SEMIFIVE mass-produces Bertha chip

SEMIFIVE said on September 8 that it had started mass production of HyperAccel’s Bertha AI inference accelerator on Samsung Foundry’s 4nm process, moving a project the companies first disclosed in 2024 into manufacturing. The chip is aimed at large language model inference in data centers, not model training, according to the companies’ statements. (prnewswire.com) SEMIFIVE described the program as its first large-scale mass-production project on Samsung’s 4nm node. The company said the initial production contract could be followed by additional purchase orders as HyperAccel expands services. ### How far along is this project, really? October 11, 2024 was the date SEMIFIVE said it had concluded a contract with HyperAccel to develop Bertha for mass production, with production then targeted for the first quarter of 2026. (hyperaccel.ai) January 10, 2024 was the earlier date the companies disclosed their collaboration to develop a generative AI chip using Samsung Foundry’s 4nm process. The September 8, 2026 announcement means the program has moved beyond design and tape-out language into actual manufacturing. June 2026 filings from SEMIFIVE’s own newsroom had already pointed to completed design and verification work on a “Big Die” chip for HyperAccel that exceeded 500 square millimeters on Samsung’s SF4X process. That matters because very large dies are harder to design and manufacture than smaller chips, with more pressure on yield, packaging and validation. (prnewswire.com) ### What exactly is Bertha? HyperAccel’s product materials describe Bertha 500, or B500, as an inference accelerator for cloud data centers. The company says the chip is designed around its own hardware architecture for transformer-based large language model inference and markets it as a system for high utilization and multi-chip scalability. (semifive.com) HyperAccel’s published specifications list 768 TFLOPS of compute, 546 GB/s of memory bandwidth and support for up to 1,024 concurrent requests. Those figures come from the company’s product page and have not been independently benchmarked in the materials reviewed here. ### Why does the “Big Die” detail matter? SEMIFIVE said Bertha is a turnkey solution for a large-area chip of more than 500 square millimeters. In semiconductor manufacturing, a die of that size generally implies a more demanding program because defects, yield management, physical design and packaging become more consequential as area rises. (semifive.com) Samsung’s 4nm process is also central to the announcement. (hyperaccel.ai) SEMIFIVE framed the production start as validation of its advanced-node and supply-chain capabilities, while earlier company materials tied Bertha specifically to Samsung Foundry’s SF4X technology. ### Why is this framed around inference instead of training? HyperAccel has consistently described Bertha as an LLM inference chip, and industry materials around the launch use that distinction repeatedly. A separate September 8 release from CNCF and SlashData said AI infrastructure is moving toward inference as cloud-native adoption matures, though that report was not about Bertha specifically. (prnewswire.com) EE Times wrote in March 2026 that HyperAccel was positioning Bertha 500 around lower cost and lower latency token generation in the data center. That framing fits the companies’ emphasis on serving deployed model workloads rather than frontier-model training clusters. (prnewswire.com) ### What should readers watch next? SEMIFIVE said on September 8 that follow-on purchase orders are anticipated as HyperAccel’s services expand. The next concrete checkpoint is whether those orders materialize and whether HyperAccel discloses customer deployments or shipment volumes tied to Bertha 500. HyperAccel’s product page remains the clearest public source for commercial specifications, while SEMIFIVE’s newsroom tracks manufacturing and partnership milestones. (prnewswire.com) Any future update is likely to come from one of those two companies or from Samsung Foundry ecosystem events where Bertha has already been highlighted. (hyperaccel.ai) (eetimes.com)