Adaptive error control for nanometer scale network-on-chip links

The authors present an adaptive error control method for switch-to-switch links in nanoscale networks-on-chip to manage reliability, throughput and energy. Unlike previous works, the proposed method adjusts both error detection and correction simultaneously at runtime. For a given application or predicted noise scenario, an appropriate error control scheme is selected for reliable message transmission. When link conditions degrade, more powerful error detection and correction are temporarily provided to recover the previous message. To achieve this adaptation, the authors create a configurable M-error correction, 2M-error detection code, combined with a hybrid automatic repeat request retransmission policy. Simulation results show that the proposed method can reduce residual flit error rate by over three orders of magnitude and achieve up to 75% higher average throughput compared to other error control methods. Further, average energy per successfully transmitted flit is reduced by up to 15% compared to fixed error control in a 65-nm technology. Compared to a recent adaptive error detection method, a 34% energy reduction can be achieved in high noise environment, at the expense of moderate area overhead.