eCommerceNews New Zealand - Technology news for digital commerce decision-makers
Story image
High demand for Google’s machine learning processor in NZ
Thu, 22nd Mar 2018
FYI, this story is more than a year old

Strong local demand for Google's new generation of Cloud TPUs (Tensor Processing Units) means Kiwi businesses and consumers will benefit from better services and greater productivity.

Google's Tensor Processing Units, released into public beta last month, are hardware accelerators optimised to speed up and scale up Machine learning (ML) workloads.

Google Cloud New Zealand's Brendan Bain says interest from New Zealand businesses shows there is real appetite locally to start leveraging the technology.

“The benefit to business is much faster training of Machine learning models, in some cases reducing from several weeks to several hours. That's like moving from basic dial-up to broadband overnight, and the productivity boost will be significant.

“Kiwi consumers can expect smarter, faster services across the board, from online shopping to old-fashioned big box retailing,” Bain says.

A large number of Kiwi businesses are already using Machine learning on a daily basis.

Trade Me, which uses it to improve its search function and make recommendations to users, is one business planning to use Google's TPUs.

Lead data scientist Lester Litchfield says they offer a lot of opportunities to improve the site.

“TPU's will help us iterate and explore ideas quicker, and present fresher models to our users. We'll be able to deliver better experiences faster than we can with traditional CPU and GPU powered Machine learning,” Litchfield says.

Why TPUs for machine learning?

Traditionally, writing programs for custom ASICs and supercomputers has required deeply specialised expertise.

By contrast, Cloud TPUs can be programmed with high-level TensorFlow APIs, and Google has open sourced a set of reference high-performance Cloud TPU model implementations to help users get started:

  • ResNet-50 and other popular models for image classification
  • Transformer for machine translation and language modeling
  • RetinaNet for object detection

Google continuously test these model implementations both for performance and for convergence to the expected accuracy on standard datasets. Over time, Google is planning to open source additional model implementations.

Adventurous ML experts may be able to optimise other TensorFlow models for Cloud TPUs on their own using the documentation and tools provided.