AI: Machine Learning / Deep Neural Network
DeepRed: Machine Vision & DNN development Platform
DeepRed development platform is specially designed for industry grade deployment of Deep Neural Network using Xilinx ZYNQ SOC. Complex DNN design can be synthesized using IBM Power AI Inference Engine without the need of HDL coding.
The primary DNN framework supported is Caffe framework, the supported layers for HW accelerations includes: Convolutional Layer, Maxpool Layer, Fully Connected Layer, Batch Norm Layer and ReLU activation layer. Data Scientists need to quantize the pre-trained network using Caffe Ristritto and the HDL IP generation will be generated automatically using IBM Power AI Inference Engine.
Automation Tool -- AccDNN:
An end-to-end automation tool for generating convolutional neural network in FPGA without programming
This work has been presented in ICCAD’2018, San Diego, and wins the Best Paper Award.
Motivations and Features of AccDNN
1) An end-to-end automation tool which provides an integrated design flow from deep learning frameworks to FPGA board-level implementations.
2) A flexible support of quantization to address the limited resource issues, Our design supports flexible quantization for weights and activations either within a layer or across layers in DNN. It also supports binary and ternary networks.
3) A fine-grained layer-based pipeline architecture that can achieve high throughput even without batch processing.
4) An unified and flexible Processing Engine (PE) that provides a two dimensional parallelism scheme for implementing major layers in DNNs including convolutional layer and fully-connected layer.
5) An automatic resource allocation management scheme (A-REALM) that provides resource allocation across network layers based on the external memory access bandwidth, data reuse behaviours, computation resource availability, and network complexity
Layer based pipeline structure