The NVIDIA Spark Roadmap: Vera Rubin and Rosa Feynman Spark Generations Next

NVIDIA put a three-generation plan on the table at Computex 2026, and the NVIDIA Spark roadmap now reads like a miniature version of the company's data-centre map. The current Grace Blackwell Spark gives way to a Vera Rubin Spark, then a Rosa Feynman Spark, each named after the same scientist-pairing scheme NVIDIA uses for its big AI systems. For anyone weighing a desktop AI machine, that progression tells you what is coming and when.

Quick Answer

The Spark line runs in three confirmed steps: Grace Blackwell Spark (the GB10 generation shipping this year), Vera Rubin Spark around 2027 to 2028 with a move to LPDDR6 memory, and Rosa Feynman Spark around 2029 to 2030. Each step swaps both the CPU and GPU architecture in lockstep, mirroring NVIDIA's data-centre roadmap.

Where the Roadmap Starts

The current generation is Grace Blackwell Spark, the GB10-class part, expected to ship in the second half of 2026 with up to roughly 1 petaflop of AI performance in a desktop-sized form. It pairs NVIDIA's Grace CPU architecture with Blackwell GPU silicon, the same naming convention you see on the data-centre side. That pairing is the key to reading the whole roadmap: NVIDIA advances the CPU and GPU together, and the product name reflects both halves.

This matters because Spark is aimed at people who want serious local AI compute, model fine-tuning, large-context inference, and heavy creative pipelines, without renting cloud time. It is a different animal from a gaming PC, and the roadmap signals long-term commitment to that local-AI desktop category.

Generation Two: Vera Rubin Spark

The next step, slated for roughly 2027 to 2028, is Vera Rubin Spark. The CPU architecture moves from Grace to Vera and the GPU from Blackwell to Rubin. The headline hardware change is memory: a transition from LPDDR5X to LPDDR6, which brings higher bandwidth and better power efficiency. For AI workloads that are constantly shuttling model weights and context in and out of memory, bandwidth is often the real ceiling, so this is more than a spec-sheet bump.

The timing also lines up with the data-centre side, where NVIDIA confirmed Vera Rubin entered full production in mid-2026. The desktop Spark version follows the big systems by a generation, which is the usual pattern: prove the architecture at scale, then bring it to the deskside box.

Generation Three: Rosa Feynman Spark

Further out, around 2029 to 2030, is Rosa Feynman Spark. The CPU moves to Rosa and the GPU to Feynman. NVIDIA has named the generation and placed it on the timeline, but concrete specifications are not yet public, so treat anything beyond the names and rough window as unconfirmed. What the announcement does establish is intent: NVIDIA is planning the local-AI desktop category out to the end of the decade, not treating Spark as a one-off.

What It Means for Buyers

A published roadmap is useful precisely because it sets expectations. If your work genuinely needs the local compute, the Grace Blackwell generation is the one you can actually buy into this year, and it will stay capable for years regardless of what follows. If you are merely curious, knowing that Vera Rubin and its LPDDR6 jump sit a generation away helps you decide whether to buy now or wait. The broader AI PC range at Evetech covers machines built for local model work today, and the current GPU options in the GPU best sellers cover the cards most local-AI builders reach for while the Spark line matures.

Frequently Asked Questions

What comes after Grace Blackwell Spark?

Vera Rubin Spark, expected around 2027 to 2028, followed by Rosa Feynman Spark around 2029 to 2030. Each generation changes both the CPU and the GPU architecture together.

What is the big change in Vera Rubin Spark?

The memory move from LPDDR5X to LPDDR6, alongside the architecture shift from Grace to Vera and Blackwell to Rubin. Higher memory bandwidth directly helps AI workloads that are limited by how fast data moves, not just raw compute.

Why are the generations named after scientists?

NVIDIA pairs a CPU codename and a GPU codename, both named after scientists, and the same scheme runs across its data-centre roadmap. Spark simply inherits that naming, so each Spark generation matches a data-centre architecture pairing.

Should I wait for the next Spark generation?

If you need local AI compute now, the Grace Blackwell generation is the buyable option and stays capable for years. Waiting only makes sense if your need is not urgent and the LPDDR6 bandwidth of the Vera Rubin generation specifically matters to your workload.