NVIDIA is hosting its annual GPU Technology Conference (GTC 2022) this week, where we should be introduced to their next-gen Hopper GPU architecture.
We should see NVIDIA introduce a next-gen MCM-based (multi-chip module) GPU with the flagship GH100, which should be a monster GPU with up to 140 billion transistors -- yeah, that's 140,000,000,000 transistor -- around 2.6x times the largest GA100 GPU from NVIDIA.
The full GH100 GPU should have 140 billion transistors, 144 Streaming Multiprocessors, and an insane 18,432 CUDA cores. The dual GPU chiplet design would have 2 x 144 SMs (288 SMs) and a mind-boggling 36,864 CUDA cores. It would be absolutely bonkers, and the 1000W+ power figures in rumors of the last few months come into play... and man, I can't wait. Bring it on, NVIDIA.
- Read more: NVIDIA's next-gen GPU after Hopper: Blackwell GPU aka Ampere Next Next
- Read more: NVIDIA Hopper GPU: bonkers 140 billion transistors on next-gen AI chip
- Read more: NVIDIA's next-gen GH100 Hopper GPU: 1000mm2 and monster 1000W+ power
- Read more: NVIDIA 'GPU-N': GH100 GPU, up to 233GB HBM2e at 6.3TB/sec (!!!)
NVIDIA's new Hopper GPU architecture whether it be in monolithic or MCM GPU design, will be going up against AMD's next-gen Aldebaran GPU -- a new MCM-based GPU with up to 128GB of HBM2e memory and on TSMC 6nm -- as well as Intel's upcoming Ponte Vecchio GPU (which will use an Intel 7 base GPU tile, TSMC N7 for Xe-Link, and TSMC N5 for the compute tile) with up to 128GB of HBM2e memory.