The Google TPU paper is out: https://cloudplatform.googleblog.com/2017/04/quantifying-the-performance-of-the-TPU-our-first-machine-learning-chip.html

As we have been predicting for 10 years, in SoC you can achieve > 10x more performance that current GPUs and > 100x more performance per watt.

All in 8-bits, because really, you do not need more than 8 bits for inference and the reason we did not use 8-bits more is because they do not give any advantage on current hardware!

Well done!
Shared publiclyView activity