Decision Mechanics

Insight. Applied.

Google reveals performance of its custom AI chips

By

chart showing relative efficiency of different kinds of processing units

Google has just release performance data for its Tensor Processing Unit (TPU) custom machine learning chips.

TPUs are

  • 15-20x faster than GPUs and CPUs when running deep learning applications
  • 30-80x more power efficient (operations per Watt)
  • frugal in terms of lines of code required to control them (their deep learning applications are implemented using 100-1500 lines)

Google claim that without these chips they would have to double the number of data centers they run.

Parsey McParseface

By

Personally I think that the UK’s National Environment Research Council (NERC) missed a trick in ignoring the results of their own public poll when naming their new polar research vessel.

The poll results where overwhelmingly in favour of naming the vessel RRS Boaty McBoatface. Unfortunately, ministers went for the much safer fourth-placed option of naming it after beloved nature broadcaster Sir David Attenborough.

In terms of engaging the public with science, every time RRS Boaty McBoatface was mentioned by the media people would have paid attention. Instead the message is “So much for democracy.”—reinforcing the opinion many people hold about politics today.

However, NERC’s loss is Google’s gain. They’ve named a component of their TensorFlow open source machine intelligence framework “Parsey McParseface”. Parsey McParseface is a pre-trained parser for English text.

I’m not likely to forget that now, am I?