Embedding programmable logic devices within deep learning processors provides flexible reconfigurable support functions.
A reconfigurable processing element uses multiplexer circuitry to switch between weight stationary and output stationary modes in a systolic array.
Dynamic mode switching between row-major and column-major configurations maximizes processing element utilization in MxN systolic arrays.
A computing architecture with sequential memory locations processes data streams using dedicated processing elements.
A systolic array structure processes data reuse schemes across two dimensions to accelerate computation.