While cutting-edge silicon chips have been instrumental in advancing artificial intelligence, the question now is whether AI can reciprocate by revolutionizing chip design itself.
Cognichip is developing a deep learning model designed to assist engineers in creating new computer chips. This initiative addresses a long-standing industry challenge: the immense complexity, prohibitive cost, and time-consuming nature of chip design. Cutting-edge chips typically require three to five years from initial concept to mass production, with the design stage alone often consuming up to two years before the physical layout process even begins. For context, Nvidia’s latest Blackwell GPUs incorporate 104 billion transistors, highlighting the scale of this intricate alignment task.
Faraj Aalaei, Cognichip’s CEO and founder, points out that the extensive development cycle for new chips means market conditions can shift dramatically, potentially rendering significant investments obsolete. Aalaei aims to introduce AI tools, similar to those that have accelerated software engineering, into the realm of semiconductor design.
Aalaei shared with TechCrunch that, “These systems have advanced to a point where, with simple guidance on desired outcomes, they can generate exceptional code.”
He claims that the company’s technology has the potential to decrease chip development costs by over 75% and shorten the development schedule by more than 50%.
Having exited stealth mode last year, the company announced on Wednesday that it secured an additional $60 million in funding, spearheaded by Seligman Ventures. This round also saw significant involvement from Intel CEO Lip-Bu Tan, who invested via his firm Walden Catalyst Ventures and is slated to join Cognichip’s board. Umesh Padval, a managing partner at Seligman, will also take a seat on the board. Since its inception in 2024, Cognichip’s total funding has reached $93 million.
However, Cognichip has yet to showcase a new chip developed using its system and has not publicly identified any of the customers with whom it claims to have been collaborating since September.
The company asserts its competitive edge stems from utilizing its proprietary model, specifically trained on chip design data, instead of relying on a generic large language model. This approach necessitated acquiring access to highly specialized training data, a considerable challenge. Unlike software developers, who frequently share extensive codebases openly, chip designers meticulously protect their intellectual property, thus limiting the availability of the vast open-source data pools commonly used to train AI coding assistants.
To address this, Cognichip has independently created its own datasets, incorporating synthetic data, and has also licensed data from collaborators. Furthermore, the firm has established protocols enabling chipmakers to securely train Cognichip’s models using their confidential proprietary data, ensuring no exposure of sensitive information.
In instances where proprietary data is scarce, Cognichip has leveraged open-source alternatives. During a demonstration last year, the company hosted a hackathon where electrical engineering students from San Jose State University utilized the model. Participating teams successfully designed CPUs based on the RISC-V open-source chip architecture, a publicly accessible design suitable for anyone to develop upon.
Cognichip faces competition from established industry giants such as Synopsys and Cadence Design Systems, along with other well-financed startups like ChipAgents, which completed an extended Series A round of $74 million in February, and Ricursive, which secured a $300 million Series A funding round in January.
Padval noted that the present surge of investment into AI infrastructure is the most significant he has witnessed throughout his four decades in investing.
He asserted, “If we are experiencing a super cycle for semiconductors and hardware, then it is also a super cycle for companies akin to [Cognichip].”