NVIDIA, Feynman GPU 탑재한 집적 광학 기술 5년 앞당겨 먼저 출시

NVIDIA’s Feynman GPUs will be the first to feature Co-Packaged Optics (CPO), marking a strategic pivot for the AI leader. CPO, also known as silicon photonics, is a next-generation architecture designed to reduce reliance on copper by harnessing light for signal transmission. By packaging optical components directly alongside hardware accelerators like GPUs, CPO delivers lower interconnect latency and establishes high-bandwidth pathways between CPUs and GPUs, making it a foundational technology for next-generation AI data centers.

Under previous roadmaps, CPO commercialization was targeted for 2033. However, NVIDIA has accelerated its optics timeline by five years, targeting 2028 with the launch of its Feynman GPUs. According to a report from Nikkei Xtech, as AI infrastructure scales, the physical distance between deployed platforms can exceed 10 kilometers, requiring data transmission speeds of several hundred gigabits per second or higher. Conventional copper-based interconnects struggle to meet these demands, necessitating a shift toward optical solutions as the primary data transfer medium.

This industry-wide transition is supported by the Optical Compute Interconnect Multi-Source Agreement (OCI-MSA), established in March and backed by major technology firms including NVIDIA, Broadcom, AMD, Meta, OpenAI, and Microsoft. As the market leader, NVIDIA will debut its first CPO-integrated solution within the Feynman GPU lineup in 2028.

During GTC 2026, NVIDIA confirmed that the Feynman architecture will incorporate 3D Die Stacking technology, potentially marking the company’s first deployment of vertically stacked GPU dies. Additionally, NVIDIA is expected to partner with Intel as a foundry provider, utilizing Intel’s advanced packaging capabilities, including Embedded Multi-die Interconnect Bridge (EMIB), to manufacture Feynman chips.

In memory subsystem developments, NVIDIA has shifted from standard next-generation High Bandwidth Memory (HBM) specifications to a custom HBM solution for the Feynman GPUs. While the preceding Rubin architecture utilizes HBM4 and Rubin Ultra relies on HBM4E, NVIDIA’s approach appears to be either an enhanced variant of HBM4E or a proprietary HBM5 configuration, differentiating it from standard HBM5 implementations.

NVIDIA also revealed the architecture name for its next-generation data center CPU. Rather than relying on the previously rumored Vera platform, Feynman will integrate a new processor codenamed Rosa, named after American physicist and Nobel laureate Rosalyn Sussman. While specific technical specifications remain undisclosed, NVIDIA’s historical performance trajectory suggests significant architectural advancements.

Alongside the Feynman rollout, NVIDIA will maintain a comprehensive portfolio of AI platform components, including BlueField-5, NVLink 8 CPO, Spectrum 7 204T, CPO, and CX10. The Rosa CPU and Feynman GPU solutions are both scheduled to arrive in 2028. Meanwhile, AMD is reportedly developing its own Co-Packaged Optics architecture in collaboration with Global Foundries, with initial integration anticipated across its MI500 GPUs on a comparable 2028 timeline.