Scalable Matrix Multiplication With Hybrid CMOS-RSFQ Digital Signal Processor

We report an RSFQ digital signal processor design based on hybrid RSFQ-CMOS memory suitable for a general matrix-on-matrix multiplication algorithm. The DSP consists of an RSFQ multiply-accumulate unit, memory caches and synchronization block, partitioned into multiple chips, and a large CMOS memory. The parameters of the RSFQ DSP are a 10times10 bits multiplication with rounding to 14 bits, an 18-bit accumulator length and a 3.7 Kb memory cache. The maximum simulated clock frequency is equal to 24 GHz for HYPRES 4.5 kA/cm² process and optimum communication bandwidth with the CMOS memory is 2 Gbps. The simplified version of the RSFQ DSP consisting of a 4times4 MAC with rounding to 5 bits and 17times6 memory caches has been designed for HYPRES 4.5 kA/cm² process.