NVIDIA continues its push into the AI chip industry, developing components specifically for developers. At GTC Washington, they showed off their latest prototype component, the NVIDIA Vera Rubin Superchip.
As the name suggests, the Vera Rubin AI Superchip combines a newly developed Vera CPU with two Rubin AI graphics processing chips on a single motherboard to create a powerful superchip for machine learning and AI inference.
The Vera CPU features 88 physical ARM cores with 176 power threads and is connected to the Rubin GPU components using the NVIDIA C2C Interconnect with a data transfer capacity of 1.8TB/s.
The Rubin GPU components are also equipped with 288GB of HBM4 memory built into each GPU, and the combination of the two GPUs is reported to be capable of 50 PFLOPS using the FP4 data processing format.
According to NVIDIA, the combined Vera Rubin Superchip, called NVL144, can offer 3.6 Exaflops using the FP4 format and up to 1.2 Exaflops with the FP8 format, which is suitable for machine learning training.
In the meantime, they also confirmed that a single NVL144 superchip comes with 3.3 times greater inference performance than the previous GB300 NVL72.
The NVIDIA Vera Rubin superchip is currently being tested in NVIDIA laboratories, where they confirmed that mass production and sales of the component will begin in 2026.
