Abstract: Digital signal processors (DSPs) are suitable for a wide variety of computationally intensive real-time applications. This paper describes the architectural features of DSPs for intelligence applications, and the node configuration of the IX-n general purpose neuro-computer, based on the commercially available DSP. DSPs provide high computing power by employing a high level of on-chip parallelism, integrated hardware multipliers, carefully tailored instruction sets, memory organization schemes, hardware support for loop execution, and specific sophisticated addressing modes. High-precisions control and fault-tolerance are achieved by exploiting the high-speed arithmetic, on-chip peripherals, direct memory access (DMA) controllers, multiprocessor support and bit manipulation capabilities of DSPs. Fast multiply/ accumulate time, integrated on-chip random access memory (RAM), large address space, high precision and multiprocessor support are necessary for efficient virtual implementation of neural networks. DSP architectural features make them applicable to both symbolic and connectionist AI models.
Keywords: DSP, Artificial intelligence, Neural network, CISC microprocessor, DSP architecture, FFT butterfly.