What happens when AI inference has to run continuously inside enterprise servers? An accelerator architecture targets higher inference efficiency without replacing existing infrastructure.
Axelera AI has launched its Europa AI Processing Unit (AIPU), designed to handle inference workloads in enterprise and data centre environments.
The accelerator supports workloads including agentic AI, vision-language models, generative AI, machine vision and computer vision. It is available as a chip for custom board designs and as PCIe accelerator cards in half-height, half-length and full-height, full-length formats.
The key advantage is its claimed efficiency. Axelera says Europa can deliver up to six times more tokens per second per watt than GPU-based solutions, targeting workloads where power consumption and operating costs can become significant as AI usage grows.
The architecture is designed to add inference capacity to existing server infrastructure through standard PCIe connections. This allows organisations to deploy AI acceleration without rebuilding their servers around a different platform.
The accelerator is aimed at enterprise applications where inference needs to run within controlled power, cost and data environments. Local deployment can also help organisations retain control of sensitive data in sectors such as healthcare, financial services, legal, defence and government.
Europa is also supported by Axelera’s Voyager Toolchain, which helps developers optimise AI models for the hardware and move inference pipelines from development to deployment. Voyager Wingman can also port existing pipelines, create new ones and optimise code, while AxeleraScript (AxScript) provides a model-compilation approach for different inference workloads.
“We built our architecture around some of the hardest constraints in computing: power, energy, cost and the need to process data locally,” says Fabrizio Del Maffeo, CEO and co-founder of Axelera AI. “Europa applies those same principles, expanding from Physical to Enterprise AI.”
For the official announcement, click here.
