With the continuous advancement of artificial intelligence (AI), the field has evolved from traditional manual feature engineering to data-driven learning. This shift has led to significant breakthroughs in areas such as machine vision, speech recognition, and natural language processing. Among the many technologies driving this progress, Convolutional Neural Networks (CNNs) have gained widespread popularity. As one of the most influential architectures in deep learning, CNNs are especially powerful in image processing tasks.
As these networks grow in size and complexity, the demand for high-performance computing resources increases. To address this challenge, developers are turning to FPGA (Field-Programmable Gate Array) technology. FPGAs combine the flexibility of software with the efficiency of ASICs, offering high throughput and low latency. Additionally, their rich I/O interfaces make them ideal for protocol conversion and interface integration.
Recently, KORTIQ introduced AIScale, a CNN accelerator IP designed for Xilinx FPGAs. This solution allows users to take pre-trained CNN models like ResNet, AlexNet, Tiny Yolo, and VGG-16 and compress them into deployable binary files. These files can be implemented on any Xilinx programmable logic device. The Zynq SoC and Zynq UltraScale+ MPSoC devices can serve as the processing system (PS), feeding data to the AIScale accelerator (PL), which then processes and sends the results back.
The compressed CNN model uses minimal resources, making it suitable for on-chip memory deployment. This enables faster and more flexible switching between different CNN models.
*Figure 1: Schematic diagram of AIScale in computer vision applications*
At the heart of AIScale is the AIScale RCC (Reconfigurable Compute Core). Users can customize the number of RCC modules based on their specific needs. The core supports convolution preprocessing, pooling/sampling, weight operations, and fully connected layer processing. Devices like the Zynq SoC and UltraScale+ MPSoC, which offer more resources, can integrate multiple RCC modules, significantly boosting performance. Alternatively, systems can be optimized by selecting a suitable number of modules based on cost, power consumption, and performance requirements.
*Figure 2: Multiple AIScale RCC Module Cascading Connections*
KORTIQ currently focuses on embedded systems, computer vision, Industrial 4.0, and the Internet of Things (IoT). In the future, they plan to expand the capabilities of AIScale, including features like image classification, object detection, face recognition, speech recognition, and natural language processing. By integrating advanced AI models into automated production and control systems, KORTIQ aims to enhance productivity across various industries and deliver better services to end users.
Ni-Fe Battery 700~1200ah
Nife Battery,48V 800Ah Nife Battery,Ni-Fe Battery 700~1200Ah,Ni-Fe Battery 700~1200Ah
Henan Xintaihang Power Source Co.,Ltd , https://www.taihangbattery.com