PDF Basket
Today’s computers are highly precise and ultra-conservative, but achieving such an incredibly high standard of accuracy requires increasing amounts of time and energy.
This means computers are hitting the power wall, where power efficiency is being lost as faster and more expensive cores are needed. And the race to the exascale – systems capable of a billion calculations per second – is being held back by power demands.
The EU-funded OPRECOMP project is exploring a different approach to computing. The team aims to create a pioneering framework for computing that involves switching from ultra-precise systems to what is known as transprecision computing.
Inspired by the human brain, this new system is being designed to expend the minimum amount of energy on a large variety of operations. Sacrificing a few digits of accuracy may significantly lower costs, save energy and improve efficiency, without compromising the quality of the final result – resulting in more efficient computers and longer lives for our devices.
‘The innovation of OPRECOMP relies on finding the right balance and combination of approximation and accurate calculation, so that power and energy costs decrease without compromising accuracy,’ says project coordinator Cristiano Malossi of IBM Research in Switzerland. ‘It’s acceptable to lose a degree of precision, as most people don’t need 16 digits of precision. However, the final answer provided by the program must be accurate, reliable and usable.’
Making marketable solutions
The OPRECOMP team is working to demonstrate the breakthrough potential of the transprecision approach in two real-life computing scenarios.
One strand is aiming for at least one order of magnitude improvement in energy efficiency for an open-source computing system. This system is aimed at the Internet of Things (IoT) domain – enabling everyday household objects such as cameras, alarms, vehicles, drones and sensors to connect, collect and exchange data over the web. There is limited power available in this domain.
The other strand is working towards the same energy-efficiency boost for traditional bigger and more complex computing systems, using IBM’s OpenPOWER technology.
Both hardware and software are being redesigned to run new transprecision algorithms efficiently. The project is also exploring mathematical theory and the basis of physical computing, with an eye to possible improvements in upcoming technology.
So far, members of OPRECOMP have produced new functional units, processing elements and memory hierarchy structures that make use of transprecision characteristics.
They have also developed early transprecision computing support for two platforms to be used in IoT applications. One of these is the ultra-low power GAP8 Application Processor marketed by GreenWaves Technologies in France. Some of OPRECOMP’s results and software have already been shared with the wider community via open-source platforms.
Educating for change
‘Even after exascale computing is achieved, designing future systems will become increasingly troublesome,’ says Malossi. ‘A shift from traditional computing paradigms is now essential. OPRECOMP aims to demonstrate to society that transprecision computing is the way to go to reduce power costs and to improve the performance of next-generation systems.’
Running alongside the core work of the project, yearly summer schools in Perugia, Italy, are aiming to spread the word by introducing others to the concept of transprecision computing.
OPRECOMP received funding through the EU’s Future and Emerging Technologies (FET) programme.