Social media and movie streaming companies have a zero problem.
You may not have heard about it, but it turns out that the algorithms that analyze social media relationships or make recommendations on which movie you should watch waste huge amounts of time and energy trying to multiply and add zeroes together as they do their work. Software developers, in fact, have to build software just to remove the zeroes. The answer to adding or multiplying zeroes together, for the record, is zero.
WSU researchers have developed a computer chip design that saves energy and time by better managing such useless information.
In a simulation, their hardware design worked three times faster and was 11 times more energy efficient than Graphical Processing Units (GPUs), the processors used in modern phones and laptops, in processing relationships data. They recently presented their work at the Design, Automation, and Test in Europe conference. The work is led by graduate student Aqeeb Arka along with Professors Jana Doppa and Partha Pande in the School of Electrical Engineering and Computer Science, and Biresh Joardar, who is now a Computing Innovation Fellow at Duke University. Prof. Krishnendu Chakrabarty from Duke University is collaborating with the WSU team on this project.
To come up with recommendations for movies or to connect relationships through social media, companies currently use a sophisticated machine learning tool called a Graph Neural Network.
However, when relationships are represented mathematically, there is a huge amount of data. As one looks at a very large system of relationships, only one or two will be calculable data. The rest—those with whom we don’t have relationships—are represented by zeroes, which computers spend huge amounts of time counting. In fact, to do such relationship-based analytics, companies have to use huge super computers that are continuously doing the computations.
“It is very much useless and wastes precious time,” Arka said.
To solve the computation problem, the WSU research team created a system that uses a state-of-the-art memory technology called resistive random-access memory (ReRAM). The main advantage of the ReRAM system is that, unlike our computers that store memory and do calculations in different places, the ReRAM system can do both in the same place, Arka said. Because they don’t have to constantly go back and forth to stored memory looking for data, they’re very fast in their computation.
“So, we don’t need to worry about having so many zeroes,” Arka said. “By not looking for those zeroes and doing efficient computation, it becomes much faster than general computation systems, such as GPUs.”
In addition to speeding up computation, the researchers also improved the communications system on the chip.
“With our architecture, we could scale down those huge super computers into small chips that could be used in our phones,” he said.
Because their system is a design, the researchers used a simulator to demonstrate its success. It will need further testing in the future, using a manufactured computer chip. The researchers are also working to improve on their architecture design and communications.
“There has been a lot of research in the past few years,” said Joardar. “This is a very promising technology that we hope to see in the next five to 10 years.”
The work was funded by the National Science Foundation and the U.S. Army Research Office.
