An Energy-efficient Adaptive Sub-THz Wireless Interconnect with MIMO-Beamforming between Cores and DRAMs

Abstract

Future data-oriented computing requires intensive memory access with burden of I/O interconnection. Traditional wired interconnect has limited bandwidth and energy-efficiency with no feature of configurability. This paper introduces an adaptive sub-THz wireless interconnect between cores and main memory (DRAMs) using MIMO beamforming. To satisfy the requirement of energy efficiency and latency, a configuration problem for wireless channel management at memory controller is formulated to find the optimal value of the beamforming matrix. A low latency beamforming algorithm for sub-THz based interconnect is proposed. To meet the requirement of optimal power efficiency and nanosecond level beamforming latency, a gradient-free Mesh Adaptive Direct Search (MADS) based method is designed to find the sub-optimal solution for beamforming without knowing the channel information. The simulation shows that the proposed MADS-based algorithm converges rapidly and achieves average 90.3% optimal power efficiency when compared to the solution provided by gradient-based convex optimization algorithm. Moreover, we shows that wireless interconnect can provide better energy efficiency compared with the PCB trace wired counterparts under the same data rate.