📜 ⬆️ ⬇️

Prototype inaccurate processor created

For 50 years, processor manufacturers have been trying to increase reliability and stability, they thoroughly test and implement several levels of error correction - naturally, this leads to a complication and cost of construction. In other words, the mass of energy and performance is spent on achieving perfect accuracy. But after all, there are areas of use where 100% reliability is not needed, and 92-99% is enough. For example, processing analog signals, photos and videos, pattern recognition.


enlarge image

The illustration shows the frames from the video processing software. On the left is a normal frame, in the middle is a frame processed by an “inaccurate” processor with a relative error of 0.54%, on the right, an even more truncated version of the processor with a relative error of 7.58%.

A group of scientists from Rice University (Houston), the University of California, Nanyang Technological University (Singapore) and the Center for Electronics and Microtechnologies (Switzerland) set the goal of making the processor as cheap and energy efficient as possible by sacrificing the accuracy of the calculations.
It turns out that a microprocessor with such tolerances can work several times more efficiently , taking advantage of speed, power consumption and size.
')


Work on the creation of such a processor at Rice University has been going on since 2003, and this week they showed the first working prototype. The presentation took place at the ACM International Conference on Advanced Computing Development (ACM International Conference on Computing Frontiers) in Cagliari, Italy, winning the award for the best scientific work. Earlier, it entered the list of the 10 most promising technologies in 2008 , according to the MIT Technology Review.

To maximize the simplicity of the design, scientists have removed all rarely used circuits from the microchip. The first tests conducted in 2011 . It turned out that after simplifying the design, the performance of the processor is doubled and it consumes half the energy at half the size.

Recent tests have shown an even better result: a reduction in energy consumption of 3.5 times with an error tolerance of just 0.25%. If you allow up to 8% of errors, then the efficiency increases by several times.

Source: https://habr.com/ru/post/144144/


All Articles