Cerebras CS-4: 30x Faster Than GPUs
cerebras.ai
2 threads
Just some quick math, the video says each rack contains 3 wafer scale chips. Each chip is 4 trillion transistors. In total, there are 12 trillion transistors for compute inside each rack.
A Vera Rubin rack contains 72 Rubin GPUs, containing 2 dies for each GPU, and have an estimated 24 trillion transistors for compute.
So a Vera Rubin rack has 2x more transistors for compute. Not only that but each Vera Rubin rack contains 36 Vera CPUs totaling 3,168 CPU cores and an additional 8 trillion transistors.
It seems like an Nvidia rack contains vastly more transistors - not counting memory chips.