Announced last year during Google I/O 2017Second-generation TPU acceleratorLater, Google also announced the launch of its third-generation accelerator TPU 2018 at Google I/O 3.0, hoping to promote higher-level artificial intelligence technology computing.
However, compared to last year when some time was spent explaining the second-generation TPU, this year's keynote speech only briefly introduced it, perhaps meaning that the changes in the basic structure of the third-generation TPU are not as obvious as those of the second-generation product.
However, the third-generation TPU boasts a claimed 100 PFLOPS computing performance, nearly eight times that of the second-generation product. This is likely due to the fact that the heat generated during operation cannot be reduced through standard air cooling, necessitating the introduction of a water cooling system to dissipate the heat.
Judging from the previous TPU design, it is still mainly positioned as a TensorFlow accelerator component, so the system still needs to be configured with a mixed use of CPU and GPU.
