SemiFive, HyperAccel complete design of LLM processor

Contract design work on HyperAccel’s a 4nm generative AI processor IP called ‘Bertha’ has been completed with silicon IP now targeting mass production in 1Q26.

HyperAccel Co. Ltd. (Seoul, South Korea) is a 2023 startup that has developed a large language model processing unit (LPU), an AI processor specifically designed for LLM inference, offering low latency and domain-specific features. The LPU was original defined as a Latency Processing Unit.

The company’s business model is as a dedicated processor IP licensor and provider of servers for generative AI workloads.

SemiFive Co. Ltd. (Seongnam, South Korea), a high-performance chip designer specializing in AI, announced it has concluded its design contract with HyperAccel. This was work that started in January 2024.

Up until now graphics processing units (GPUs) with more general applicability parallel processing of arithmetic have been used with great success in the datacenter for the fast-moving AI sector. This has benefited Nvidia Corp. greatly.

HyperAccel asserts that with arrival of generative AI and the use of LLMs the LPU is expected to replace the GPU as it can offer a 2x improvement in price/performance. HyperAccel is one of series of AI startups that have emerged from Korean academia.

SemiFive specializes in SoC platforms and ASIC design solutions and is developing its chiplet platform in collaboration with partners.

“By collaborating with SemiFive, we are excited to offer customers AI semiconductors that provide more cost-effective and power-efficient LLM features than GPU platforms,” said Joo-Young Kim, CEO of HyperAccel, in a statement issued by SemiFive.

Brandon Cho, CEO and co-founder of SemiFive, said: “As the demand for LLM computation is skyrocketing, HyperAccel has the potential to become a new powerhouse in the global processor infrastructure.”

SemiFive has a strong but not exclusive partnership with Samsung Foundry.