SEMIFIVE starts Bertha mass production

SEMIFIVE said Tuesday it has started mass production of HyperAccel’s Bertha artificial-intelligence inference chip on Samsung Foundry’s 4nm process, moving a project first announced in 2024 into manufacturing. The South Korean custom-silicon company said the work covers HyperAccel’s large language model inference accelerator and follows an initial mass-production contract signed in October 2024. HyperAccel markets the product as Bertha 500, or B500, for cloud data-center inference. The announcement came in a company release and was echoed in prior SEMIFIVE statements that had targeted first-quarter 2026 for the start of production. This matters because Bertha is not a general-purpose AI chip launch in the usual sense. SEMIFIVE said it is delivering a turnkey manufacturing program for a “Big Die” design larger than 500 square millimeters, a size threshold that points to a large, complex chip with tighter yield and packaging demands than smaller accelerators. SEMIFIVE said the production run validates its ability to take advanced-node custom chips from design into volume manufacturing on Samsung’s 4nm process. (prnewswire.com) ### Why is the 500 mm² figure getting so much attention? SEMIFIVE highlighted the “more than 500 mm²” die size because large dies are harder to manufacture economically, especially on leading-edge nodes. The company said the Bertha program was completed as a full turnkey engagement, covering the path from development to mass production. In chip manufacturing, a bigger die generally means fewer chips per wafer and greater exposure to defects, which is why companies tend to emphasize large-area production only when they believe it demonstrates process and integration capability. (prnewswire.com) That last point is an inference based on standard semiconductor economics and SEMIFIVE’s own framing of the project as proof of advanced-process execution. ### What exactly is Bertha built to do? HyperAccel describes Bertha 500 as an accelerator for AI inference in cloud data centers rather than for model training. On its product page, HyperAccel says B500 supports up to 1,024 concurrent requests, delivers 768 TFLOPS of compute, 546 GB/s of memory bandwidth and averages 90% hardware utilization during inference. The company positions the chip around serving transformer-based large language models more efficiently, continuing the architecture it previously described as an LLM Processing Unit, or LPU. (prnewswire.com) ### How long has this program been in the works? SEMIFIVE first disclosed the HyperAccel collaboration in January 2024, saying the chip would be developed on Samsung Foundry’s 4nm technology. In October 2024, SEMIFIVE said it had concluded a mass-production contract for Bertha and expected manufacturing to begin in the first quarter of 2026. Tuesday’s announcement is the first confirmation that production has in fact started. (hyperaccel.ai) ### What does SEMIFIVE say comes next? SEMIFIVE said it has secured an initial mass-production contract for Bertha and expects follow-on purchase orders as HyperAccel’s services expand. The company did not disclose unit volumes, wafer commitments, pricing or customer names for deployments. HyperAccel’s website currently lists Bertha 500 as part of its inference hardware lineup, providing the clearest public marker for the next step beyond this manufacturing start. (semifive.com) ### Where does Samsung fit in? Samsung Foundry is the manufacturing platform named in both the January 2024 development announcement and the September 2026 production start. SEMIFIVE has framed the Bertha program as a demonstration of its platform-based custom-silicon model on Samsung’s 4nm node, while HyperAccel is the chip designer and end-product owner. The next concrete milestone will be whether SEMIFIVE or HyperAccel discloses shipment volumes, customer deployments or additional purchase orders tied to Bertha 500. (prnewswire.com) (semifive.com)