At the NVIDIA (NVIDIA) annual conference GTC October 2025 in Washington, the United States, NVIDIA CEO Jensen Huang unveiled the next-generation Vera Rubin superchip (Superchip) for the first time, showing that the motherboard integrates a Vera CPU and two huge Rubin GPUs, up to 32 LPDDR memory slots, and HBM4 high-bandwidth memory.
Huang Renxun said that Rubin GPU has returned to the laboratory for testing, and this is the first batch of samples produced by TSMC. Each GPU has 8 HBM4 interfaces and two GPU core chips (Reticle-sized dies) that are the same size as the reticle. In addition, Vera CPU is equipped with 88 customized Arm architecture cores and can support up to 176 threads.
According to NVIDIA’s plan, Rubin GPU is expected to enter mass production in the third or fourth quarter of 2026, which is roughly equivalent to or earlier than the current Blackwell Ultra’s GB300 Superchip platform being fully mass-produced.
In addition, NVIDIA’s Vera Rubin NVL144 platform will use a combination of two new chips. The Rubin GPU is composed of two Reticle-sized cores, has an FP4 precision computing power of 50 PFLOPS, and is equipped with 288 GB HBM4 high-bandwidth memory. The matching Vera CPU provides 88 customized Arm cores, 176 threads, and the NVLINK-C2C interconnect bandwidth can reach 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 computing power, which is approximately 3.3 times higher than the GB300 NVL72. The system has a total memory bandwidth of 13 TB/s and a fast storage capacity of 75 TB, which are 60% higher than the previous generation. It also has double NVLINK and CX9 communication capabilities, with maximum speeds of 260 TB/s and 28.8 TB/s respectively.
NVIDIA also plans to launch a higher-end Rubin Ultra NVL576 platform in the second half of 2027. The system will expand the NVL scale from 144 to 576, the CPU architecture will remain unchanged, and the GPU will be upgraded to four Reticle-sized cores, with FP4 precision computing power of up to 100 PFLOPS, and equipped with 1 TB HBM4e high-bandwidth memory.
In terms of performance, the Rubin Ultra NVL576 platform can achieve 15 Exaflops of FP4 inference and 5 Exaflops of FP8 training computing power, which is 14 times higher than the GB300 NVL72. Its HBM4 high-bandwidth memory has a bandwidth of 4.6 PB/s and a fast storage capacity of 365 TB, which is 8 times that of the previous generation. The NVLINK and CX9 communication capabilities are increased to 12 times and 8 times, and the maximum speed reaches 1.5 PB/s and 115.2 TB/s respectively.