New paper on Scalable Memory Interconnects for DNN Accelerators to appear at FPL 2018

A new paper, co-authored with Yongming Shen, Tianchu Ji, and Mike Ferdman has been accepted to appear at FPL2018.

A preprint is available here, and an extended version is available on arXiv.

Abstract—To cope with the increasing demand and computational intensity of deep neural networks (DNNs), industry and academia have turned to accelerator technologies. In particular, FPGAs have been shown to provide a good balance between performance and energy efficiency for accelerating DNNs. While significant research has focused on how to build efficient layer processors, the computational building blocks of DNN accelerators, relatively little attention has been paid to the on-chip interconnects that sit between the layer processors and the FPGA’s DRAM controller.

We observe a disparity between DNN accelerator interfaces, which tend to comprise many narrow ports, and FPGA DRAM controller interfaces, which tend to be wide buses. This mismatch causes traditional interconnects to consume significant FPGA resources. To address this problem, we designed Medusa: an optimized FPGA memory interconnect which transposes data in the interconnect fabric, tailoring the interconnect to the needs of DNN layer processors. Compared to a traditional FPGA interconnect, our design can reduce LUT and FF use by 4.7x and 6.0x, and improves frequency by 1.8x.

Citation: Yongming Shen, Tianchu Ji, Michael Ferdman, and Peter Milder. “Medusa: A Scalable Memory Interconnect for Many-Port DNN Accelerators and Wide DRAM Controller Interfaces.” To appear at the 28th International Conference on Field Programmable Logic and Applications (FPL), 2018.