Google officially released during the Google I/O 2018 Developer Conference Third Generation AI Artificial Intelligence/Machine Learning Processor TPU 3.0.
TPU is the Tensor Processor Unit, Google's custom-built special chip (ASIC) for machine learning, designed specifically for Google's deep learning framework TensorFlow.
Compared to traditional GPU graphics chips, it uses 8-bit low-precision calculations to save transistors, has a small impact on accuracy but can significantly save power, speed up, as well as pulsating array design, optimizing matrix multiplication and convolution operations, and Use larger upper memory to reduce the dependence on system memory.
TPU 3.0's computing power is up to 100 PFlops (1000 trillion floating-point calculations per second), which is more than 8 times that of last year's second generation.
It also used liquid cooling for the first time , Can more effectively serve the data center, easy to customize the hardware program.
More technical details to be announced...
