Nvidia unveiled next-generation CPUs (Rosa) and stacked GPUs (Feynman) with advanced optical interconnects, alongside support for Groq LPUs with NVFP4.

Nvidia presented its updated data center product roadmap at its GPU Technology Conference, revealing plans to introduce a brand-new GPU architecture every couple of years and update its AI GPU family annually. The company plans to roll out its Vera Rubin platform this year, based on the Vera CPU and Rubin GPU, accompanied by five additional processors: the Groq LP30 low-latency inference accelerator, the BlueField-4 data processing unit, the NVLink-6 switch, Spectrum-X Ethernet with co-packaged optics, and the ConnectX 9 1600G SuperNIC. The Vera Rubin platform is notable for Nvidia's integration of Groq LPUs into its hardware portfolio, with the company favoring LPUs to the point that it no longer mentions its own Rubin CPX processors on the roadmap.

In 2027, Nvidia plans to update its offerings with Rubin Ultra AI accelerators, which will feature four compute chiplets and be equipped with 1 TB of HBM4E memory, dramatically increasing performance compared to this year's Rubin. These GPU accelerators will be paired with the Groq LP35 LPU, which will support the NVFP4 data format to improve performance. Nvidia will also introduce the Kyber NVL144 rack-scale solution, packing 144 Rubin Ultra GPU packages enabled by an NVLink 7 switch, offering at least 4X performance improvement compared to Oberon NVL72 racks with 72 Blackwell GPU packages.

The next generation, Feynman, arriving in 2028, will feature fundamental architectural improvements. As Jensen Huang, Nvidia's chief executive, stated at the GTC: "The next generation from here is Feynman. Feynman has a new GPU, of course; it also has a new LPU LP40 […] now uniting the scale of Nvidia and the Groq building together LP40, it is going to be incredible." The platform will include a brand-new CPU called Rosa, short for Rosalyn, along with BlueField-5, ConnectX 10, and both Kyber copper scale-up and Kyber CPO scale-up solutions. This represents Nvidia shortening its CPU development cycle from four years to two years, putting it on par with AMD and Intel.

Feynman's technical innovations center on multiple areas. The Feynman data center GPU will adopt die stacking, enabling a new way to scale performance, and will use custom high-bandwidth memory, most likely a variant of C-HBM4E, enabling Nvidia to boost HBM capacity beyond 1 TB per GPU package and increase memory bandwidth. The Feynman platforms will be powered by Rosa CPUs, Nvidia's next-generation in-house processors focused on ultimate single-thread performance. The platform will integrate the LP40 LPU, supporting Nvidia's NVFP4 format and connecting to other system components using the NVLink protocol, thereby integrating Groq hardware with Nvidia's GPUs.

For the first time, Feynman will adopt NVLink switches with co-packaged optics, enabling optical interconnections using the NVLink protocol and making them significantly easier and cheaper to implement. This will allow Nvidia to increase the scale-up world size of its rack-scale solutions to 576 GPU packages using Oberon chassis or even 1152 GPU packages using Kyber chassis, making the company's rack-scale systems more competitive against alternatives like AMD's Instinct or custom accelerators deployed by hyperscalers. Nvidia's data center portfolio will achieve quantitative improvements in 2027 through increased GPUs per rack and a new LPU with NVFP4 support, while 2028's all-new architectures will bring qualitative improvements to the company's products.