NVIDIA Rubin Architecture

Rubin GPUs are built on TSMC’s 3 nm process, use HBM4 memory and are slated for mass production in late 2025 with availability in early 2026; they deliver roughly 50 petaflops of FP4 compute (versus 20 petaflops in Blackwell) and about 1.2 exaflops of FP8 training performance, thanks to enhanced NVLink and memory bandwidth