Jensen Huang explained that the Rubin GPUs have returned to the lab for testing and that these are the first samples produced by TSMC. Each GPU has eight HBM4 interfaces and two reticle-sized GPU cores (the same size as the photomask). In addition, the Vera CPU is equipped with 88 custom Arm architecture cores, supporting up to 176 threads.
The Rubin GPU is expected to enter mass production in Q3 or Q4 of 2026 — roughly the same or even earlier than the full production of the current Blackwell Ultra GB300 Superchip platform.
The NVIDIA Vera Rubin NVL144 platform will use a combination of the two new chips: the Rubin GPU comprises two reticle-sized cores, delivering 50 PFLOPS of FP4 compute and paired with 288 GB of HBM4 high-bandwidth memory. The accompanying Vera CPU offers 88 custom Arm cores and 176 threads, with NVLINK-C2C interconnect bandwidth reaching 1.8 TB/s.
In terms of performance, the Vera Rubin NVL144 platform can achieve 3.6 Exaflops of FP4 inference and 1.2 Exaflops of FP8 training compute—approximately a 3.3× increase over the GB300 NVL72 platform.
NVIDIA plans to launch the higher-end Rubin Ultra NVL576 platform in the second half of 2027, raising performance to 15 Exaflops.