In its newest iteration, Intel’s Gaussian and Neural Accelerator (GNA) supports vision neural networks. Although this coprocessor has operated behind the scenes for years, the company revealed its architecture at the recent Linley Spring Processor Conference.

GNA provides low-power execution of small AI models while the main CPU sleeps. It’s capable of 38 billion operations per second (GOPS). Typical audio applications include speech recognition, acoustic context awareness, dynamic noise suppression, and speaker identification and separation; the most prominent workload is wake-word recognition. The new GNA version 3.1 can process inputs from visual sensors to wake the system and authenticate users.

When it officially launched as part of the Ice Lake processor for laptop PCs, GNA 1.0 implemented Gaussian mixture models (GMMs) to assist with audio functions such as keyword recognition. Over time, Intel added neural-network processing. The company launched GNA 2.0 with Tiger Lake and 3.0 with Alder Lake. Although the coprocessor still executes GMMs for backward compatibility, new models are mainly for neural networks. Clover Falls is a standalone chip that employs GNA 3.1 for proximity detection and facial recognition to wake a PC and authenticate a user.

The GNA core is available only inside Intel; version 3.1 is shipping now in Clover Falls. Whereas the company’s designers can instantiate it on a chip, customers can gain access only by purchasing the company’s chips and SoCs, using OpenVino tools to compile AI models to the core.