Nvidia has shared fresh details about its upcoming AI data‑center platform, Rosa Feynman, confirming that its next‑generation Feynman GPU architecture will use 3D die‑stacking technology and a custom high‑bandwidth memory (HBM) solution.
The company announced the shift during GTC 2026, revealing that Feynman GPUs will adopt advanced 3D stacking, which could make them the first from Nvidia to use fully stacked GPU dies. Another key update is the memory design. Instead of standard next‑gen HBM, NVIDIA’s Feynman architecture will employ custom HBM technology.
Alongside the GPU details, Nvidia also confirmed that the platform will use a new data center CPU architecture called “Rosa”, named after physicist Rosalyn Sussman. While specific CPU specifications haven’t been shared yet, the change marks a departure from earlier plans to pair Feynman with the Vera CPU used in prior generations. The broader ecosystem around Rosa Feynman is set to include a full suite of AI‑focused chips, such as BlueField‑5, NVLink 8 CPO, Spectrum 7 204T, and CX10, reinforcing NVIDIA’s push into large‑scale AI data‑center solutions.
According to NVIDIA’s roadmap, the Rosa Feynman lineup is expected to arrive in 2028.
