Machine studying and synthetic intelligence (ML and AI) acceleration specialist Axelera AI has introduced an upgraded tackle its Metis M.2 module — delivering double the reminiscence bandwidth to spice up the efficiency of workloads like on-device giant language fashions (LLMs): the Metis M.2 Max.
“We proceed to set the price-performance ratio benchmark for the AI accelerator market,” claims Aexlera AI chief govt officer Fabrizio del Maffeo, although it have to be identified that the corporate has not but introduced pricing for its new {hardware}. “Our objective is to make it attainable for our prospects to deploy transformative edge AI functions at scale. Metis M.2 Max and our unique Metis M.2 card allow prospects to decide on the answer that most closely fits their necessities – from laptop imaginative and prescient to giant language fashions to the newest of transformer fashions. We’re additionally including enhanced safety to raised deal with the rising menace panorama on the edge.”
The Metis M.2 Max guarantees extra reminiscence bandwidth for as much as a doubling in efficiency for LLM and VLM workloads. (📷: Axelera AI)
The Metis M.2 Max is, because the title suggests, an upgraded model of the corporate’s earlier Metis M.2 — a compact accelerator board for small kind issue techniques and single-board computer systems, which delivers a claimed 214 tera-operations per second (TOPS) of minimum-precision compute for machine studying, laptop imaginative and prescient, and synthetic intelligence workloads. Whereas nonetheless that includes the identical kind of Metis “Synthetic Intelligence Processing Unit” (AIPU) at its coronary heart, the brand new mannequin comes with elevated reminiscence bandwidth — bringing a claimed 33 per cent efficiency uplift for convolutional neural networks (CNNs) and a doubling of token throughput for giant language fashions and imaginative and prescient language fashions (LLMs and VLMs).
Different enhancements within the new accelerator embody a lower-profile however still-optional heatsink, an onboard energy probe for automated efficiency and energy draw tuning, and the mixing of a root of belief system designed to make sure firmware integrity — a safety characteristic that, the corporate says, operates transparently and requires no consumer intervention, however blocks the usage of unofficial firmware on the gadget.
Extra data on the Metis M.2 Max accelerator is on the market on the Axelera AI web site. The gadget is anticipated to start out delivery within the fourth quarter of 2025 in 1GB, 4GB, 8GB, and 16GB variants with or with out industrial-grade temperature scores, although on the time of writing pricing had not but been confirmed; the usual Metis M.2 accelerator launched at $149.
