At GTC 2025, Nvidia founder spoke about Rubin, Rubin Ultra and Feynman accelerators

The speech of the founder and permanent CEO of Nvidia Jensen Huang at the GTC 2025 technology conference lasted more than two and a half hours. The head of the company spoke about the company’s new products towards the end of his presentation, and for the first time the name of the successor to the Rubin family of accelerators was announced, which will enter the market in 2028.

Image source: NVIDIA

By that time, Nvidia’s newest accelerators will have switched to the Feynman architecture, obviously named after the American physicist Richard Feynman, who won the Nobel Prize and helped create the nuclear bomb in the mid-1940s. The characteristics and expected performance level of the Feynman-generation accelerators have not yet been given, but Nvidia mentions that they will use the next generation of HBM memory after HBM4E. Curiously, these accelerators will only be accompanied by chips with the Vera architecture, which will debut in 2026 or 2027 at the same time as the Rubin accelerators.

In fact, separating the first and last names of outstanding scientists is becoming the norm for Nvidia in the accepted system of conventional designations. Following Blackwell, solutions named after the American astronomer Vera Rubin should appear on the market. She will involuntarily share her last name with the computing accelerators that will be released in 2026 and 2027, and the central processors will inherit her name. Thus, the successors to the Vera processors will probably receive the designation Richard, but they are not mentioned in Nvidia’s plans yet. Vera processors will combine 88 cores with an Arm-compatible architecture of their own design and 176 threads.

Image source: NVIDIA

During his speech, Jensen Huang admitted that when designating the B200 accelerators of the Blackwell family, his company made a small miscalculation, using the combination NVL72. Since two crystals are located next to each other on one substrate, it would be more correct to call this arrangement NVL144. This will become the norm for similar representatives of the Rubin family.

The Vera Rubin NVL144 rackmount system will debut in the second half of next year, offering 3.3 times the performance of the GB300 NVL72 and 288GB of HBM4 memory. Memory bandwidth will increase from 8TB/s to 13TB/s compared to the B300. NVLink interfaces for data exchange within the system will also increase their bandwidth by approximately two times.

Image source: NVIDIA

Rubin Ultra accelerators and NVL576 rack-mount solutions based on them will be released in the second half of 2027. The performance in FP4 calculations will increase fourfold compared to Rubin NVL144. Rubin Ultra chips will combine four crystals in one package. The amount of RAM when switching from Rubin NVL144 to Rubin Ultra NVL576 will increase from 75 to 365 TB. Rubin Ultra accelerators will switch to using HBM4E memory, but each chip will have no more than 8 TB/s of bandwidth, but the memory capacity will increase to 1 TB. The NVLink7 interface in the Rubin Ultra generation will be accelerated sixfold to 1.5 PB/s, CX9 interconnects will be accelerated fourfold to 115.2 TB/s.