Semifive, a global AI custom semiconductor specialist, has begun large-scale production of a data-center AI inference accelerator. Semifive said on Tuesday it will mass-produce AI chipmaker HyperAccel's inference accelerator, Verda, using Samsung Foundry's 4-nanometer process.

The production run is Semifive's first large-scale mass production using Samsung Foundry's most advanced 4-nanometer process. The company said it expects to secure additional purchase orders as HyperAccel expands services following an initial production contract, and forecast production volumes will continue to increase.

The product is an accelerator optimized for AI inference computing such as large language models, and is a large-area chip, or Big Die, with an area of more than 500 square millimeters. As chips get larger, the burden of managing power, heat and yield rises, requiring advanced technology. Semifive said it brought the project to mass production with a turnkey solution spanning front-end design and verification, packaging, testing and mass-production supply.

Semifive said it has progressed from starting mass production of Hanwha Vision's AI chip for security cameras, Wisenet9, in the third quarter of last year to mass production of an AI chip for high-performance computing in Japan in the second quarter of this year, and now mass production of a data-center AI inference accelerator in the third quarter.

Semifive's value of new mass-production orders in the first half of this year was 42.3 billion won, about double the 21.2 billion won it secured for the whole of last year. It rose 71 percent quarter-on-quarter to 26.7 billion won in the second quarter from 15.6 billion won in the first quarter, and the overseas share in the second quarter was 45 percent.

Semifive CEO Cho Myung-hyun (조명현) said demand for ASIC accelerators in the data-center market is rising rapidly as AI shifts from training to inference. He said the production success reflects the combination of the customer's architecture and Semifive's turnkey capabilities, covering a large-area chip of more than 500 square millimeters from design to mass production. He said the company will carry the mass-production growth momentum proven in the first half into the second half.