Semifive starts mass production

SEMIFIVE said on Tuesday it had begun mass production of HyperAccel’s Bertha AI inference chip, moving a South Korean custom-silicon design program into manufacturing on Samsung Foundry’s 4nm process. The announcement gives HyperAccel a production milestone for a chip aimed at large language model inference in data centers. SEMIFIVE said it is providing a turnkey production solution for the device and called the program its first large-scale mass-production project on Samsung’s 4nm node. (prnewswire.com) The chip is notable for its size as well as its timing. (prnewswire.com) SEMIFIVE has previously described Bertha as a “Big Die” design exceeding 500 square millimeters, and HyperAccel markets a Bertha 500 product for AI inference workloads. The companies disclosed in October 2024 that they had signed a contract to develop Bertha for mass production using 4nm technology, with production then targeted for the first quarter of 2026. (prnewswire.com) ### Why is this chip getting attention now? September 8 is the date SEMIFIVE said mass production had started. That matters because it turns an earlier development contract into a manufacturing program and gives HyperAccel a named production partner and process node for its inference accelerator. SEMIFIVE said follow-on purchase orders are anticipated as HyperAccel expands services tied to the chip. (prnewswire.com) Yahoo Finance’s pickup of the release said the project marks SEMIFIVE’s first large-scale mass-production effort on Samsung Foundry’s 4nm process. That aligns with the company’s own framing of the program as a test of its ability to move a large AI die through advanced-node manufacturing. (semifive.com) ### What exactly is Bertha? HyperAccel says Bertha is an AI inference accelerator built around what it calls an LLM Processing Unit, or LPU, designed for transformer-based large language model workloads. In product material for Bertha 500, HyperAccel says the chip is intended for cloud data center inference and lists 768 TFLOPS of compute, 546 GB/s of memory bandwidth and support for 1,024 concurrent requests. (prnewswire.com) May 29, 2026, is when SEMIFIVE separately highlighted its work on Bertha at Samsung’s SAFE Forum in San Jose. In that release, the company said it had completed design and verification work for a Big Die exceeding 500 mm² on Samsung’s SF4X 4nm process. (finance.yahoo.com) ### What is SEMIFIVE selling in this deal? SEMIFIVE said it delivered a “complete turnkey solution” for Bertha, covering the path from chip design into production. The company has described itself as a platform-based custom silicon provider and, in its 2024 announcement with HyperAccel, said its Samsung Foundry partnership and broader ecosystem let it offer a one-stop SoC design solution. (design-reuse-embedded.com) Samsung Foundry is the manufacturing partner named for the 4nm process. The companies did not disclose production volumes, pricing or a customer shipment date in the September 8 announcement. ### How does this connect to the earlier 2024 contract? (semifive.com) October 11, 2024, is when SEMIFIVE first announced the HyperAccel contract for Bertha mass production. At that point, the companies said production was targeted to begin in the first quarter of 2026 and identified Bertha as a generative AI chip using 4nm process technology. The September 8 update shows the program has now reached production, though later than the initial first-quarter 2026 target. (prnewswire.com) SEMIFIVE said additional purchase orders could follow as HyperAccel’s service footprint expands, while HyperAccel continues to market Bertha 500 as an inference accelerator for cloud data centers. (prnewswire.com)