blackwell RTX Spark pairs a Blackwell RTX GPU with 6,144 CUDA cores and a 20-core Grace CPU, offering up to 128GB of unified memory; Tom's Hardware and PC Gamer report NVIDIA's roadmap slide adds a Vera Rubin Spark generation with LPDDR6 memory in 2027-2028 and a Rosa Feynman Spark generation in 2029-2030. For practitioners, the takeaway is that NVIDIA is committing every future PC generation to a Spark-class chip, giving OEMs and toolchain teams multi-year visibility for planning on-device agent workloads.
The more consequential detail for hardware and platform practitioners is not the initial chip NVIDIA announced in May, but its follow-up commitment at Computex Taipei to a multi-generation Spark roadmap running through 2030, a rare public hardware cadence that gives OEMs, ISVs, and toolchain teams multi-year planning certainty for on-device AI PCs.
What happened
NVIDIA and Microsoft jointly announced NVIDIA RTX Spark on May 31, 2026 via official NVIDIA and Microsoft channels, a superchip built to reinvent Windows PCs for on-device personal AI agents. The Grace Blackwell RTX Spark pairs a Blackwell RTX GPU with 6,144 CUDA cores and a custom 20-core Grace CPU built with MediaTek, connected via NVLink-C2C, delivering up to 1 petaflop of AI performance and up to 128GB of unified LPDDR5X memory, according to NVIDIA's press release. At Computex Taipei, Tom's Hardware, PC Gamer, VideoCardz, and ServeTheHome report NVIDIA followed the launch with a roadmap slide extending the Spark platform through 2030: a Vera Rubin Spark generation in 2027-2028 adding LPDDR6 memory, then a Rosa Feynman Spark generation in 2029-2030.
Technical context
The move from LPDDR5X to LPDDR6, as reported by Wccftech and Heise, typically raises sustained memory bandwidth and lowers energy per bit for unified CPU/GPU memory pools, which matters for on-device workloads that share large tensors between CPU and GPU. NVIDIA's press release states RTX Spark can run 120-billion-parameter language models with up to 1 million tokens of context locally, which points to on-package memory bandwidth, not discrete DRAM, as the constraint that later generations are being designed around.
For practitioners
NVIDIA committing every future PC platform generation to a Spark-class chip reduces planning uncertainty for OEMs and ISVs building around Windows on Arm and on-device agent runtimes such as NVIDIA OpenShell. Teams optimizing for RTX Spark today should expect the underlying memory architecture, not just GPU core counts, to shift with each generation, and should track JEDEC LPDDR6 finalization and early silicon benchmarks before committing to memory-bandwidth-sensitive designs.
What to watch
Watch for OEM design wins and shipping RTX Spark laptops and desktops this fall from ASUS, Dell, HP, Lenovo, Microsoft Surface, and MSI; early benchmarks of unified-memory throughput on mixed CPU/GPU ML workloads; and LPDDR6 module availability as the Vera Rubin Spark generation approaches in 2027-2028.
Key Points
- 1NVIDIA and Microsoft launched RTX Spark on May 31, 2026, a superchip built to run personal AI agents locally on Windows PCs.
- 2At Computex Taipei, NVIDIA outlined a roadmap extending Spark through 2030, with LPDDR6-based Vera Rubin and Rosa Feynman generations following the initial Blackwell chip.
- 3OEMs and toolchain teams gain multi-year planning visibility, but each generation's memory architecture shift will require new bandwidth and thermal validation work.
Scoring Rationale
An official NVIDIA/Microsoft product launch (RTX Spark) followed by a multi-generation hardware roadmap through 2030, corroborated by NVIDIA's own press release and independent tech press (Tom's Hardware, PC Gamer, Heise, VideoCardz, ServeTheHome), is notable for OEM planning and on-device AI hardware practitioners, though it is a platform/hardware update rather than a new model or research breakthrough.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
