Listen to this article

​

  • Clemson University develops pTPADTP polymer to reduce AI energy use
  • New material enables memristors that store and process data together
  • p-bits in polymer tech may improve AI computing efficiency
  • Research published in Advanced Physics Research journal

 

The hunt for a way to meet artificial intelligence‘s demand for energy has made its way to Clemson University, where a team has created a new material aimed at transforming how computers process and store information.

The new material, pTPADTP, is a polymer that could be key to making AI more energy efficient and cost effective, Stephen Foulger, Clemson’s Gregg-Graniteville Endowed Chair and professor, said in a news release.

“The energy costs are not trivial, and I don’t think people understand how severe the energy costs are,” he said in the release. “Breaking the standard technology we have right now into something new to support AI and the growth of AI is what this material is all about.”

Electricity demand from data centers — the physical hubs of AI — is on course to more than double worldwide by 2030 to about 945 terawatt-hours, slightly more than Japan’s electricity consumption, according to an April report from the International Energy Agency.

Options for helping meet the demand have ranged from restarting nuclear reactors to sinking data centers in the ocean to cool them. The new material developed at Clemson adds an innovative and potentially transformative approach, the university says.

Foulger, who led the development of the material, said it could upend how computers have worked since the 1950s.

In classical computers, memory and processing are separate. One part of the chip stores the data, and another part processes it. The computer needs to move data back and forth between these two parts, which slows things down and eats up energy.

But pTPADTP could be used to build a memristor — a device that can both store and process information in the same spot. That means less data shuttling, cutting down on energy use, the release stated.

What sets the Clemson team’s work apart is its focus on probabilistic bits, or p-bits, a middle ground between the classical bits used in today’s computers and the quantum bits that power quantum computing.

Unlike regular bits, which are either 0 or 1, p-bits flip randomly between 0 and 1, but in a controlled way. This randomness is valuable for certain types of computing, such as running AI models.

The pTPADTP material developed at Clemson allows for p-bit behavior using a polymer, which is cheaper and easier to work with than traditional materials such as magnetic tunnel junctions. By creating these switches, the research opens the door to new kinds of energy-efficient computing systems that embrace randomness as a tool rather than a flaw.

pTPADTP is an acronym for poly-4-((6-(4H-dithieno[3,2-b:2′,3′-d]pyrrol-4-yl)hexyl)oxy)-N,N-diphenylaniline.

Kyle Brinkman, chair of the Department of Materials Science and Engineering, said the team’s work is a creative way of addressing one of the most challenging and urgent topics in artificial intelligence.

“This work shines a spotlight on the crucial role materials science and engineering plays in advancing AI while underscoring the innovative ways we tackle real-world challenges in the department,” he said in the release. “I congratulate Dr. Foulger and his team on publishing their findings.”

The team’s work was published in the journal Advanced Physics Research. The paper’s title was “Polymeric Memristors as Entropy Sources for Probabilistic Bit Generation.”

In addition to Foulger, the co-authors were Yuriy Bandera, Igor Luzinov, Travis Wanless, all of Clemsons’ Center for Optical Materials Science and Engineering Technologies (COMSET).

C