Researchers from the College of Engineering at Oregon State University (OSU) have presented a microchip that can halve energy consumption when working with large language models, such as Gemini and GPT-4. This is an important step towards reducing the energy consumption of state-of-the-art AI applications.
The development was announced by Ramin Javadi, head of the OSU Mixed Signal Circuits and Systems Lab, and Tejasvi Anand, assistant professor of electrical engineering, at the IEEE Custom Integrated Circuits Conference that was held in Boston.
It is widely known that the processing and transmission of large amounts of data required for the operation of language models causes enormous energy consumption by data centers. With the data transfer rate constantly rising and energy efficiency technologies lagging behind, the data centers’ load on power grids is growing exponentially. This has become a very sensitive issue for major AI operators.
Mr. Javadi explains that the biggest issue is caused by distortions: at high data rates, signal quality often deteriorates and needs to be restored. This is typically achieved by equalizers, special microcircuits that successfully eliminate distortion while consuming substantial amounts of energy.
The new microchip offers an alternative: it uses a built-in classifier based on artificial intelligence, which is trained to recognize and correct errors in the transmitted signal. This method allows data to be restored with more accuracy, consuming less energy in the process.
“Essentially, to save energy when dealing with AI, we are using those AI principles on-chip by training the on-chip classifier to recognize and correct the errors,” Mr. Javadi says.
The project has already received support from the Defense Advanced Research Projects Agency under the U.S. Department of Defense, as well as the Semiconductor Research Corporation and the Center for Ubiquitous Connectivity. The research team is currently working on the next version of the chip, which is expected to be even more energy efficient.
