More machine learning – ScaledML

27 – 28 March 2019

The ScaledML conference is growing up; from a Saturday at Stanford to a two day event at the Computer History Museum with sponsors.

Two big new themes emerged

  • Concern for power efficiency (Simon Knowles, Graphcore, talked about Megawatts; Pete Warden, Tensorflow talked about milliwatts and energy harvesting
  • Development platforms – Adam D’Angelo, Quora, was particularly clear on how Quora operate development to efficiently support a small number of good developers

David Paterson gave the first talk on Domain Specific architectures for Neural Networks – an updated version of this talk

The roofline performance model is a useful way to visualize comparative performance. For future performance improvements functionally specific architectures are the way forward; this requires both hardware updates (what Google is doing with the TPUs) and improved compiler front and back ends.

Fig 3 from the Domain Specific Architectures paper linked above.

Intel recognizes this trend – Wei Li described the work his team is doing to incorporate domain specific support into Xeon processors. This blog post has the gist of what he presented.

Most of the talks are here on YouTube