NVIDIA announces the H100 Hopper GPU with a Transformer Engine and FP8

NVIDIA

NVIDIA announced Hopper architecture and the H100, an 80-billion-transistor GPU on TSMC's 4N process. The Transformer Engine uses FP8 precision to speed up transformer model training and inference by up to 6x, and the system delivers 32 petaflops of AI performance.

• H100 features Transformer Engine accelerating transformer networks up to 6x without accuracy loss
• FP8 precision enables 32 petaflops of AI performance in DGX H100 configuration
• Supports HBM3 memory (3TB/s bandwidth), PCIe Gen5, and nearly 5 terabytes per second external connectivity