npu

A neural processing unit (NPU) is a low-power accelerator integrated into a phone, laptop, vehicle, or edge system-on-chip to run neural-network inference without sending every request to a data center. **NPUs center on quantized matrix engines.** Arrays of multiply-accumulate units execute INT8, INT4, and other compact tensor formats, supported by on-chip SRAM and dedicated data-movement hardware. This is far more energy-efficient than using a CPU for the same regular workload. **The design target is sustained efficiency.** Edge inference is constrained by battery, heat, latency, and memory rather than maximum floating-point throughput. Compilers partition supported graph operators onto the NPU and leave unsupported work on the CPU or GPU. **TOPS is not the whole story.** Real performance also depends on memory bandwidth, operator coverage, precision, sparsity support, compiler quality, and how often data crosses processor boundaries.