Axelera AI successfully attracted a massive investment of 36 trillion Korean Won… Achieving a breakthrough in AI chip performance
Dutch edge device AI chip supplier Axelera AI announces it has raised over $250 million (approximately 3.6 trillion KRW). The round was led by Amsterdam-based venture capital firm Innovation Industries, with participation from investment firms such as BlackRock and Samsung Catalyst Fund.
Axelera AI’s flagship product, Metis, is an AI chip capable of performing 2.14 quadrillion operations per second. The processor consumes about 10 watts, making it suitable for battery-powered connected devices. For example, warehouse robots can use built-in Metis processors to run AI navigation software.
Metis is driven by Axelera AI’s architecture called Digital In-Memory Computing (D-IMC). This technology reduces power consumption by limiting data movement within the chip. AI processors typically use two different circuits to store and process data. During processing, data needs to periodically move between these circuits, which consumes power. Axelera AI’s D-IMC architecture uses high-speed SRAM memory to store and process data at the same location, reducing the need to transfer data across different parts of the chip.
D-IMC organizes SRAM modules into configurations called crossbar switch arrays. These arrays can perform calculations on vectors and matrices— the two fundamental units of data used in AI models. Axelera AI claims this innovation aims to address power and cooling limitations faced by data centers. Since edge AI solutions must operate close to data sources, they must adhere to strict energy and bandwidth constraints.
Axelera AI has integrated Metis into two types of accelerator cards that can be added to connected devices. The first card can contain up to four chips and connects to the host system via PCIe. The other card features a single Metis unit with an M.2 interface, which is a space-efficient alternative often used in low-power devices.
Companies can run AI models on Metis using a software toolkit called Voyager SDK. This toolkit leverages the open-source compiler Apache TVM to optimize customer algorithms. Additionally, Axelera AI offers a collection of pre-packaged AI models called Model Zoo.
Currently, Axelera AI is developing its second-generation chip, Europa. Europa can perform 629 trillion operations per second, twice the performance of Metis. The chip includes 8 AI-optimized cores, 16 CPU cores, and 128MB of memory. According to the company, Europa’s performance per watt can reach up to three times that of competitors, making it especially suitable for running computer vision models. Internal testing shows the chip can process over 13,168 frames per second.
Axelera AI plans to use the funds from this round to expand chip manufacturing and strengthen its customer success team and partner network.