By pressing a button, businesses can make their AI models many times smaller and faster on smartphones with no accuracy loss
PALO ALTO, Calif., July 12, 2022 /PRNewswire/ — CoCoPIE today launched XGen, the first AI acceleration toolchain that co-optimizes the models and code of deep neural network (DNN), making AI models many times smaller and faster for use on devices while maintaining accuracy. The current release is mainly focused on smartphones. It is the first major milestone of CoCoPIE’s efforts to clear the long-standing obstacles to AI deployment on end devices.
Now that DNN technology in the cloud has shown its potential, the market demand to deploy AI models directly on end devices has become increasingly strong in recent years. Service providers and device makers want to run AI models on end devices to provide better experiences to their customers and save operations costs. Sometimes, it is mandatory to deploy AI on end devices when it comes to data privacy that avoids data transmission to the cloud or data locality that avoids service failure due to cloud connection. It is, however, challenging to deploy AI on end devices. The current important AI models are mostly designed for cloud deployment, with billions of parameters and operations, relying on the cloud servers’ high computation power to run them. Running the models on end devices requires times of reduction in both the model size and the model execution time. Existing tools have isolated solutions that either optimize the AI model size or the execution speed, leading to unsatisfactory results and leaving the AI capabilities largely locked out on devices.
CoCoPIE XGen solves the challenge through its proprietary co-design technology in the AI model’s compression, compilation, and runtime. The technology compresses AI models in high ratio while still maintaining the model representation friendly for highly efficient code generation. It meanwhile tailors the generated code to best map and schedule the compressed model to the hardware.
With CoCoPIE XGen, an input AI model can turn into a directly deployable code with large speedups, much smaller size, and significant power savings while keeping the model accuracy to users’ satisfaction. CoCoPIE XGen is a low-code tool. Users just need to input their AI requirements through its intuitive user interface; XGen then automatically conducts the model and code co-optimization and generates the desirable code ready to be integrated into end applications. In the first release, it supports the integration with applications of Android and iOS. Future releases will support more target platforms and systems. CoCoPIE XGen also offers full controllability so that users can fully customize the optimizations as they desire.
The superior optimization capability of CoCoPIE XGen makes ubiquitous AI deployment possible on off-the-shelf devices without relying on dedicated hardware support. An example is the cross-platform real-time video super-resolution (CRSR), an AI technology that enhances low-resolution video on smartphones by upscaling the resolution on the fly in real-time. Before this CoCoPIE technology, super-resolution was possible only on large servers or high-end smartphones through special hardware support. Thanks to the superior optimizations by CoCoPIE XGen, CRSR provides the first-of-its-kind solution that delivers high-quality videos in real-time on all the current smartphones, including low-end ones.
By removing the key roadblock to AI on end devices, CoCoPIE XGen allows businesses to capitalize on the huge market of device AI that had been beyond their reach.
"Through CoCoPIE XGen, businesses can significantly improve their AI capabilities on end devices," said Dr. Xiao-Feng Li, Co-founder and CEO of CoCoPIE. "Smartphones are an essential part of consumers’ everyday lives. By removing the deployment barrier, the first release of CoCoPIE XGen will help unlock the huge market potential of mobile AI. We are excited to work with our partners to improve their competitive advantages with CoCoPIE XGen, and to accelerate the fast-growing device AI industry."
The first release of CoCoPIE XGen is an on-premise software toolset. Users can install and use it on their own machines and hence do not need to worry about privacy and data confidentiality. XGen can be deployed on either a single computer for individual uses or a cluster of computers for scalable uses.
CoCoPIE is working with its partners globally to deploy the technology to meet the various needs of AI on devices. For free trials or more information, please visit www.cocopie.ai.
CoCoPIE is the leading software provider enabling real-time AI for off-the-shelf mobile and IoT devices through its proprietary AI model compression-compilation co-design technology. CoCoPIE products significantly improve the AI device time-to-market deployment and reduce development costs. Based in Palo Alto, CA, CoCoPIE’s team consists of the world’s leading professors and researchers on AI optimizations for end devices. Learn more at www.cocopie.ai.
SOURCE CoCoPIE Inc