Most Resource Efficient Matrix-Vector Multiplication on FPGAs

16. August 2023
Speaker: Ralf Müller
Date and Time: Wednesday August 16, 2023 at 2pm
Room: Seminar Room 389-2 (Room # CG 04 02)

Abstract: Fast and resource-efficient inference in artificial neural networks (ANNs) is of utmost importance and drives many new developments in the area of new hardware architectures. In this work, we present a novel method for lowering the computation effort for ANN inference utilizing ideas from information theory. Weight matrices are sliced into submatrices of logarithmic aspect ratios. These slices are then factorized. This reduces the number of required computations without compromising on fully parallel processing. We also provide a tool to map these sliced and factorized matrices efficiently to reconfigurable hardware. By comparing to the state of the art FPGA implementations, we lower hardware resources measured in look-up-tables (LUTs) by a factor of four to six. Our method does not rely on any particular property of the weight matrices of the ANN. It works for the general task of multiplying an input vector with a constant matrix and is also suitable for convolutional neural networks as well as digital signal processing beyond ANNs.

Bio: Ralf Müller received his diploma in electrical engineering in 1996 and his PhD in 1999 from the Friedrich-Alexander University of Erlangen. Then he spent a year as a postdoctoral researcher at Princeton University in New Jersey. From 2000 to 2004, he led a research group at the
Telecommunications Research Center Vienna (ftw.). In 2005, he was appointed full professor at the Department of Electronics and Telecommunications at the Norwegian University of Technology and Natural Sciences (NTNU) in Trondheim. In 2013, he moved to the Institute of Digital Transmission at Friedrich-Alexander University Erlangen-Nuremberg.

Ralf Müller received the Leonard G. Abraham Award (shared with Sergio Verdú) for the publication „Design and analysis of low-complexity interference mitigation on vector channels“ from the IEEE Communications Society. His dissertation „Power and bandwidth efficiency of multiuser system with random spreading“ was awarded both the Vodafone (then Mannesmann) Foundation Grant for Research in Mobile Communications and the ITG Grant. For the publication „A random matrix model for communication via antenna arrays“ he received the prize of the ITG. The Philipp-Reis Prize was awarded to him jointly with Robert Fischer. Prof. Müller was Associate Editor of the IEEE Transactions on Information Theory from 2003 to 2006.