Aitomatic and the other collaborators under the Foundation Models workgroup of the AI Alliance unveiled a new semiconductor-focused artificial intelligence (AI) model earlier this month. Dubbed SemiKong, the large language model (LLM) is built on the domain-expert agents (DXA) architecture and has been trained on high-quality data about the industry. The model is expected to assist human professionals across the entire process of developing semiconductor devices and processes. The foundation model of SemiKong is Meta's Llama 3.1 70B, which has been fine-tuned with semiconductor industry documents, research papers, and anonymised design and manufacturing data.
SemiKong Can Reduce Chipset Design Time by Up to 30 Percent
The new AI model has been detailed in a blog post on Meta. Developed by Palo Alto-based AI firm Aitomatic and several other collaborators from the AI Alliance, SemiKong is a semiconductor-focused model. This means SemiKong only contains data about the specific industry, including designing, manufacturing, fabrication, as well as development of semiconductor devices and processes.
It is an open-source AI model is available on GitHub and Hugging Face with the Apache-2.0 license for individual and commercial use cases. It is a bilingual language model trained on three trillion multilingual tokens and is available in four variants — SemiKong-8B, SemiKong-70B, SemiKong-8B Instruct, and SemiKong-70B Instruct.
Aitomatic's CEO Christopher Nguyen said that the AI model was aimed at tackling the knowledge gap in the semiconductor industry which has witnessed many veterans retire in recent years without the knowledge being replaced sufficiently. Additionally, Nguyen claimed that SemiKong can promote innovation and collaboration “across the industry to adopt AI in accelerating many mission-critical manufacturing and operational procedures.”
SemiKong is based on a neurosymbolic agentic AI architecture dubbed Domain-Aware Neurosymbolic Agents (DANA). It takes a three-pronged approach where expert knowledge is structured, human expertise is augmented with synthetic knowledge to train a DXA, and trained DXAs are connected to semiconductor companies' manufacturing execution systems to automate technical analyses and decisions.
Based on Aitomatic's internal testing, SemiKong can offer a 20 to 30 percent reduction in time-to-market for new chip designs. It is also said to offer up to 25 percent improvement in “first-time-right” rates in chip manufacturing. Additionally, it is also claimed to help young professionals speed up the onboarding process by 40 to 50 percent.
