Machine learning applications have different requirements at each level
Zynq SoC demonstrates responsiveness with low latency

Machine learning technology, which remained stagnant over nearly 70 years starting from 1950, has significantly expanded in the past 2-3 years. It is now being used in diverse applications including translation, ADAS, autonomous vehicles, image classification, robotics, and hearing aids.

Machine learning is a field of artificial intelligence that predicts the future by analyzing vast big data such as the generation volume, cycle, and format of data. Through the combination of technologies such as deep learning, neural networks, and natural language processing, it is expected to converge into various fields including not only physical devices but also intelligent apps and mesh devices.

Different applications using machine learning have different requirements in cost, power consumption, performance, and efficiency. For example, ADAS requires high accuracy and low power consumption, while hearing aids prioritize power and latency, using smaller networks.

Xilinx enabled machine learning with Zynq All Programmable Logic to train neural networks using popular frameworks including Caffe. By configuring an ARM-based software scheduler, it generates a .prototxt file to drive a pre-optimized custom neural network (CNN) inference accelerator on the programmable logic.

Xilinx launched the SDSoC development environment based on C, C++, and OpenCL languages to reduce design time and reduce dependency on hardware experts, but it still fell short of widespread adoption and deployment as well as addressing the complexity of machine learning. To supplement this, the reVISION stack includes various development resources for platform, algorithm, and application development. It also supports the most commonly used neural networks such as AlexNet, GoogLeNet, SqueezeNet, SSD, and FCN.

Zynq SoC can process algorithms intended to be implemented in parallel. From sensors to inference control, it demonstrates superior responsiveness with low latency. When compared in performance with embedded GPUs, it showed 6x better images/second/watt in machine learning. Additionally, it demonstrated 42x higher frames per second in computer vision processing. In real-time applications, latency is the most critical factor, and in this regard, it showed 1/5 level performance.

The advantage of this response time can be found in the fundamental architecture of Zynq SoC. Embedded GPUs frequently need to access external memory from sensors through vision, machine learning, and control processing. In contrast, Zynq SoC achieves consistent response time through dataflow batches implemented with programmable logic and significantly increased internal memory.

Jung Woong, DSP Specialist at Xilinx, stated, "When used in machine learning, GPUs are commonly used for training and Zynq has advantages in inference," and explained, "In training, FPGAs are used to accelerate some functions, while Zynq SoC is used for system implementation in embedded form at the edge. Since training itself is performed at the cloud level rather than the edge, FPGAs are used in large-scale systems and GPUs are used at the general PC level."