At the GTC 2025 conference of NVIDIA in California, California, CEO Jensen Huang Disclosure The company is planning to several new AI-ACCLErating GPUS in the coming months and years. He also revealed more explanations about the previously announced chips.
The center’s announcement was Vera Rubin, First teased The computer is scheduled for release in 2024 and now in the second half of 2026. Famous astronomerTen Tambitis will display memory and it will come with a custom NVIDIA designed CPU called CPU called Vera.
According to NVIDIA, Vera Robin will improve performance compared to his predecessor, Grace BlackwellEspecially for AI training and diagnosis.
Explanations for Vera Rubin presented during his GTC 2025 key note.
Vera Rubin includes two GPUs on a dye, which provides 50 Petaflaps of FP4 Incration Performance per chip. When a complete NVL144 rack is formed, the system provides 3.6 Exphillus of the FP4 Influx 3.3 times more than the 1.1 -Exphms of Blackwell Ultra in a similar rack.
Vera CPU has 88 customs arm core, with 176 threads, including a high -speed 1.8 TB/SNV link interface from Rubin GPU.
Huang also announced Robin Ultra, which will follow it in the second half of 2027. Robin Ultra will use the NVL576 Rack Configure and feature a four -retail -sized individual GPUS, which will supply FP4 precision 100 Petaflops (AI models a 4 -bit floating point format).
At the rack level, the Robin Ultra FP4 will provide 15 Exphillus and 5 Exphms of FP8 Training Performance – which is four times more powerful than Robin NVL 144 sequences. Each Robin Ultra GPU will contain 1TB HBM4E memory, which contains a full rack 365TB fast memory.