arstechnica.com
Qualcomm’s Hexagon NPUs evolved from their digital signal processors (DSPs), initially used for audio and modem signal processing. As AI advanced, DSPs adapted for parallel processing tasks like LSTMs and CNNs, focusing on matrix functions crucial for generative AI. While NPUs share architectural roots with DSPs, they’re optimized for parallelism, transformer processing, and handling large parameter sets. NPUs aren’t strictly required for edge AI; CPUs can handle lighter workloads efficiently, while GPUs excel at data-intensive tasks, even if they consume more power. Choosing the right processor depends on the specific application and power constraints.
