Product Launchgpu acceleratorsdie stackinghbm4envlink
Nvidia Reveals Feynman GPUs With Die Stacking
9.2
Nvidia at its GPU Technology Conference this week unveiled a multi-year data-center roadmap detailing 2027 Rubin Ultra accelerators and 2028 Feynman platforms. Rubin Ultra will include four compute chiplets and 1 TB of HBM4E enabling Kyber NVL144 racks, while Feynman adopts die stacking, custom C‑HBM4E, Rosa CPUs, LP40 LPUs and NVLink co‑packaged optics to scale racks to 576–1,152 GPUs.
Key Points
- 1Announces Rubin Ultra with four compute chiplets and 1 TB HBM4E memory for 2027 accelerators.
- 2Introduces Feynman architecture using die stacking and custom C‑HBM4E to boost memory capacity and bandwidth.
- 3Enables rack-scale designs with NVLink co‑packaged optics, scaling to 576–1,152 GPU packages for higher density.
Scoring Rationale
High novelty and broad industry impact; slight limitation from reporting-based details lacking full technical disclosure and timelines.
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
