Earlier this month, Aitomatic, along with partners from the Foundation Models workgroup of the AI Alliance, introduced a new artificial intelligence (AI) model focused on the semiconductor sector. Named SemiKong, this large language model (LLM) utilizes the domain-expert agents (DXA) architecture and has been meticulously trained on high-quality industry-specific data. The model aims to enhance the capabilities of professionals involved in the development of semiconductor devices and processes. Built on Meta’s Llama 3.1 70B foundation, SemiKong has undergone fine-tuning using a vast array of semiconductor industry documents, research papers, and anonymized design and manufacturing data.
SemiKong Can Reduce Chipset Design Time by Up to 30 Percent
This innovative AI model was discussed in detail in a blog post published by Meta. The collaboration with several members of the AI Alliance has positioned SemiKong as a dedicated semiconductor model, focusing exclusively on the nuances of the industry, including design, manufacturing, and the development of semiconductor devices and processes.
Available as an open-source AI model, SemiKong can be accessed on GitHub and Hugging Face under the Apache-2.0 license, catering to both individual and commercial applications. It features a bilingual language model trained on three trillion multilingual tokens, with four distinct variants — SemiKong-8B, SemiKong-70B, SemiKong-8B Instruct, and SemiKong-70B Instruct.
Christopher Nguyen, CEO of Aitomatic, emphasized that this AI model aims to bridge the knowledge gap in the semiconductor industry, particularly as many experienced professionals retire without adequate knowledge transfer. He also stated that SemiKong has the potential to foster innovation and collaboration within the industry, promoting the adoption of AI to accelerate critical manufacturing and operational processes.
The architecture of SemiKong is based on a neurosymbolic agentic AI framework known as Domain-Aware Neurosymbolic Agents (DANA). This approach is designed to structure expert knowledge, augment human expertise with synthetic information to train DXAs, and link these trained DXAs to the manufacturing execution systems of semiconductor companies for the automation of technical analyses and decision-making.
According to Aitomatic’s internal assessments, SemiKong could achieve a 20 to 30 percent reduction in time-to-market for new chip designs. It is also reported to enhance “first-time-right” rates in chip manufacturing by up to 25 percent and expedite the onboarding process for junior professionals by up to 40 to 50 percent.
