Axelera AI Unveils the Metis M.2 Max, a Higher-Performance Accelerator Tuned for LLMs and VLMs
Looking to run a large language model on a small form factor or power-constrained device? The Metis M.2 Max could help.
Machine learning and artificial intelligence (ML and AI) acceleration specialist Axelera AI has announced an upgraded take on its Metis M.2 module — delivering double the memory bandwidth to boost the performance of workloads like on-device large language models (LLMs): the Metis M.2 Max.
"We continue to set the price-performance ratio benchmark for the AI accelerator market," claims Aexlera AI chief executive officer Fabrizio del Maffeo, though it must be pointed out that the company has not yet announced pricing for its new hardware. "Our goal is to make it possible for our customers to deploy transformative edge AI applications at scale. Metis M.2 Max and our original Metis M.2 card enable customers to choose the solution that best fits their requirements – from computer vision to large language models to the latest of transformer models. We are also adding enhanced security to better address the growing threat landscape at the edge."
The Metis M.2 Max is, as the name suggests, an upgraded version of the company's earlier Metis M.2 — a compact accelerator board for small form factor systems and single-board computers, which delivers a claimed 214 tera-operations per second (TOPS) of minimum-precision compute for machine learning, computer vision, and artificial intelligence workloads. While still featuring the same type of Metis "Artificial Intelligence Processing Unit" (AIPU) at its heart, the new model comes with increased memory bandwidth — bringing a claimed 33 per cent performance uplift for convolutional neural networks (CNNs) and a doubling of token throughput for large language models and vision language models (LLMs and VLMs).
Other improvements in the new accelerator include a lower-profile but still-optional heatsink, an onboard power probe for automatic performance and power draw tuning, and the integration of a root of trust system designed to ensure firmware integrity — a security feature that, the company says, operates transparently and requires no user intervention, but blocks the use of unofficial firmware on the device.
More information on the Metis M.2 Max accelerator is available on the Axelera AI website. The device is expected to start shipping in the fourth quarter of 2025 in 1GB, 4GB, 8GB, and 16GB variants with or without industrial-grade temperature ratings, though at the time of writing pricing had not yet been confirmed; the standard Metis M.2 accelerator launched at $149.