With the continuous development of artificial intelligence (AI), it has evolved from early artificial feature engineering to now learning from massive data, and major breakthroughs have been made in the fields of machine vision, speech recognition and natural language processing. CNN (Convolutional Neural Network) is more and more popular in the field of artificial intelligence. It is one of the most representative network structures in deep learning technology, especially in the field of image processing. . As networks become larger and more complex, we need a lot of computing resources to train them, so people are turning their attention to FPGA (Field Programmable Gate Array) devices, FPGAs. Not only has software programmability and flexibility, but also ASIC high throughput and low latency, and due to its rich I / O interface, FPGA is also very suitable for use as a protocol and interface conversion chip.
Recently, KORTIQ has launched a Xilinx FPGA CNN Accelerator IP, AIScale, which can utilize the well-trained CNN network, such as industry standard ResNet, AlexNet, Tiny Yolo and VGG-16, and compress them to output binary. Description files that can be deployed to Xilinx's full range of programmable logic devices. The Zynq SoC and Zynq UltraScale+ MPSoC device PS can provide data to the AIScale CNN Accelerator (PL), which is sorted to output data to the PS. The compressed CNN network occupies a relatively small amount of resources and can be deployed in on-chip memory to switch the CNN network more quickly and flexibly.
Figure 1: Schematic diagram of AIScale in computer vision applications
The core of the AIScale accelerator is AIScale RCC (Re-configurable Compute Core). Users can flexibly customize the number of AIScale RCC modules according to their requirements. AIScale RCC supports convolution preprocessing, pooling/sampling, weighting and full connection layer processing. The more resource-rich Zynq SoC and UltraScale+ MPSoC can integrate more AIScale RCC modules, which will give the AIScale accelerator even greater performance gains. Of course, it is also possible to integrate certain AIScale RCC modules according to cost, system power consumption, and performance requirements, and select a suitable Xilinx FPGA device.
Figure 2: Multiple AIScale RCC Module Cascading Connections
KORTIQ currently focuses on embedded and computer vision, Design Industry 4.0 and Internet of Things (IoT), etc., and will provide more advanced features such as image classification, object recognition and tracking, face and AIScale CNN accelerators in the future. Speech recognition, natural language processing, etc., the application of advanced artificial intelligence networks to automated production, control and other scenarios to improve the productivity of related industries to bring better services to users.
Longkou Libo Insulating Material Co.,Ltd. , https://www.liboinsulation.com