Residue number matrix multiplier
Arithmetic circuits and methods that perform efficient matrix multiplication for hardware acceleration of neural networks, machine learning, web search and other applications are disclosed herein. Various arrays of multiplier-accumulators may be coupled to form a matrix multiplier which processes data using high precision, fixed point residue number arithmetic.
1. A residue number matrix multiplier comprising:
a plurality of digit processing elements, each of the digit processing elements comprising a modular accumulator having one or more registers;
wherein the modular accumulator operates in a modular accumulator mode for an initial accumulation, and operates in a binary accumulator mode for a final accumulation.
2. The residue number matrix multiplier of claim 1 , wherein the initial accumulation stores one or more final product terms in the one or more registers, and results in a summation of one or more initial product terms of a matrix dot product.
3. The residue number matrix multiplier of claim 2 , wherein the final accumulation results in a summation of the one or more final product terms.
4. The residue number matrix multiplier of claim 1 , wherein the modular accumulator operates on a predefined modulus.
5. A residue number matrix multiplier comprising:
a plurality of digit processing elements, each of the digit processing elements comprising a modular accumulator associated with a modulus and responsive to a control input, wherein:
the modular accumulator operates in a modular accumulation mode for an initial accumulation responsive to assertion of a control input;
the modular accumulator operates in a binary accumulation mode for a final accumulation responsive to de-assertion of a control input and;
at least a subset of the modulus are unique;
wherein said modular accumulation is transferred from the modular accumulator to an output register before overflow of the modular accumulator.
6. The residue number matrix multiplier of claim 5 , wherein the modular accumulator performs operations on congruent operands.