Semifive starts production of Bertha

SEMIFIVE said on September 8 that it had started mass production of HyperAccel’s Bertha AI inference accelerator on Samsung Foundry’s 4nm process, turning a 2024 development contract into a manufacturing program. The South Korean custom-chip company said the project covers a turnkey build for a die measuring more than 500 square millimeters. HyperAccel describes Bertha 500 as a data-center inference accelerator aimed at large language model workloads. The start of production gives SEMIFIVE a live reference design in one of the busiest parts of the chip market: custom silicon for AI inference. ### Why is a foundry-services company announcing this at all? SEMIFIVE is not the end customer for the chip; it is the design and manufacturing partner that helps customers turn chip concepts into producible silicon. The company says it is an official Samsung Foundry SAFE Design Solution Partner, a role Samsung defines as part of its design-service alliance for customers building custom silicon on its process nodes. (prnewswire.com) Samsung says those partners provide design services and supply-chain support around foundry manufacturing. In that context, SEMIFIVE’s announcement is less about launching a branded processor of its own and more about showing it can carry a customer design through advanced-node production. ### What exactly is Bertha? HyperAccel markets Bertha 500, or B500, as an AI inference accelerator for cloud data centers. (semifive.com) On its product page, HyperAccel says the chip is designed for high-throughput inference and lists 768 TFLOPS of compute, 546 GB/s of memory bandwidth and support for 1,024 concurrent requests. The September 8 announcement ties that product to Samsung’s 4nm manufacturing process and says SEMIFIVE is handling the program as a complete turnkey engagement. (semiconductor.samsung.com) That matters because advanced AI chips increasingly require not just logic design, but packaging, ecosystem coordination and production management across EDA, IP and manufacturing partners. Samsung’s SAFE framework is built around that model. (hyperaccel.ai) ### Why does the chip’s size keep coming up? SEMIFIVE said Bertha is a “Big Die” chip of more than 500 square millimeters. That is a notable figure because larger dies are generally harder to manufacture efficiently than smaller ones, with more exposure to yield and integration risk as area increases. The company presented the size as evidence that it can handle large-area designs on an advanced process node. (prnewswire.com) The company also called the program its first large-scale mass-production project on Samsung’s 4nm node. For a design-services firm, that is a credential as much as a shipment milestone: it shows a customer was willing to take a large AI accelerator through production using SEMIFIVE’s flow. (prnewswire.com) ### How long has this been in the works? SEMIFIVE and HyperAccel disclosed their manufacturing agreement in October 2024. At that time, the companies said Bertha would use 4nm process technology and target the start of mass production in the first quarter of 2026. The September 8 update confirms that production has now begun, though later than that initial first-quarter target. (prnewswire.com) Neither the PRNewswire release surfaced in search nor the company pages reviewed here gave a reason for the timing. ### What happens next for the program? (prnewswire.com) SEMIFIVE said it has secured an initial mass-production contract for Bertha and expects follow-on purchase orders as services expand and validation work is completed. The company also said the program broadens its portfolio from security and high-performance computing into data-center AI. (prnewswire.com) HyperAccel’s public product page continues to position Bertha 500 for cloud data-center inference deployments. The next concrete milestones are likely to come from customer validation and additional production orders, which SEMIFIVE said it expects after the current manufacturing phase. (hyperaccel.ai) (prnewswire.com)