Microchip introduces Software development kit and neural network IP

     With the rise of artificial intelligence, machine learning technologies, and the Internet of Things, applications are beginning to migrate to the edge of the network where data is collected. To reduce volume, reduce heat production, and improve computing performance, these edge applications require energy-efficient solutions. Microchip Technology Inc. 's intelligent embedded vision solution is designed to make it easier for software developers to execute algorithms inside PolarFire field programmable gate arrays (FPgas). And then meet the increasing demand for energy-saving inference function in edge applications. An important new addition to Microchip's embedded solution portfolio, the VectorBlox Accelerator Software Development Kit (SDK) helps software developers create flexible, low-power overlay neural network applications using Microchip PolarFire FPGas without having to learn the FPGA tool stream.

     Fpgas are ideal for edge AI applications, such as inference in power-constrained computing environments, because FPgas have higher computing power (GOPS) and consume more power than central processing units (cpus) and graphics processors (Gpus). However, FPgas require operators with specialized hardware design skills. Microchip's VectorBlox Accelerator SDK helps developers program energy-efficient neural networks in C/C++ without prior FPGA design experience.

     This highly flexible toolkit enables models to be executed in TensorFlow and Open Neural Network Exchange (ONNX) formats to maximize interoperability of the framework. ONNX supports many frameworks such as Caffe2, MXNet, PyTorch, and MATLAB®. Unlike other FPGA solutions, the VectorBlox Accelerator SDK is available on both Linux and Windows operating systems and includes an emulator with Bit-level accuracy. With simulators, users can verify the accuracy of hardware in a software environment. In addition, with the neural network IP included in the VectorBlox Accelerator SDK, users can load different network models at runtime.

     Bruce Weyer, vice president of Microchip's FPGA business group, said: "To ensure that software developers can take full advantage of the energy savings of FPGas, we need to eliminate the need for developers to learn new FPGA architectures and proprietary tool streams, while giving them the flexibility to connect to multi-framework and multi-network solutions. With the VectorBlox Accelerator SDK and neural network IP core, software and hardware developers can deploy extremely flexible overlay neural network convolutional architectures on PolarFire FPGas. With PolarFire FPGas, they can more easily build and deploy AI edge systems that are best-in-class in terms of form size, heat generation and power consumption."

     When performing inference functions at the edge, the PolarFire FPGA's total power consumption is 50 percent lower than competing products, while the math module capacity is 25 percent higher than competing products, and the number of operations per second (TOPS) is up to 1.5 trillion. Developers can also customize and differentiate themselves with PolarFire FPGas' inherent ease of upgradeability and ability to integrate different functions onto a single chip. The PolarFire FPGA Neural network IP is available in a variety of sizes to balance performance, power consumption and package size to meet application requirements, with packages as small as 11×11mm.