1 min read
Add as a preferred source on Google

Axelera Unveils Titania Architecture for a Six-Kilowatt AI Server

The design scales to a 400-watt chip and is intended to fit existing air-cooled enterprise racks. Axelera’s Europa chip has been shipping since September.

Compact inference hardware installed beside storage and networking equipment / TokenPost.ai
Compact inference hardware installed beside storage and networking equipment / TokenPost.ai

Axelera AI unveiled Titania, a third-generation chip architecture designed to support a six-kilowatt server that could let enterprises add AI inference hardware to existing air-cooled data centers without building a separate server room.

The architecture scales the company’s design up to a 400-watt chip. Axelera says the resulting server is built to fit into existing enterprise racks rated for 20 to 30 kilowatts. Its Europa chip has been shipping since September.

Manuel Botija, Axelera’s vice president and head of product, said companies could place the inference system alongside storage, networking and CPU equipment. Inference is the processing of new inputs by a trained AI model.

Botija described the cost and power requirements of GPU racks as barriers to companies owning their AI infrastructure. He said a rack costs at least $1 million, often far more, and usually needs liquid cooling.

Axelera’s design uses digital in-memory computing, a technique that grew out of research by the company’s founders. Botija said the approach could help companies own their infrastructure, spread the investment over time and reduce operating expenses.

Simon Yoon

Reporter

Simon Yoon reports on blockchain technology for TokenPost. Send corrections or tips to info@tokenpost.com.

Loading…