HyperAccel, a 2023 startup based in Seoul, South Korea, has developed an AI processor specifically designed for large language model (LLM) inference, called the Latency Processing Unit (LPU). The LPU is expected to revolutionize generative AI workloads by offering low latency, cost-effectiveness, and domain-specific features, potentially replacing traditional graphics processing units (GPUs) in data centers. HyperAccel claims the LPU provides up to twice the performance at a significantly better price/performance ratio compared to current GPU-based solutions.

In a strategic partnership, HyperAccel has teamed up with SemiFive, a South Korean design solution provider specializing in system-on-chip (SoC) platforms and ASIC design solutions. Together, they are developing Bertha, a generative AI chip built on 4nm process technology, with mass production planned for the first quarter of 2026. SemiFive, known for its platform-based custom silicon solutions, will be responsible for the production of Bertha, leveraging its collaboration with Samsung Foundry.

Joo-Young Kim, CEO of HyperAccel, expressed excitement over the partnership, stating that the collaboration with SemiFive would enable the delivery of AI semiconductors that outperform traditional GPU platforms in terms of cost-effectiveness and power efficiency. "This advancement will significantly reduce operational expenses for data centers and expand our business into industries reliant on LLMs," said Kim.

Brandon Cho, CEO and co-founder of SemiFive, praised HyperAccel's innovative LPU technology, stating that it positions the company as a potential leader in the global processor infrastructure as demand for LLM computation continues to rise. Cho added, "We are thrilled to be the production partner for HyperAccel's groundbreaking Bertha chip and contribute to another innovative success story built on the SemiFive platform."

This collaboration aims to enhance the global AI semiconductor market by providing a more efficient and scalable solution for LLM computation, potentially reshaping the future of AI chip technology.