Architectural Trade-Off Analysis for Accelerating LSTM Network Using Radix-r OBC Scheme

This paper presents architectural trade-off analysis for accelerating two (Type I, II) fixed-point long short-term memory (LSTM) network based on circulant matrix-vector multiplications (MVMs) using radix- $r$ offset binary coding (OBC) scheme. Type I MVM architecture rotates the weights with the proposed modulo-cum interleaver and uses partial product generators (PPGs) with a single generation unit across a column. It is hardware-optimized using a single adder tree through time-multiplexing. Meanwhile, Type II MVM architecture rotates the inputs with the proposed store-cum interleaver and uses single PPGs with a single generation unit across a row. It is time-optimized by unfolding shift-accumulate unit to a shift-add tree followed by pipelining. A new design for element-wise multiplication using radix- $r$ PPG is also presented. Both the designs are extended to their block-circulant variants for certain accuracy requirements. Post-synthesis of Type I and II architectures for a different model, kernel, radix sizes and clock frequencies result in several efficient designs. Compared with the prior scheme, Type I architecture for $128 \times 128$ with $r=2$ on 28 nm FDSOI technology at 800 MHz occupies 32.27% lesser area, consumes 67.89% lesser power at the same throughput, while Type II architecture at the expense of area and power provides $40\times $ higher throughput.