Google Applies Deep Learning to AI-Chip Development

Google and the University of California, Berkeley have created PRIME, a deep-learning algorithm that helps design fast, compact processors for artificial intelligence tasks.

Presenting PRIME, a data-driven approach for architecting hardware accelerators that trains a #DeepLearning model on existing accelerator data, improves runtime and chip area usage by 1.2 — 1.5X, and can generate accelerators for unseen applications → https://t.co/E0PcQMg3d4 pic.twitter.com/NdQWQgZ4AA

— Google AI (@GoogleAI) March 17, 2022

The new approach designs AI-chip architectures from existing designs and performance metrics.

The team said PRIME-built chips have latency up to 50% lower than those produced by classical approaches. The deep-learning approach also cut the time to generate designs by up to 99%.

The researchers compared the performance of PRIME-built chips with EdgeTPU accelerators across nine AI applications, including image-classification models MobileNetV2 and MobileNetEdge. They emphasised that the designs were optimised for each application.

The PRIME approach improved latency by a factor of 2.7 and reduced die area by a factor of 1.5. This could lower chip costs and reduce power consumption, the researchers said.

In addition, the AI-designed chips outperformed in all nine applications tested. Only three applications showed higher latency than the designs created via modelling.

According to the researchers, PRIME holds promising prospects, including creating microchips for applications requiring solutions to complex optimisation problems, and using designs of low-performance chips as training data.

In June 2021, Google described using reinforcement learning to speed up chip design from several months to six hours.

In October, the company unveiled the Pixel 6 and Pixel 6 Pro with a Tensor chip for machine learning of its own design.

In August, Samsung began using artificial intelligence to automate the process of designing computer chips.

Subscribe to ForkLog AI on Telegram: ForkLog AI — all the news from the world of AI!