Adder with Reduced Latency and Minimized Interconnect for Streaming Inner Products

Addition of streaming inner products is fundamental to computation intensive applications. Particularly, adders form the bottleneck of convolution operations in deep neural networks in terms of latency. In this regards, bit-serial generation of most significant digits first is exploited. However, separation cycles are introduced to conventional online adders for separating the computation of successive inputs. Consequently, the throughput for addition of inner products is bounded by the separation cycles. In this work, we introduce pipeline architecture for online adder (POA) for reduced latency. The proposed POA is evaluated with existing works on basis of execution cycles, hardware constraints and synthesis profile. For addition of sixteen streaming inner products, upto 75% reduction in execution cycles is achieved with POA as compared with conventional online adder.