It is 1.2-1.7 times faster than the supercomputing built by the Nvidia A100, and Google says TPUv4 can better meet the needs of generative AI.
Thanks to CTOnews.com netizen Xiao Zhan cut the clue delivery! CTOnews.com April 6 news, Google researchers recently published a paper, want to prove that they have enough strength to create a supercomputer, to meet the increasing demand for generative AI applications.
CTOnews.com learned from the paper [PDF] that supercomputers built with Google TPUv4 are 1.2-1.7 times faster and consume 1.3-1.9 times less power than supercomputers built with Nvidia A100.
At present, NVIDIA dominates the field of artificial intelligence computing, with more than 90% of artificial intelligence project development using its chips. Google is trying to compete with Nvidia in this market. TPU is called Tensor Processing Units, which is Google's own chip.
TPU v4 has 2.1 times higher performance than TPU v3 and 2.7 times higher performance/power consumption. The TPU v4-based supercomputer has 4096 chips and is about 10 times faster overall.