IBM’s Zurich lab leads effort to solve a major challenge for the next generation of IoT devices, computers and supercomputers: reducing energy consumption.
In 2013, the first scientists started to think about the first exascale high performance computer, it was already clear that the power and energy consumption were the main problems. Cristiano Malossi, current Project Coordinator and Artificial Intelligence (AI) Automation Manager at IBM Research, started exploring ideas on how to reduce the energy consumption of such machines. Together with Luca Benini, Professor at the ETH, and eight other European partners, the Zurich lab of IBM Research submitted a proposal to the Horizon 2020 (H2020) Future and Emerging Technologies (FET) Proactive funding instrument. The project was granted €5.99m in a four year period, which will end by the end of December.
The mission of Open transPREcision COMPuting (OPRECOMP) is to demonstrate that reducing precision of calculation can be done systematically for almost any type of computing application, from deep learning to simulations, without losing the quality of the result. The transprecision analytics, as they call it, could be a new foundation of computing, which saves energy by using approximations instead of “ultra-conservative precisions”.
In other words, OPRECOMP’s programmes take fewer binary digits into account and can be, therefore, faster and more efficient without compromising the result. Too much information and an excessively precise computation makes it slower and consume more power. Moreover, the OPRECOMP team assessed software modifications and evaluated the best possible trade-off for the programme. With a screening, the team can demonstrate how few bits are needed to run an application that can still produce the same results as a version run with the usual energy consumption. Interestingly enough, the studying of results show that an application can reduce substantially the amount of intermediate bits used without changing final results. As Florian Scheidegger, IBM Researcher, emphasised, this understanding of transprecision computing should be shared with the research community.
One of the outcomes of OPRECOMP is an open source library of computer code called FloatX that enables researchers to easily emulate transprecision computing on traditional computers that do not (yet) support it. The library is free of charge and has the goal of accelerating the adoption of this approach that the project has pioneered. Transprecision computing can be applied in various fields from simulations, to big data analytics and artificial intelligence. Among them, deep learning is certainly one of the applications where transprecision has the highest potential, particularly when considering models to be used for Internet of Things (IoT), in applications such as bridge and other infrastructure monitoring. More broadly, commercial applications of transprecision computing could help improve the battery-life of everyday devices by implementing a new “transprecision mode”. As an example: classical energy-save button on today smartphones save energy by turning off functionalities (e.g., GPS, Bluetooth), however a transprecision mode could reduce energy consumption of the device without turning off any service.
Dionysios Diamantopoulos, IBM researcher and OPRECOMP team member, explained how new projects reuse and develop further the results of OPRECOMP. Among new Horizon projects, one named APROPOS will explore new ways of reducing energy consumption and optimise accuracy trade-offs, which can be improved by up to 50 times. IBM Research is going to provide funding for students, who are using the OPRECOMP outcome. Another project, EVEREST, is working on a holistic approach for co-designing computation and communication in Big Data analytics to make it scalable and secure at the same time. The EVEREST project will use and adapt a demonstrator from the OPRECOMP and thus be the foundation for further projects. This recently started project also involves important collaboration with French, German, Italian, and Czech universities. The partnerships with companies like NUMTECH or Virtual Open System, both French, or the Italian DUFERCO ENERGIA SPA firm are contributing with practical expertise.
IBM Research, the lab coordinating OPRECOMP, has long been deeply involved in the EU research programmes and knows the importance of international collaboration. According to Malossi, the programmes enable access to infrastructure and networks, which otherwise are hard to find. Throughout the four years of OPRECOMP, various different partners have contributed to the success of OPRECOMP. IBM Research have, since they expanded to Switzerland, always a strong link to ETH, the Swiss Federal Institute of Technology in Zurich. He argued for Switzerland as a location that attracts excellence, which creates an inspiring environment for new ideas. However, the OPRECOMP project is founded on close collaboration also with the University of Kaiserslautern, Bologna, Belfast, Jaume or Perugia, as well as with the CEA research institution in France and the CINECA computing centre in Italy. From the private sector, companies like GreenWaves Technologies supported the project by giving different angles and transferring the transprecision computing paradigm to the marketplace.