an ASIC IA that does more than ‘distinguish cats from dogs’ in photos

A SoC with 32 processing units and 23 billion transistors available as a PCIe card.

IBM Research introduces its SoC AIU (Artificial Intelligence Unit). According to the company, it is an ASIC (application specific integrated circuit) “designed to perform and train deep learning models faster and more efficiently than a CPU”. This solution takes the form of a PCIe card.

Jeffrey Burns and Leland Chang, the authors of the publication, specify that the SoC has 32 processing cores and 23 billion transistors. This amount is roughly equivalent to that of the z16, an “AI and cyber-resilience” platform based on IBM’s Telum processor. In fact, Jeffrey Burns and Leland Chang explain that the AIU SoC was not built from scratch, but rather is “a scaled-up version of an already proven artificial intelligence accelerator, integrated to our Telum chip”. Thus, “the 32 cores of IBM’s AIU are very similar to the AI ​​cores embedded in the Telum chip”. However, Telum uses transistors fabricated on a 7nm process, while those of AIU benefit from 5 nm etching.

IBM unveils the first wafer engraved in 2 nm

Improve AI hardware efficiency by a factor of 2.5 every year

These are the only technical details shared by IBM for now. However, the company aims to contribute to the rapid development of AI; not to detect dogs or cats in photos, as we can read in the article, but rather to face the great challenges of recent years. That task falls to researchers at the AI ​​Hardware Center, which was founded in 2019.

Here are the terms used by Jeffrey Burns and Leland Chang:

“The AI ​​cores embedded in Telum, and now our first dedicated AI chip, are the product of the AI ​​Hardware Center’s aggressive roadmap to increase IBM’s hardware firepower. ‘IA. Due to the time and expense involved in training and running deep learning models, we have barely scratched the surface of what AI can offer, especially for businesses.

We enabled the AI ​​Hardware Center in 2019 to fill this void, with a mission to improve AI hardware efficiency by 2.5 times every year. By 2029, our goal is to train and run AI models a thousand times faster than three years ago.

Deploying AI to distinguish cats from dogs in photos is a fun academic exercise. But that won’t solve the pressing problems we face. For AI to tackle the complexities of the real world – things like predicting the next Hurricane Ian, or knowing if we’re headed for a pullback – we need professional-grade, industrial-scale hardware. . Our IAU brings us closer. We may soon inform you of its release.

Source: IBM Research

Leave a Comment