Startup Ricursive to Create an End-to-End AI Model for Chip Design

PALO ALTO, Calif. — Startup Ricursive, founded by the former leads of Google's AlphaChip project, announced plans to develop an end-to-end AI model for chip design, aiming to transform the industry by enabling custom chip creation for customers without in-house expertise. The company has raised $335 million, primarily for GPU compute hours, and will not use standard EDA toolchains or compete with EDA companies, focusing instead on workload-specific chip design for third parties.

PALO ALTO, Calif. — Startup Ricursive, founded by the two leads from Google’s famous AlphaChip project, aims to develop an end-to-end AI model for chip design. The new AI frontier lab is focused on hardware workload co-optimization, with the first stages of its rollout targeting workload-specific chip design for third parties. The company has raised $335 million, which will largely be spent on GPU compute hours as it trains and tests its model. In an exclusive interview with EE Times, Ricursive co-founder and CEO Anna Goldie, as well as co-founder and CTO Azalia Mirhoseini, stressed that the startup is not an EDA company, will not compete with EDA companies, and will not utilize standard EDA toolchains. “What we are doing is different to EDA , we are doing end-to-end chip design,” Mirhoseini said. “We want to transform the chip design industry, such that we can enable a new platform that can make chips for any customer that doesn’t have their own teams of chip experts in-house, but has an algorithm that they are serving at scale. We can make the chips for them such that they can do so significantly more efficiently than using an off-the-shelf chip.” Goldie and Mirhoseini have been working together for 10 years, starting the machine learning persistence team at Google Brain, one of whose projects was AlphaChip https://deepmind.google/blog/how-alphachip-transformed-computer-chip-design/ . AlphaChip, developed in 2018, was heralded as one of the first reinforcement learning RL approaches used to solve a real-world engineering problem. It was used for macro placement across four generations of Google TPUs and was adopted externally by chip companies, including MediaTek. View All https://www.eetimes.com/category/sponsored-content/ “What we’ve done in this company goes way beyond AlphaChip in terms of the breadth of what we’re taking on in the chip design flow, the performance improvements we can achieve, and the speed of the way, way higher,” Goldie said, confirming that Ricursive will not license or use any Google IP. Google had offered to make an Alphabet spinout for the pair to continue their work on AlphaChip and related projects, Goldie said, but it seemed like a middle ground. “We love Google, we grew up there, Google feels like home,” she said. “ But we’ve had more resources here at Ricursive , we’ve had more focus here, we’ve had much higher velocity here.” Ricursive wants to help build chips beyond Alphabet, and that requires being an independent company, Goldie said. “We wanted to have that broader impact,” she said. “We also thought that as a standalone company, not only would other chip makers trust us more and know that they can send us their data, but we also thought we can move faster, because we have this company whose entire mission is to accelerate this process and close this loop. That’s never really existed before.” Now is the perfect time to start an AI company, Mirhoseini added, since the needs of the industry, combined with the state of AI technology, are creating the perfect opportunity. End-to-end design Ricursive has planned three stages for its technology rollout. The company is currently working on phase one. In this phase, Ricursive will take on portions of chip design for chip companies, improving performance and helping companies get to market faster. “The goal in phase one is to accelerate the chip design process, basically, to take on the long pulls, physical design, and design verification,” Goldie said. “We believe that if you don’t accelerate both of those, then you can’t accelerate the process end-to-end. And we think it’s really, really important to be able to move more quickly from architecture to GDSII that you can actually implement.” Phase two will combine the stages of chip design into an end-to-end model that can ingest workloads and spit out GDSII files ready for manufacturing. It will enable fast, custom chip designs, which are particularly useful in AI accelerators given the scale of their deployment, but can be applied across any workload. “Our thesis is that if we customize the compute architecture to the model architecture, we can achieve massive performance improvements, but we can only unlock that performance if we can quickly implement that architecture,” Goldie said. “It can’t take a year or two to do that, or it would be obsolete.” Ricursive also wants to democratize chip design for companies that have at-scale workloads but have not previously considered custom hardware. Ricursive will do the design work for them using its AI and foundry relationships to help customers get custom chips across the line. This could apply to chips for scientific discovery and healthcare workloads, such as DNA sequencing, Mirhoseini said. “We may not need those applications at scale, but we could enable them,” she said. “By lowering the cost and the time that it takes to build a chip , we can have more applications, not just be more efficient, but potentially even enable applications that otherwise wouldn’t have been possible because of their latency or power requirements.” Hardware-workload co-design Phase three is the ultimate vision for Ricursive, in which model, workload, and hardware are tightly co-designed. “If we have this capability to quickly build highly performant chips, why not build our own chips, train our own models, and co-evolve them?” Goldie said. “Frontier models are about the tradeoff of cost versus capability, and we think we can be on a totally different Pareto-optimal curve through the co-evolution of chips and models.” Ricursive’s models will be generally intelligent not just for chip design , so the eventual chip will also apply to more general AI workloads. The intention is for AI to eventually handle everything from chip design to system and infrastructure design—the entire stack. “In the third phase, we want to make a frontier AI, period,” Mirhoseini said. “Why not co-evolve the chips and AIs together? This is not an AI for chip design, this is just AI.” This is similar to what OpenAI and Anthropic are doing, Mirhoseini said, noting that they handle code generation, math, and language with the same model. Mirhoseini was on a team that originated the original mixture-of-expert models; the eventual Ricursive model could follow this structure, with a common level of intelligence that supports different specialties, she said. “The way we think about phase three is that the chips are enablers of models,” Mirhoseini said. “If you can co-evolve the chip and model together, you can speed up training time by multipliers, you can speed up your explorations in post-training and RL by a multiplier. That means the inner loop of designing this new model can be sped up by a multiplier. That’s how it enables the AI to evolve faster. That’s how we are thinking about chips as the catalyst for self-improvement and evolution.” Ricursive is training its model on generally available web data and open-source chip data, but can also create and use synthetic data to avoid hitting the limits of what is available in the real world today. Advanced AI models need less data, Goldie said, since knowledge can transfer across tasks. Ricursive is also working with chip companies to demonstrate early results on their real chip designs and get their feedback. This is valuable since open-source data is not representative of real designs, Goldie said. While hardware-workload co-design is happening to some extent today, Goldie said it is extremely slow. Researchers working on new model architectures have to think about what runs quickly on today’s hardware, while chip makers have to think about what will be prevalent in two to three years. “It’s a deadlock,” she said. “We want to tighten that loop dramatically.” A faster end-to-end flow will enable a spectrum of custom chips and enable new applications, Goldie said. “Long term, we believe we’re at a local optimum,” Goldie said. “We have these chips that are pretty good, but we think the global optimum is potentially a very different computer architecture, and a very different neural architecture, and we want to find that.” See also: The Magic of Agentic AI Will Come From a Holistic Approach to Chip Design https://www.eetimes.com/the-magic-of-agentic-ai-will-come-from-a-holistic-approach-to-chip-design/ How to Plan Agentic AI Deployment for Chip Design https://www.eetimes.com/how-to-plan-agentic-ai-deployment-for-chip-design/ Physical AI Pushes Chipmakers Up the Value Chain https://www.eetimes.com/physical-ai-pushes-chipmakers-up-the-value-chain/