Nvidia Feynman GPU는 3D die stacking 및 사용자 정의 HBM을 특징으로 하며, 차세대 Rosa CPU를 발표했습니다.

NVIDIA has revealed additional details regarding its next-generation AI Data Center solution, Rosa Feynman, which incorporates 3D Die-Stacking technology. The Feynman GPU architecture was initially confirmed at GTC 2025, when NVIDIA announced the Feynman GPU alongside next-gen HBM, Vera CPU, and several connectivity chips forming the foundation of the AI data center.

At GTC 2026, NVIDIA unveiled significant new information and modifications to its plans. Most notably, NVIDIA confirmed that Feynman GPUs will adopt 3D Die Stacking technology, marking what could be the first use of 3D stacked GPU dies from the company. To manufacture Feynman chips, NVIDIA will utilize Intel as a Foundry partner and leverage their advanced packaging technologies, such as EMIB.

Rather than implementing next-gen HBM as originally planned, NVIDIA is now designing custom HBM technology for its Feynman GPUs. While Rubin leverages HBM4 and Rubin Ultra leverages HBM4E, NVIDIA's solution appears to be either a custom or heightened version of HBM4E or a custom HBM5 solution, distinguishing it from standard HBM5 offerings.

A key announcement concerns the next-generation Data Center CPU architecture, which NVIDIA has named Rosa rather than Vera. The CPU is named after American physicist and Nobel Prize winner Rosalyn Sussman. No specific technical details have been disclosed yet, though given NVIDIA's trajectory, significant improvements are anticipated. NVIDIA will continue rolling out a comprehensive spectrum of chips for its AI platforms, including BlueField-5, NVLink 8 CPO, Spectrum 7 204T, CPO, and CX10. The Rosa Feynman solutions are expected to arrive in 2028.