Oregon State Develops AI Chip: 50% Energy Reduction for Large Language Models

Oregon State University College of Engineering researchers have developed a revolutionary chip that consumes half the energy of traditional designs, offering a crucial solution to the vast electricity demands of AI applications like Gemini and GPT-4.

Doctoral student Ramin Javadi and Associate Professor Tejasvi Anand presented their breakthrough technology at the IEEE Custom Integrated Circuits Conference in Boston on May 8, 2025.

"The problem is that energy required to transmit a single bit isn't being reduced at the same rate as data rate demand increases," explained Anand, who directs OSU's Mixed Signal Circuits and Systems Lab. "That's causing data centers to use so much power."

The chip itself uses AI principles to reduce electricity consumption for signal processing. Large language models require tremendous data transmission over wireline, copper-based communication links in data centers, demanding significant energy.

When data transmits at high speeds, it becomes corrupted and requires cleanup. Conventional systems use power-hungry equalizers for this task. "We're using AI principles on-chip to recover data in a smarter, more efficient way by training the on-chip classifier to recognize and correct errors," Javadi explained.

The project, supported by the Defense Advanced Research Projects Agency, Semiconductor Research Corporation, and Center for Ubiquitous Connectivity, earned Javadi the Best Student Paper Award at the conference.

Instead of traditional equalizers that consume significant power, the new chip employs trained on-chip classifiers that intelligently identify and correct data transmission errors. This AI-driven approach reduces processing overhead while maintaining data integrity.

Javadi and Anand are developing the next iteration, expecting further energy efficiency gains. This breakthrough represents a crucial step toward sustainable AI infrastructure, potentially reducing the environmental impact of large-scale AI deployments while improving performance.