What's New

Feature Article

Feature Article

A Domain-Specific Architecture for Deep Neural Networks

N. Jouppi, et al., Google

Communications of the ACM, September 2018

Google engineers apply new Tensor Processing Unit (TPU) chips to neural network applications, up to 30 times faster and up to 80 times more energy efficient than competing chips.

Read the article here.

These improvements have been applied to real-time data center applications of deep neural networks, such as image recognition, language translation, and search. The applications are designed using open-source TensorFlow software.

A brief video of some of the authors discussing this work is also available here.

Technology Spotlight

Technology Spotlight

The Accelerator Age

At the recent DARPA ERI Summit, Bill Dally of NVIDIA spoke about machine learning applications of new GPU chips, enabling acceleration of performance based on computer architecture, rather than faster devices. See here for the video of his presentation here.

Slides for the talk are available here (PDF, 2 MB).

He provided an example whereby a speedup to 15,000 was enabled by co-design of algorithms and hardware (memory distribution) on a Volta GPU chip. He also described how DARPA sponsorship and teaming with university researchers is promoting the further development of this technology.

The agenda for the ERI Summit is available here, with links to many other presentations. Other videos from the Summit are available at DARPAtv.