Exploiting Reduced Precision for Machine Learning on FPGAs

Wednesday, May 23, 2:10 PM - 2:40 PM
Summit Track: 
Enabling Technologies
Location: 
Exhibit Hall A-3

Machine learning algorithms such as convolutional neural networks have become essential for embedded vision. Their implementation using floating-point computation requires significant compute and memory resources. Research over the last two years has shown that reducing the precision of the representations of network inference parameters, inputs and activation functions results in more efficient implementations with a minimal reduction in accuracy. With FPGAs, it is possible to customize hardware circuits to operate on these reduced precision formats: 16 bit, 8 bit and even lower precision. This significantly reduces the hardware cost and power consumption of inference engine implementations. In this talk, we will show detailed results of the accuracy and implementation cost for several reduced-precision neural networks on a set of embedded platforms. From these design points, we extract the pareto-optimal results for accuracy versus precision of both weights and activations, ranging from 16 bit to 8 bit, and down to only a few bits.

Speaker(s):

Kees Vissers

Distinguished Engineer, Xilinx

Kees Vissers graduated from Delft University in the Netherlands. He worked at Philips Research in Eindhoven, the Netherlands, for many years. The work included Digital Video system design, HW –SW co-design, VLIW processor design and dedicated video processors. He was a visiting industrial fellow at Carnegie Mellon University, where he worked on early High Level Synthesis tools. He was a visiting industrial fellow at UC Berkeley where he worked on several models of computation and dataflow computing. He was a director of architecture at Trimedia, and CTO at Chameleon Systems. For more than a decade he has been heading a team of researchers at Xilinx, including a significant part of the Xilinx European Laboratories. The research topics include next generation programming environments for processors and FPGA fabric, high-performance video systems, machine learning applications and architectures, wireless applications and new datacenter applications. He has been instrumental in the High-Level Synthesis technology.

See you at the Summit! May 20-23 in Santa Clara, California!
Register today and reserve your hotel room!