The Thetis Core, first taped out in December 2021 with the support of imec.IC-link, is a test vehicle demonstrating the high performance and efficiency of the in-memory computing technology developed by Axelera AI.
The in-memory computing engine delivers 39.3 tera operations per second (TOP/s) with an efficiency of 14.1 TOPs/W at an INT8 precision in less than 9 square millimetres. In the case of high-sparsity activations, the peak energy efficiency reaches 33 TOPs/W. Moreover, by increasing the nominal clock frequency the peak throughput reaches 48.16 TOPs.
Multiple enhanced versions of the in-memory computing engine will be integrated inside the company’s first product, which will be announced later this year and is expected to be available in early 2023.
Axelera AI is a European company looking to accelerate artificial intelligence workloads at the edge.
Based in the AI Innovation Center of the High Tech Campus in Eindhoven, the company has R&D offices throughout Europe in Leuven (BE) and Zurich (CH) and presence in Germany, Greece, Ireland, Italy, Portugal, and Romania.
Since being set up in 2021, the company has recruited a team of more than 50 senior engineers and developers from AI companies and research centres, including Intel, Qualcomm, IBM, and imec.
“Thetis Core is an important milestone for Axelera AI. It demonstrates that our in-memory computing technology can deliver extremely high throughput, high precision, and efficiency in an extremely compact footprint,” Axelera AI CEO and co-founder Fabrizio del Maffeo said. “This chip, taped out in just a few months, shows the strength of our fast-growing team which has a unique mix of competencies, from analogue and digital in-memory computing to dataflow, RISC-V, machine learning and software.”
“The breakthrough performance of Thetis manifests the efficacy of Axelera’s in-memory computing approach. Such a novel computing paradigm is perfectly timed to meet the stringent operational requirements of the AI applications at the edge,” Axelera AI co-founder and CTO Evangelos Eleftheriou said. “The high-energy efficiency together with the high precision and the versatile software stack will address the user pain points of AI application development at the edge.”
According to the company, it’s solutions will help empower thousands of applications of AI at the edge, making the use of AI more efficient and accessible. Axelera AI’s products will be fully integrated with the leading open-sourced AI frameworks when it is launched to select customers and partners in early 2023.
The company is also currently raising its Series A fund, expected to close later this year.