Energy-efficient AI model could be a game changer, 50 times better efficiency with no performance hit

Researchers made AI models 50 times more efficient without performance loss by using unconventional arithmetic and custom hardware.

: Researchers at UC Santa Cruz improved AI efficiency by 50 times using custom hardware and ternary values in neural networks. By avoiding matrix multiplication, they reduced hardware demands significantly. The new approach works on standard GPUs and could bring full AI capabilities to mobile devices.

Researchers from the University of California, Santa Cruz, have revolutionized AI efficiency by creating an AI model 50 times more efficient than current models, while maintaining performance. By replacing conventional matrix multiplication with ternary arithmetic, their approach reduces computation intensity and power consumption drastically.

The team also developed custom hardware utilizing field-programmable gate arrays (FPGA) to maximize energy-saving features of the neural network. This hardware can run the new model using only 13 watts, compared to the 700 watts needed for traditional large language model processing using typical GPUs.

Their new method works seamlessly with standard GPUs, making it accessible without custom hardware and reducing memory usage tenfold compared to traditional approaches. This could potentially allow advanced AI functionalities on mobile devices, leading to significant advancements in AI applications.