PIM MAC Operator Using FP-to-Fixed Accumulation for Deep Learning

The patent merges memory and processor functions into a single integrated device. The memory device includes memory cells for storage and MAC operators for arithmetic operations, eliminating the need for separate memory and processor components. This integration allows data to be processed directly within the memory device, removing communication bottlenecks while maintaining manufacturing simplicity through a unified semiconductor structure.

The MAC operator performs multiplication and accumulation operations directly within the memory device using the stored data itself, without requiring external processor intervention. The memory cells hold input data, weights, and partial products, and the MAC operator uses these stored values to compute results, enabling the system to serve its own computation needs internally and reducing the exponential growth of external computation requirements.

The floating-point multiplication is segmented into distinct operational phases: exponent addition, mantissa multiplication, and result normalization. The MAC operator processes these segments sequentially, with each stage handling a specific aspect of the calculation. This segmentation allows the use of simpler circuitry for each individual stage while maintaining overall floating-point precision, rather than requiring a single complex multiplier circuit.

The MAC operator acts as an intermediary component within the memory device that bridges the gap between stored data and computation results. It receives input data and weights from memory cells, performs arithmetic operations using these stored values, and directly writes results back to memory cells. This intermediary function eliminates the need for data to travel to and from external processors, maintaining simple hardware structure while dramatically improving data processing speed.