Bilkent University
Department of Computer Engineering


Reuse-Centric CNN Accelerator for Efficient Inferences


Nihat Mert Çiçek
MS Student
Computer Engineering Department
Bilkent University

Reuse-centric CNN acceleration speeds up CNN inference by reusing computations for similar neuron vectors in CNN’s input layers or activation maps. This new paradigm of optimizations is however largely limited by the large overhead in neuron vector similarity detection, an important step in reuse-centric CNN. This presentation will present the in-depth exploration on architectural support of reuse-centric CNN. It proposes an accelerator, which reduces the overhead of neuron vector similarity detection by several orders of magnitude and boosts the speed and reduce the energy consumption of reuse-centric CNN inference by a number of factors. Chisel hardware construction language was used to implement that accelerator as a parameterizable generator, allowing us to set network and memory subsystem related parameters. Design exploration was performed through synthesis and implementation using Xilinx’s Vivado tool.


DATE: 25 November 2019, Monday @ 16:40