CNN Plus Accelerator IP Core

Description

The Lattice Semiconductor CNN Plus Accelerator IP Core is a calculation engine for Deep Neural Network with fix point weights. It calculates full layers of neural network including convolution layer, pooling layer, batch normalization layer and full connect layer by executing sequence code with weight values which is generated by the Lattice SensAI Neural Network Compiler tool. The engine is optimized for convolutional neural network, which is why it can be used for vision-based application such as classification or face detection and key phrase detection. The engine is not a simple accelerator which requires an extra processor; it can do all required calculations by itself.

Features:

• Selectable 2 implementation types: Compact CNN and Optimized CNN
• Support for convolution layer, max pooling layer, global average pooling layer, batch normalization layer, and full connect layer
• Configurable bit width of activation (16/8-bit)
• Configurable number of memory blocks for tradeoff between resource and performance
• Optimized for 3x3 2D convolution calculation
• Dynamic support for various 1D convolution from 1 to 9 taps
• Supports 2x2 max pooling
• Supports global average pooling by full connect engine
• Configurable input byte mode (signed, unsigned, disable – Compact CNN only)
• Partial DRAM access
• Configurable maximum burst length (32, 256)

Devices Supported

LIFCL

References

• User Guide

Revision History