Artificial intelligence has advanced more quickly thanks to the most sophisticated silicon circuits. Can AI now repay the favor?
In order to assist engineers in creating new computer chips, Cognichip is developing a deep learning model. It is attempting to address an issue that the industry has been dealing with for decades: chip design is incredibly slow, costly, and complex. From conception to mass production, advanced chips require three to five years; the design stage alone might take up to two years before actual layout starts. Blackwell, the newest generation of Nvidia GPUs, has 104 billion transistors, which is a significant number.
According to Faraj Aalaei, CEO and founder of Cognichip, the market may shift within the time it takes to develop a new chip, rendering all of that effort unnecessary. The objective of Aalaei is to introduce AI techniques that software developers use to expedite their work in semiconductor design.
Aalaei said, “These systems have become intelligent enough that they can actually produce beautiful code by just guiding them and telling them what the result is that you want.”
According to him, the company’s technology can decrease the timetable by more than half and the cost of chip development by more than 75%.
After coming out of stealth last year, the company said on Wednesday that it had raised $60 million in new funding, led by Seligman Ventures. Notable contributors included Intel CEO Lip-Bu Tan, who will join Cognichip’s board and invest in his venture firm Walden Catalyst Ventures. Seligman managing partner Umesh Padval will also become a member of the board. Since its establishment in 2024, Cognichip has now raised a total of $93 million.
Cognichip did not reveal any of the clients it claims to have been working with since September, and it is still unable to identify a new chip created using its technique.
Instead of beginning with a general-purpose LLM, the company claims a benefit from employing its own model trained on semiconductor design data. Accessing domain-specific training data was necessary for that, which is a difficult task. The kind of open source resource that usually educates AI coding helpers is largely unavailable since chip makers rigorously protect their intellectual property, in contrast to software engineers who freely share enormous volumes of code.
In addition to licensing data from partners, Cognichip has had to create its own databases, including synthetic data. Additionally, the company has created protocols that enable chipmakers to safely train Cognichip’s models on their own confidential data without disclosing it.
Cognichip has relied on open source alternatives in situations where private data is unavailable. Last year, Cognichip invited San Jose State University electrical engineering students to participate in a hackathon to test the idea. Using the concept, the teams were able to create CPUs based on the RISC-V open-source chip architecture, which is a freely accessible design that anybody may expand upon.
In addition to well-funded companies like ChipAgents, which closed a $74 million extended Series A round in February, and Ricursive, which secured a $300 million Series A round in January, Cognichip faces competition from established firms like Synopsys and Cadence Design Systems.
According to Padval, the present infusion of money into AI infrastructure is the biggest he has witnessed in his forty years of investment.
“If it’s a super cycle for hardware and semiconductors, it’s a super cycle for companies like [Cognichip],” he stated.