Breakthrough
AI acceleration hardware that doesn't break budgets

Our single Metis AIPU is capable of running multiple HD camera streams and reaches up to 214 TOPS @ INT8, which equals up to 3.200 FPS for ResNet-50.

YOLOv5 object detection runs at 480 FPS (16 Full HD streams at 30 FPS) on the Metis Evaluation Kit which is for sale today, while the manufactured product is projected to reach 800 FPS.

We are able to offer the best price, because the Metis AIPU is based on the combination of D-IMC, RISC-V and standard CMOS technologies for neural network optimization.

Our proprietary D-IMC technology is a quantization method that doesn’t lead to model accuracy loss and excels in efficiency. It uses crossbar arrays of memory devices to store a matrix and perform matrix-vector multiplications without intermediate movement of data. Because each memory cell effectively becomes a compute element, the Metis AIPU excels in efficiency. That’s why Axelera AI products are the most cost-effective solutions for AI acceleration.

The Metis AIPU, based on Digital In-Memory Computing, standard CMOS and RISC-V technologies, minimizes power consumption and is 4-5 times more energy efficient than a GPU. As you can see in the graph below, our M.2 AI accelerator running on the Metis AIPU is unrivaled regarding energy efficiency (15 TOPs/W; typical power consumption of 10W).

Axelera AI matches 99% of the original model’s precision, indistinguishable from GPU-based inference models. This is due to our proprietary technology D-IMC.

Vision AI is easy to integrate in original equipment with our AI acceleration hardware, because of its open source code. You’ll have your (vision) AI application up and running in no time, as you can import models from our model zoo or plugin your own model and dataset without the need for retraining, because the TVM compiler includes post-training quantization tools.

Furthermore, the software can run seamlessly on the RISC-V-based processors structured through D-IMC without users having to worry about compatibility issues.

A major advantage of the products of Axelera AI is that its accelerator technology has been implemented in standard CMOS technology. This means that our AI acceleration hardware is compatible with a broad range of electronic products and systems.

Memory technologies such as standard CMOS are also a key driver for lower lithography nodes. So, Axelera AI will be able to easily scale performance with their products when the semiconductor industry brings advanced lithography nodes into volume production. This means that our AI acceleration hardware can be integrated into increasingly compact and diverse hardware platforms.

Providing best-in-class performance, the Axelera AI M.2 module is the foundation for AI acceleration at the Edge. Powered by a single Metis AIPU and containing 512MB LPDDR4x dedicated memory, it minimizes power consumption and simplifies integration. And at just EUR 149,- it opens the door for every application to benefit from AI processing. Providing a significant boost to any vision processing system with an NGFF (Next Generation Form Factor) M.2 socket.