Google Unveils Liquid-Cooled TPU 3.0 to Accelerate AI Training
Google reveals its third-generation Tensor Processing Unit, which delivers eight times the performance of its predecessor and requires liquid cooling to handle the heat. The new chip aims to meet the growing computational demands of modern, unified neural networks.
Google introduces the third generation of its custom Tensor Processing Unit (TPU) at the Google IO developer conference, claiming a pod of these new chips operates eight times faster than the previous generation. The tech giant develops this specialized silicon because modern neural networks merge across various tasks like imaging and speech recognition, resulting in larger models that demand significantly more computational power.
Unlike the original TPU that served as a basic math accelerator with limited instructions, the TPU 3.0 represents a major leap in complexity and standalone capability. The performance gains are so substantial that the new processors generate extreme heat, forcing Google to implement liquid cooling to keep the silicon from overheating during intensive machine learning workloads.
While Google keeps the exact internal specifications of the TPU 3.0 a secret, the company typically waits until a chip is heavily deployed before sharing detailed information. It is widely expected that this latest processor will follow the path of the TPU 2.0 by becoming available to developers through Google Cloud to train advanced artificial intelligence systems.