Amazon bought AI chip developer Perceive for $80 million to develop LLM at the edge

Xperi has announced an agreement to sell its Perceive division, which develops specialized AI chips. The buyer is Amazon, the deal is worth $80 million in cash. The takeover is planned to be completed by the end of 2024.

Perceive, headquartered in San Jose, California, USA, creates AI solutions for running large language models (LLMs) on edge devices. Such equipment usually has limited capabilities in terms of computing resources, connectivity and data storage. Perceive chips are designed taking into account the features of edge platforms.

In particular, Perceive has developed the Ergo AI processor. It claims to allow running “data center-class neural networks” even on devices with the most stringent power requirements. These could be conferencing systems and wearable gadgets. It is noted that Ergo AI can support a wide variety of AI workloads, from classification and object detection to audio and language processing.

Image Source: Perceive

How exactly Amazon intends to use Perceive’s solutions is not specified. But it is known that Amazon is acquiring the AI ​​chip developer through its Devices & Services division, which includes the Alexa voice assistant, smart speakers, and Echo and Fire TV series devices.

The transaction is not expected to require regulatory approval. Following the acquisition, most of Perceive’s 44 employees will join Amazon. Perceive, led by co-CEOs Murali Dharan and Steve Teig, has employees in the US, Canada, Ireland, Romania and Estonia. The Perceive laboratory in Boise (Idaho, USA) will continue to operate.

It should be noted that the Amazon Web Services (AWS) division has been developing its own hardware solutions for AI tasks and cloud platforms for several years. These are, in particular, the Graviton family of chips and Trainium AI accelerators. Graviton4 debuted in late 2023 with 96 Arm cores for a wide range of workloads. A powerful Trainium 3 accelerator is being prepared for release, the power consumption of which can reach 1000 W.