NVIDIA CEO Huang Renxun announced the next artificial intelligence (AI) chip architecture "Rubin" in COMPUTEX 2024, as an iteration of the "Blackwell" architecture just released in March this year. The new Blackwell series is still in production and is expected to be officially shipped in late 2024, while the next generation Rubin architecture is expected to be officially launched in 2026.
Huang Renxun stated that Nvidia has promised to launch new AI chips at a "generation by generation" pace, with a faster update frequency compared to the previous two generations, highlighting Nvidia's efforts to maintain a leading position in the fiercely competitive AI chip market.
According to Huang Renxun, the Blackwell Ultra product will be launched in 2025, the first generation Rubin product will be launched in 2026, and the Rubin Ultra will be launched in 2027.
The NVIDIA Rubin architecture will support 8-layer HBM4 high bandwidth storage for the first time, while the Rubin Ultra will support 12 layer HBM4. Meanwhile, Nvidia also showcased a CPU codenamed "Vera", which will be launched simultaneously with the Rubin GPU to form the Vera Rubin superchip, replacing the current Grace Hopper.
It is reported that other prominent features of the Rubin platform include the next-generation NVLink 6 Switch, which can reach up to 3600 GB/s, and the CX9 SuperNIC component, which can reach up to 1600 GB/s.
