Expedera Inc, a leading provider of scalable Deep Learning Accelerator (DLA) semiconductor intellectual property (IP), announced that a global consumer device maker is now in production with its Origin™ DLA solution.
Also Read: MediaTek Launches Dimensity 8000 5G Chip Series for Premium 5G Smartphones
Many consumer devices include video capabilities. However, at resolutions of 4K and up, much of the image processing must now be handled on the device rather than in the cloud. Functions such as low light video denoising require that data must be processed in real time, but at higher image resolutions, it is no longer feasible to transfer the volume of data to and from the cloud fast enough. To meet the expanding need for advanced on-device image processing and other new deep learning applications, device manufacturers are adding highly efficient specialized accelerators such as Expedera’s.
“I am delighted to announce the first shipping consumer product with Expedera IP,” said Da Chuang, founder and CEO of Expedera. “A key advantage of our DLA architecture is the capability to finely tune a solution to meet the unique design requirements of new and emerging customer applications. Our ability to adapt our IP to any device architecture and optimize for any design space enables customers to create extremely efficient solutions with industry-leading performance.”
In a recent Microprocessor Report, editor-in-chief Linley Gwennap noted, “Expedera’s Origin deep-learning accelerator provides industry-leading performance per watt for mobile, smart-home, and other camera-based devices. Its architecture is the most efficient at up to 18 TOPS per watt in 7nm, as measured on the test chip.”
Expedera takes a network-centric approach to AI acceleration, whereby the architecture segments the neural network into packets, which are essentially command streams. These packets are then efficiently scheduled and executed by the hardware in a very fast, efficient and deterministic manner. This enables designs that reduce total memory requirements to the theoretical minimum and eliminate memory bottlenecks that can limit application performance. Expedera’s co-design approach additionally enables a simpler software stack and provides a system-aware design and a more productive development experience. The platform supports popular AI frontends including TensorFlow, ONNX, Keras, Mxnet, Darknet, CoreML and Caffe2 through Apache TVM.