SEMIFIVE begins mass production of Bertha

SEMIFIVE said on September 8 that it had started mass production of HyperAccel’s Bertha, a large language model inference accelerator being made on Samsung Foundry’s 4nm process. The South Korean custom-silicon company described the project as its first large-scale mass-production program on Samsung’s 4nm node and said the chip was built as a turnkey “Big Die” design for data-center AI workloads. HyperAccel, also based in South Korea, is positioning Bertha as a dedicated inference processor for large language models rather than a general-purpose GPU. The announcement ties back to an October 11, 2024 agreement under which SEMIFIVE said it would develop Bertha for mass production using 4nm process technology, with production then targeted for the first quarter of 2026. That earlier release identified Bertha as a generative AI chip for HyperAccel and framed the work as part of a push into LLM-specific silicon. (prnewswire.com) ### Why is this chip notable beyond another AI press release? SEMIFIVE said Bertha is a chip of more than 500 square millimeters, putting it in the category often described as a “Big Die.” The company said it delivered a complete turnkey solution for that design, including work needed to handle the power, thermal and yield challenges that rise as die size increases. (prnewswire.com) HyperAccel’s own product page for Bertha 500 says the chip is designed for AI inference in cloud data centers and supports up to 1,024 concurrent requests, with 768 teraFLOPS of compute and 546 GB/s of memory bandwidth. The company says the architecture is built around its LLM Processing Unit, or LPU, which it markets as specialized for transformer-based inference workloads. ### Why does Samsung’s 4nm process matter here? (prnewswire.com) Samsung Foundry’s 4nm node is the manufacturing process named in both the 2024 contract announcement and the new mass-production release. SEMIFIVE says it is a Samsung Foundry SAFE design solution partner, a role it describes as giving customers access to Samsung’s process technologies and broader ecosystem support. (hyperaccel.ai) The manufacturing choice matters because Bertha is aimed at inference, where power efficiency and throughput are central selling points. HyperAccel said in the 2024 announcement that the chip was intended to offer more cost-effective and power-efficient LLM features than GPU platforms. ### What does this say about SEMIFIVE’s business? (prnewswire.com) SEMIFIVE said the Bertha program expands its portfolio from security and high-performance computing into data-center AI. The company also said the mass-production ramp follows 42.3 billion won in first-half order intake, a figure it cited in the September 8 release. (prnewswire.com) Brandon Cho, SEMIFIVE’s chief executive and co-founder, said in the 2024 announcement that demand for LLM computation was rising and described HyperAccel as having scalable LPU technology for that market. Joo-Young Kim, HyperAccel’s chief executive, said at the time that the companies expected Bertha to reduce data-center operating expenses and broaden HyperAccel’s reach into other industries using LLMs. (prnewswire.com) ### What happens next for Bertha? SEMIFIVE said the current award is an initial mass-production contract and that follow-on purchase orders are anticipated as HyperAccel expands its services. The September 8 release did not disclose unit volumes, customers or revenue tied to the production ramp. (prnewswire.com) HyperAccel’s next visible step is commercial rollout around its Bertha 500 platform, which the company is already marketing for cloud data-center inference. SEMIFIVE, for its part, is now pointing to Bertha as a live reference for large-area 4nm AI chip production on Samsung Foundry. (hyperaccel.ai) (prnewswire.com)