Axelera AI BV, a Dutch supplier of artificial intelligence chips for edge devices, today announced that it has raised more than $250 million in funding.

Amsterdam-based venture capital firm Innovation Industries led the investment. It was joined by BlackRock, Samsung Catalyst Fund and several others.

Axelera AI’s flagship product is an AI chip called Metis that can perform 214 trillion computations per second. The processor usually consumes about 10 watts, which makes it suitable for use in battery-powered connected devices. A warehouse robot, for example, could use an embedded Metis processor to run AI navigation software.

Metis is powered by an architecture that Axelera AI calls digital in-memory computing, or D-IMC for short. It lowers power usage by limiting data movement inside the chip.

AI processors usually store and process data using two different sets of circuits. Data has to be regularly shuffled between those circuits during processing, which consumes power. Axelera AI’s D-IMC architecture uses SRAM, a high-speed memory variety, to store and process information in the same place. That reduces the need to move data between different sections of the chip and thereby lowers power usage.

D-IMC organizes SRAM modules in a configuration known as a crossbar array. Such arrays can carry out calculations on vectors and matrices, the two basic units of data used by AI models. A vector is a single row of numbers while a matrix comprises several rows.

“Data centers are hitting power and cooling limits, and as analytics move closer to where data is being created, edge AI solutions must operate within strict energy and bandwidth constraints,” said Axelera AI co-founder and Chief Executive Fabrizio Del Maffeo. “We designed our architecture from the ground up to overcome these obstacles.”

Axelera AI ships Metis as part of two accelerator cards that customers can attach to their connected devices. The first card contains up to four chips and attaches to the host system via a PCIe port. The other includes a single Metis unit with an M.2 interface, a space-efficient alternative to PCIe that is often used in low-power devices.

Companies can run AI models on Metis using a software toolkit called Voyager SDK. It relies on an open-source tool called Apache TVM to optimize customer algorithms for the chip. Additionally, Axelera AI provides a collection of prepackaged AI models called the Model Zoo.

The company is currently working on a second-generation chip called Europa (pictured). It can perform 629 trillion computations per second, or more than twice as many as Metis. The chip includes eight AI-optimized cores, 16 central processing unit cores and 128 megabytes of memory.

Axelera AI says that Europa can provide up to three times the performance per watt of competing products. According to the company, the chip will be particularly suitable for running computer vision models. An internal Axelera AI test determined that it can process more than 13,168 frames per second.

The company will use its new funding to ramp up chip manufacturing. In parallel, Axelera AI plans to grow its customer success team and partner network. 

Image: Axelera AI