Artificial Intelligence (AI), Machine Learning (ML) and the Internet of Things (IoT) are driving growth in edge computing applications, requiring increasingly energy-efficient solutions to deliver higher computational performance in increasingly confined spaces, with small form factors and severe thermal constraints. Through its Smart Embedded VisionMicrochip is addressing the growing need for energy-efficient inference in Edge applications by making it easier for software developers to implement their algorithms in PolarFire® field-programmable gate arrays (FPGAs). Specifically, with the VectorBlox Accelerator Software Development Kit (SDK), Microchip is helping them leverage PolarFire FPGAs to create flexible , low-power overlay-based neural network applications without having to learn FPGA tool flows. "For software developers to benefit from the power efficiency of FPGAs, we need to remove the limitation of having to learn new Fpga architectures and proprietary flow tools, giving them the flexibility to port multi-framework and multi-network solutions," says Bruce Weyer, vice president of the Field Programmable Gate Array business unit at Microchip.
Programming neural networks without Fpga design experience
FPGAs are ideal for AI Edge applications, such as inference in power-constrained computing environments, because they can perform multiple giga operations per second (GOPS) more efficiently than a central processing unit (Cpu) or graphics processing unit (Gpu), but require specific hardware design skills. Microchip's VectorBlox Accelerator SDK is designed to allow developers to code in C/C++ and program energy-efficient neural networks without prior FPGA design experience.
The highly flexible toolkit can run models in TensorFlow and in the Open Neural Network Exchange (ONNX) format, which offers the widest interoperability of frameworks. ONNX supports many frameworks, such as Caffe2, MXNet, PyTorch and MATLAB®. Unlike competing Fpga solutions, Microchip's VectorBlox Accelerator SDK is supported by Linux® and Windows® operating systems and also includes a bit-accurate simulator that gives the user the opportunity to validate the accuracy of the hardware remaining in the software environment. The neural network IP included in the kit also supports the ability to load different network models at runtime.
"Microchip's VectorBlox Accelerator SDK and neural network IP core will offer both software and hardware developers a way to implement an extremely flexible convolutional neural network architecture on PolarFire FPGAs, from which they can then more easily build and deploy their Edge AI-enabled systems with the best form factors, and thermal and power characteristics accessible today," Weyerpoints out.
Using Edge Inference, PolarFire Fpga offers up to 50 per cent lower total power consumption than competing devices, while offering 25 per cent higher capacity math blocks, capable of delivering up to 1.5 tera operations per second (TOPS). Using Fpga, developers also have greater opportunities for customisation and differentiation through the inherent ability to upgrade devices and the ability to integrate functions on a single chip.
Microchip and Smart Embedded Vision
Microchip's Smart Embedded Vision initiative was launched last July to provide hardware and software developers with tools, intellectual property (IP) cores and boards to meet the thermal limit and small form factor requirements of Edge applications. Because PolarFire Fpga offer lower power consumption than other solutions, customers can eliminate the need for ventilation systems from their enclosures. In addition, integration for customers' custom designs is more functional in these FPGAs: for example, in a smart camera, PolarFire FPGAs can integrate the image signal pipeline, which then includes the sensor interface, DDR controller, image signal processing (ISP) IP and network interfaces, all while integrating Machine Learning Inference.
