Mixed-Signal Current-Summing Circuit for Low-Latency AI Inference

The patent replaces traditional digital circuitry with a mixed-signal system that uses analog current sources to perform weighted sum calculations. Current sources generate analog currents proportional to weight values, and these currents are summed directly in the analog domain, eliminating the need for digital multiplication and addition circuits. This substitution of analog physical quantities for digital computational operations reduces circuitry area while maintaining computational accuracy.

The patent changes the domain of computation from digital to mixed-signal by using current magnitude as a continuous parameter to represent weight values. Instead of storing weights as digital values in memory, the system uses programmable current sources where the current magnitude directly encodes the weight parameter. This parameter change enables direct analog computation of weighted sums, reducing the need for large digital memory circuits.

The patent replaces sequential digital computation with parallel analog computation. Multiple current sources simultaneously generate currents representing different weight values, and these currents are summed in real-time through analog circuitry. This parallel physical process occurs continuously without the sequential clock cycles required by digital systems, dramatically reducing inference time and latency.

The analog current-based computation allows for continuous operation without discrete sampling or clock cycles. The current sources continuously generate and sum currents representing the neural network computation, enabling real-time inference with no idle time between computational steps. This continuous action eliminates the time losses associated with digital system clocking and data transfer.

The patent substitutes energy-intensive digital logic operations with low-power analog current processing. The mixed-signal system uses current mirrors and analog summation circuits that consume significantly less power than digital multipliers, adders, and memory access circuits. The physical nature of current flow allows for direct computation without the switching and capacitive charging/discharging that dominates digital energy consumption.

The patent changes the computational domain to exploit the energy efficiency of analog parameter processing. By representing weight values as current magnitudes and performing computations through physical current flow and summation, the system avoids the high energy cost of digital bit manipulation. This parameter change to continuous analog values enables compute power with minimal energy consumption.