Nvidia accelerates co-packaged optics (CPO) deployment with Feynman GPU line, arriving five years ahead of prior roadmap.

NVIDIA's Feynman GPUs will mark the first implementation of Co-Packaged Optics, a significant departure from the original timeline. CPO, or Co-Packaged Optics, is a silicon photonics technology that reduces reliance on copper and uses light to transfer signals between hardware accelerators such as GPUs and CPUs. Originally slated for commercialization by 2033, NVIDIA has fast-forwarded its optics roadmap by five years, bringing the technology to market in 2028. According to a Nikkei Xtech report, this acceleration addresses the growing demands of AI infrastructure, where distances between installed platforms can reach north of 10km and data must be transmitted at several hundred Gigabits per second or more—a challenge that traditional copper-based methods struggle to meet.

The push toward Co-Packaged Optics reflects broader industry momentum. The Optical Compute Interconnect Multi-Source Agreement was established in March, bringing together major AI firms including NVIDIA, Broadcom, AMD, Meta, OpenAI, and Microsoft to advance the technology. NVIDIA, as the largest participant, is leading the charge with its Feynman GPU implementation in 2028.

At GTC 2026, NVIDIA revealed that Feynman GPUs will adopt 3D Die Stacking technology, marking the first use of 3D stacked GPU dies from the company. NVIDIA will leverage Intel as a foundry partner and utilize Intel's advanced packaging technologies, such as EMIB, to manufacture Feynman chips. Additionally, rather than adopting next-generation HBM in standard form, NVIDIA is deploying custom HBM technology for Feynman—a departure from Rubin's HBM4 and Rubin Ultra's HBM4E approach that suggests a custom or heightened HBM5 solution.

NVIDIA has also announced its next-generation data center CPU architecture, departing from the Vera platform in favor of a new design called Rosa, named after American physicist and Nobel Prize winner Rosalyn Sussman. While specific details remain undisclosed, the shift signals substantial improvements ahead. Beyond Feynman, NVIDIA will continue rolling out complementary components for its AI platform ecosystem, including BlueField-5, NVLink 8 CPO, Spectrum 7 204T CPO, and CX10, with Rosa Feynman solutions arriving in 2028. Concurrently, AMD is reportedly developing its own Co-Packaged Optics solution in partnership with Global Foundries, with first implementation expected in MI500 GPUs on a similar 2028 timeline.