Axelera Titania takes in-memory computing into data centers
Dutch chipmaker Axelera AI has detailed an accelerator architecture for running trained artificial intelligence models in data centers. Titania combines programmable processors with dedicated computing units and a different memory choice from many rivals. The company is planning cards and server boards that fit existing facilities, extending a product family previously focused on devices at the network edge.
Artificial Intelligence··Midday
From devices to data centers
Axelera AI, a Dutch semiconductor developer, disclosed Titania’s architecture on 12 November, taking its in-memory computing approach toward data-center inference. Inference is the work a trained model performs when answering a request. The company’s earlier Metis and Europa products concentrated on applications such as computer vision, automation and robotics closer to the devices collecting the data. Titania extends that roadmap to organizations running larger models in their own facilities or cloud infrastructure.[1], [2]
Memory shapes the design
Titania joins programmable processing cores to specialized artificial intelligence units in a chiplet design. It uses LPDDR6 memory rather than the high-bandwidth memory common in many advanced accelerators, alongside fast memory on the chip. Axelera presents this combination as a way to keep large model weights available while managing data movement. The architectural choice is central to the announcement; the energy and cost advantages the company expects still depend on the performance of forthcoming hardware.[1]
Cards, boards and existing tools
The roadmap covers accelerator cards and server baseboards, with both air and liquid cooling in the planned configurations. Axelera wants the systems to work within existing rack and facility limits. It also plans to carry its Voyager development environment and AxeleraScript tools across hardware generations, preserving customers’ software work. The new disclosure describes the architecture and deployment plans; it provides no independent measurements from production Titania systems. The practical offering therefore remains tied to delivery of the hardware and its software support.[1]
