Phase I: laconic: a deep learning accelerator for energy efficiency and high-performance for edge devices

点击翻译按钮获取中文摘要

Machine Learning is transforming commerce, medicine, science, and virtually every aspect of modern life. Canada has been at the forefront of innovation in Machine Learning as we have been steadily investing in the area for decades. Machine learning would not have been possible without the computing power that modern commodity hardware was able to deliver. Further innovation will have to rely heavily in advances in computational power and thus efficiency.** Over the course of 4+ years we have been developing computing hardware concepts and device designs that deliver an order of magnitude faster data processing capability and energy efficiency than commodity hardware for Deep Learning (DL). With our 20+ years of experience in designing techniques for high-performance computing systems, we leapfrogged over first generation optimizations and instead invested in high-risk complementary techniques. Our approach was to identify and exploit typical yet fundamental properties of the value stream of Deep Learning. The results surpassed our expectations.** Our efforts culminated into Laconic, a design that outperforms other accelerators by an order of magnitude. Laconic exploits fundamental low-level value properties in Deep Learning models and requires no help from the programmer. While Laconic can be configured to target multiple market segments, we will target "edge" (embedded/mobile) applications as our previously-funded NSERC Market Assessment has shown great potential there. Our key competitive advantage is that Laconic will allow: 1) the deployment of more sophisticated and accurate DL networks, 2) a drastic reduction in energy consumption and thus improvement in up-time for mobile devices while making the system more portable and usable in a variety of scenarios. We have proven the potential of Laconic through industry standard methodologies for early design exploration. We are seeking Phase I support to build a proof-of-concept demonstration prototype which is essential for attracting interest from partner and investment organizations (some already express strong interest).****