The aim of this collaboration is to develop the product by applying Samsung Foundry's 4nm process technology.
HyperAccel, a South Korea-based startup that was founded in January 2023, is developing a Latency Processing Unit (LPU), an AI chip specialised for transformer-based Large Language Model (LLM). The company has also developed model parallelism technology to efficiently distribute LLM to multiple LPUs and its own networking technology for data synchronisation between LPUs.
It recently released an Orion server mounted with its own LPUs. The Orion server is dedicated for emerging generative AI workloads based on LLM using Field-Programmable Gate Array (FPGA) technology.
SEMIFIVE, a specialist in the design of SoC platforms and ASIC design solutions, develops SoC design platforms for AI chips. To date, it has developed three SoC design platforms and has completed eight custom silicon tape outs using the platforms.
SEMIFIVE said that it plans to expand its roadmap to respond to customer demand for AI custom silicon. Most recently, it announced that it had begun the mass production of an AI inference custom chip for servers designed using its 14nm AI SoC platform.
Also, the mass production of NPU chip targeting HPC applications, using SEMIFIVE's 5nm HPC SoC platform is scheduled in the first half of 2024.
"We are pleased to collaborate on design with SEMIFIVE, which provides SoC platform and comprehensive ASIC design solutions," said Joo-Young Kim, CEO of HyperAccel. "By collaborating with SEMIFIVE, we are excited about the opportunity to develop semiconductors that offer lower costs and higher power efficiency compared to GPU platforms. This advancement will significantly decrease the operational expenses of data centres and expand our business scope to other industries that require LLMs."
