Semifive starts Samsung 4nm production for HyperExcel accelerator

Semifive has begun large-scale volume production of HyperExcel’s Bertha, an LLM inference accelerator for data centers, using Samsung Foundry’s 4-nanometer process. The South Korean AI ASIC design firm is handling the project as a turnkey supplier, covering front-end design and verification through packaging, testing, and final volume production.

Bertha is built for inference workloads, where completed large language models generate responses to user queries. The chip is an ultra-large big die exceeding 500 square millimeters (mm²), a scale that raises challenges in power consumption, thermal dissipation, and yield.

The production launch marks a shift for Semifive from one-time development revenue toward a pipeline of recurring production business. New volume production orders in the first half of this year totaled 42.3 billion won (approximately $31.6 million), roughly double last year’s full-year total of 21.2 billion won (approximately $15.8 million). Overseas orders accounted for 45% of Q2 bookings as Semifive expanded from security camera and HPC chips into data center accelerators.