Axelera Offers 210 TOPS For The Edge

Axelera’s Metis AI accelerator achieves over 200 TOPS using digital in-memory computing for high-end edge applications. Typical power is less than 10 W; module pricing is below that of incumbents.

Bryon Moyer

Axelera’s Metis AI accelerator chip achieves 210 TOPS of peak inference acceleration for high-end edge applications. Typical single-digit power and module prices below $600 can attract attention in a market accustomed to double-digit (or higher) power and four-digit (or higher) module pricing. The startup’s first chip positions the company among the leaders in edge-AI power efficiency.

Metis’s performance, power, and cost reflect an in-memory-computing approach that uses only digital logic to ensure manufacturability and scalability across process nodes. A data prefetching scheme hides data latency behind computing to improve performance. With samples shipping today and production expected Q2 2024, the chip is in the final stages of qualification.

Based in the Netherlands, CEO Fabrizio del Maffeo and CTO Evangelos Eleftheriou founded the company in 2021. Del Maffeo’s background is in sales at Advantech and management at AAEON and BitFury. Eleftheriou hails from IBM Research in Zurich where he oversaw IBM’s Hermes core, which implemented in-memory compute using phase-change memory. Total company funding is $50 million (plus a $6.7 million special Dutch-agency loan); its largest investor is Innovation Industries.

Most edge-oriented inference chips offer at most tens of TOPS of performance; few manufacturers address the range above 100 TOPS. Those chips tend to be available in modules, and power starts in double-digits. Module pricing typically exceeds $1,000, leaving lots of room for companies that can create a smaller die. With a 12 nm die on modules ranging in price from $150 to $550, Axelera sees an opportunity to win business with lower prices.