AI Designs Its Own Chips with Ricursive

In a significant development for the semiconductor industry, Riccursive is pioneering the use of artificial intelligence to design its own chips. Anna Goldie and Azalia Mirhoseini, co-founders of Riccursive, shared insights into their work, highlighting how AI is not only accelerating the chip design process but also enabling a new era of "designless" silicon. This approach promises to democratize access to high-performance chips, allowing companies to create customized solutions tailored to specific workloads.

Meet the Founders of Riccursive

Anna Goldie, CEO, and Azalia Mirhoseini, CTO, bring a wealth of experience from their time at Google, where they were instrumental in developing AlphaChip. Their work on AlphaChip, which utilized deep reinforcement learning for chip floorplanning and layout, demonstrated the power of AI in this complex field. Building on this foundation, they have launched Riccursive to bring these advanced AI capabilities to a broader market.

AI's Transformative Role in Chip Design

The conversation emphasized a key theme: neural networks are increasingly replacing traditional tools in chip design. Riccursive's approach involves a two-loop system: an "inner loop" that utilizes AI chip design tools to achieve speeds 100-1000x faster than commercial tools, and an "outer loop" where self-improving AI continuously learns from problem-solving instances. This creates a virtuous cycle of co-design exploration, leading to significant gains in performance and time-to-market.

AlphaChip's Impact and Riccursive's Vision

The work on AlphaChip, detailed in a Nature publication, showcased the AI's ability to generate chip layouts that significantly reduced wire length compared to human experts. Riccursive's vision extends this to a "designless" era, where clients simply share their requirements, and Riccursive delivers optimized, ready-to-build chips. This shift is expected to unlock a Cambrian explosion of specialized silicon, making custom chip design more accessible and efficient across various verticals, including AI models, drones, autonomous vehicles, defense, and wearables.

Phased Approach to Recursive Autonomy

Riccursive's strategy is outlined in three phases: accelerating chip design, enabling end-to-end co-design, and ultimately achieving recursive autonomy. Currently, the company is focused on Phase I, aiming to speed up the chip design process. The detailed chip design flow, from architecture design to physical signoff, is being optimized with AI at each stage. The company highlighted that traditional chip design cycles can take a year or more and involve thousands of human expert hours, a process that Riccursive's AI aims to dramatically shorten and improve.

High-Fidelity Timing Analysis in Milliseconds

A key technical achievement discussed was Riccursive's Static Timing Analysis (STA) engine, which performs full-graph timing analysis in milliseconds. The results show a remarkable 0.999+ epsilon correlation with leading commercial tools, while being over 1000x faster. This level of speed and accuracy is crucial for iterating on chip designs efficiently. The company also presented data on LLM-driven PPA (Power, Performance, Area) optimization, demonstrating how AI can achieve significant wire length reductions and performance improvements by learning from various workloads.

Democratizing Chip Design

The shift from a "fabless" to a "designless" era, as described by Riccursive, signifies a major change in the semiconductor industry. Traditionally, fabless companies separated design from manufacturing, relying on established foundries like TSMC. Riccursive's approach empowers clients to share their requirements, and the company delivers optimized, ready-to-build chips. This democratization is expected to expand the market by enabling a wider range of companies to access high-performance silicon, allowing them to focus on their core products rather than the complexities of chip design.