Neural Network Accelerator

The Neural-Network Accelerators (NACC) improves the inference performance of neural networks. The NACC data type is INT8, and supports im2col, convolution, depthwise convolution, average pool, max pool, fully connected, activation and matrix multiplication acceleration. Such as can accelerate ResNet, MobileNet, Alexnet, VGG16, YOLO and other structures of the network. This IP can be used as a accelerator for functions in NMSIS, the open-source TensorFlow Lite.

• Support 8-bit & 16-bit fixed-point data formats
• Support bus transmission data width 128bit or 512bit
• Supports computing power configurable from 64 GOPS to 4 TOPS
• Support local single-port SRAM cache, and the bus can independently access the SRAM function
• Support accelerating of convolution, rgb convolution, depthwise convolution, pooling, fully connected, elementwise, activation function in NMSIS
• Support accelerating of matrix multiplication function in NMSIS
• Support accelerating of im2col function in NMSIS
• Support IFM split width or height,kernel split
• Support Multicore mode
• It has shared SRAM to store IFM ,Temp data and OFM