AI accelerator chip raises performance, lowers power use, says Gyrfalcon
Raising the bar for high performance edge AI, Gyrfalcon Technology’s AI accelerator chip, the Lightspeeur 5801, is the fourth chip in the company’s Lightspeeur portfolio, based on the Matrix processing engine that is claimed to deliver the industry’s best ratio of high performance AI with low energy use.
The Lightspeeur 5801 offers 2.8 TOPS and uses only 224mW of power to deliver 12.6 TOPS/W. According to Gryfalcon, this is the highest ratio of performance to power use for edge AI. As AI is being integrated into more applications at the edge, it is becoming more important that AI combines high performance for reliability and precision, with the efficiency of low power use.
Edge devices have challenging requirements, so in addition to the performance and power advantages, the Lightspeeur 5801 is in a small chip size of 6.0 x 6.0mm and has a built-in USB interface which can eliminate the need for some interface components on device designs.
The Lightspeeur 5801 has four times the image input size when compared to the company’s first chip, the Lightspeeur 2801 (introduced in 2017). It allows higher resolution images to be supported and extends range of accurate input capture for analysing captured sensor data at the edge. The clock speed has also been enhanced to support 50 to 200MHz, which gives application developers a greater range of performance, yet with less than four milliseconds of latency, the chip is fast; which is one advantage for edge AI processing over cloud AI processing, says Gyrfalcon.
The chip is designed to reach even lower costs with high volume applications, as it is intended for mass market edge AI device designs. With that, customers can now get high performance AI for the premium applications in their device designs without opting for more expensive processors with built in AI blocks.
Targeted applications are image recognition, object detection and tracking, natural language processing, natural language understanding, business intelligence, facial recognition and visual analysis for consumer electronics, smart buildings, smart city, industrial, enterprise and data centre applications.
A development kit, the 5801 Plai Plug is also available from the company’s development portal. Here, users can access tools to build and deploy AI. AI models can be trained with a model development kit, and development of applications are possible with a software development kit (SDK) which is available for Linux X86_64, Microsoft Windows, Android platforms and also ARM v7l and ARM v8 instruction sets. It supports popular convolutional neural networks such as ResNet, MobileNet and VGG16, and TensorFlow, PyTorch & Caffe frameworks.
