Mixed signal computing-in-memory solver for integer constrained least square problems

A mixed signal computing circuit, comprising:

• a memory array having a plurality of memory cells arranged in rows and columns, wherein a first subset of rows is configured to receive analog input signals and a second subset of rows is configured to receive feedback signals;
• conversion circuitry configured to generate the analog input signals from digital input values;
• current accumulation circuitry coupled to columns of the memory array and configured to accumulate currents from activated memory cells;
• comparison circuitry configured to generate binary output states based on accumulated currents from the current accumulation circuitry;
• feedback generation circuitry configured to generate the feedback signals by modifying the binary output states with stochastic noise;
• and
• evaluation circuitry configured to compute a distance metric based on the binary output states and stored evaluation weights, wherein the evaluation circuitry comprises monotonic current generators that produce maximum current when positive and negative current contributions are balanced.

A method, comprising:

• receiving an input vector and a matrix defining an integer constrained least square (ICLS) problem;
• transforming the ICLS problem into a Hopfield network energy function;
• quantizing weights derived from the Hopfield network energy function into first quantized weights for optimization and second quantized weights for evaluation;
• programming the first quantized weights into a first memory array and the second quantized weights into a second memory array;
• iteratively performing analog current accumulation in the first memory array using the input vector and feedback signals to generate binary neuron states;
• computing distance scores using the second memory array and the binary neuron states;
• applying stochastic noise to the binary neuron states to generate the feedback signals;
• determining whether a convergence criterion is satisfied;
• and
• outputting an integer solution vector derived from the binary neuron states when the convergence criterion is satisfied.