Analog In-Memory MAC Processing Using Mirror Currents

The patent replaces the digital computer architecture with an analog neural network architecture where continuous voltage signals directly represent and process data. The neural network performs MAC operations through analog voltage multiplication and addition, eliminating the need for discrete digital processing steps and thereby significantly improving processing efficiency for deep learning applications.

The patent changes the fundamental parameter representation from discrete digital values to continuous analog voltage values. By using continuous voltage signals to represent data and weights, the system achieves parallel processing of multiple MAC operations simultaneously through the physical properties of electrical circuits, thereby enhancing productivity without requiring complex digital logic.

The patent merges the memory storage function with the processing computation function into a unified neural network architecture. The same analog circuitry that stores data as voltage levels also performs the MAC operations, eliminating the need for separate memory access steps and thereby achieving in-memory processing that simultaneously improves speed and reduces architectural complexity.

The neural network performs self-processing by directly manipulating voltage signals within its own circuitry. The analog circuits automatically execute multiplication and addition operations through their physical characteristics, without requiring external memory access or complex control logic, thereby achieving fast in-memory processing with simplified architecture.

The patent uses voltage dividers and current mirrors to create precise copies of voltage and current signals. By replicating the input voltage and weight current through controlled resistor networks, the system maintains measurement precision while enabling parallel analog processing. The voltage divider circuits accurately proportion voltages and the current mirror circuits precisely copy current levels, ensuring high precision in analog MAC operations.

The patent introduces voltage dividers and current mirrors as intermediary circuits between the input signals and the final output. These intermediary components serve as precision measurement and control elements that accurately represent the input values while enabling the analog computation. The voltage divider acts as an intermediary to precisely scale voltages, and the current mirror acts as an intermediary to precisely replicate currents, thereby maintaining measurement precision in the analog processing path.