SEMIFIVE begins mass production of Bertha

SEMIFIVE said on September 8 that it had begun mass production of Bertha, an AI inference accelerator developed for South Korean startup HyperAccel and built on Samsung Foundry’s 4nm process. The company described the chip as a data-center inference product aimed at large language model workloads and said the program was its first large-scale mass-production project on Samsung’s 4nm node. The announcement moves the Bertha program from design and contract stage into manufacturing after a 2024 agreement between the two companies. SEMIFIVE said follow-on purchase orders are expected as HyperAccel expands its services. The immediate point of the update is not just that a chip has taped out or been announced, but that SEMIFIVE says wafers are now in mass production. For AI chip companies, that is a different milestone from design wins: it means the foundry process, physical implementation and supply chain have advanced far enough to support volume manufacturing. SEMIFIVE said it delivered the project as a turnkey program for HyperAccel. (prnewswire.com) ### Why does the 500 mm² figure matter here? SEMIFIVE said Bertha is a “Big Die” chip of more than 500 square millimeters. Larger dies are harder to manufacture because yield pressure rises as chip area increases, and power and heat management become more difficult as more compute is packed onto one piece of silicon. SEMIFIVE itself highlighted those constraints in earlier materials about its big-die AI work. (prnewswire.com) Samsung’s 4nm SF4X process is also central to the pitch. SEMIFIVE said earlier this year that the HyperAccel project used Samsung Foundry’s 4nm SF4X technology and positioned it as part of its push into custom AI silicon for inference workloads. In practice, the company is using the Bertha program as proof that it can carry a large AI ASIC from design through verification and into manufacturing on an advanced node. (pr.capitalpress.com) ### Who are the companies behind Bertha? SEMIFIVE describes itself as a custom AI semiconductor, or ASIC, solutions provider. Its role in Bertha has been the design-to-production side: the company said it provided a complete turnkey solution and is now overseeing mass production. HyperAccel is the chip customer and has described its own platform as focused on LLM inference. HyperAccel’s website says its software stack supports frameworks including vLLM and Hugging Face and works with transformer-based large language and multimodal models. (semifive.com) The two companies first disclosed the production relationship in October 2024. At that time, SEMIFIVE said it had concluded a mass-production contract with HyperAccel to develop Bertha on a 4nm process, with mass production then targeted for the first quarter of 2026. The September 8 update indicates the program has now reached that manufacturing phase, though later than the original target quarter. (prnewswire.com) ### What is Bertha supposed to do? SEMIFIVE said Bertha is optimized for AI inference, including large language models. That places it in the part of the AI hardware market focused on running trained models in production rather than training them from scratch. HyperAccel has separately said its platform is designed for LLM inference across transformer-based models, which aligns with SEMIFIVE’s description of the chip. (prnewswire.com) The emphasis on inference matters because that is where service operators need hardware that can be deployed repeatedly and at scale. SEMIFIVE said anticipated follow-on orders would be tied to HyperAccel’s service rollout and expansion, linking manufacturing volume directly to customer deployments rather than to one demonstration program. (prnewswire.com) ### How does this fit into SEMIFIVE’s broader business? SEMIFIVE said the Bertha program expands its portfolio from security and high-performance computing into data-center AI. The company also said the production ramp follows first-half order intake of 42.3 billion won, using the Bertha announcement to show momentum in AI-related custom silicon work. Earlier this year, SEMIFIVE pointed to both the HyperAccel project and another 4nm program as evidence of growing traction in AI ASICs. (pr.capitalpress.com) The next concrete milestone is volume expansion. SEMIFIVE said additional purchase orders are anticipated as HyperAccel broadens its service rollout, while HyperAccel’s platform materials point to support for named LLM frameworks and transformer-based models that would sit behind those deployments. (prnewswire.com) (pr.capitalpress.com)