SEMIFIVE has started mass production of HyperAccel’s Bertha AI inference accelerator using Samsung Foundry’s 4nm process, putting another Korean AI chip into silicon.
The project is SEMIFIVE’s first large-scale production run using Samsung’s advanced 4nm node. Follow-on orders are expected as HyperAccel expands Bertha deployment in data centres.
Bertha is designed specifically for large language model inference, not training. The accelerator uses HyperAccel’s Language Processing Unit architecture, which targets the increasingly expensive business of getting trained models to produce answers without feeding every job to Nvidia.
The chip is a sizeable lump of silicon at more than 500 mm². Large dies give designers room for more compute but make power, heat and manufacturing yield considerably nastier, because defects have a much bigger target to hit.
SEMIFIVE handled the project from front-end design and verification through packaging, testing and volume supply. It says getting such a large device through Samsung’s 4nm process proves its turnkey ASIC model works.
HyperAccel has previously said its architecture concentrates on the memory bottleneck in LLM inference. Bertha uses LPDDR5X rather than expensive high-bandwidth memory, betting that a purpose-built architecture can keep language models fed without surrounding the chip with HBM.
The production deal follows SEMIFIVE’s Wisenet 9 AI ASIC for Hanwha Vision security cameras and an HPC AI processor for a Japanese customer. The company is now trying to turn design wins into recurring production revenue rather than living from one engineering contract to the next.
SEMIFIVE booked KRW 42.3 billion (€27.1 million) of new mass-production orders during the first half of 2026, nearly twice the KRW 21.2 billion (€13.6 million) booked during all of 2025. Second-quarter orders reached KRW 26.7 billion (€17.1 million), with overseas customers providing 45 per cent of bookings.
SEMIFIVE chief executive and co-founder Brandon Cho said: “As the AI landscape shifts from training to inference, demand for ASIC accelerators in data centres is surging. This successful mass production is a significant milestone that validates our unmatched execution capabilities in advanced processes.”
“By combining our customer’s innovative architecture with SEMIFIVE’s comprehensive turnkey capabilities, we seamlessly executed the entire journey from design to volume production for a large-area die of more than 500 mm². We are committed to carrying the strong mass-production momentum demonstrated in the first half through the remainder of the year,” he said.
