Conference Presentation, Product Demonstration
AI That Designs Its Own Chips: Ricursive's Anna Goldie and Azalia Mirhoseini
- Recursive Intelligence plans to evolve through three phases: initially accelerating chip design for existing manufacturers to achieve faster, cheaper, and more environmentally friendly outcomes; second, becoming a democratized platform accepting workloads like large models to design architectures end-to-end to GDS2 clean; and third, pursuing vertical integration by building proprietary chips and models to co-evolve for unmatched capability and pricing.
- The firm aims to redesign chip design tools to be 100,000x faster, anticipating that scaling compute will reduce design runtimes and enable exponential learning across large optimization spaces.
- Recursive predicts that co-designing across the stack will unlock massive performance improvements and time-to-market gains, targeting a "designless" era where clients focus on application modeling while the company provides compute and hardware.
- The outlook foresees a future where companies serving workloads at sufficient scale benefit from custom chips without needing hundreds of human experts, leading to a new era of "designless" and greater variety in chip types including low-power or high-throughput variants.
- Anticipated future demands include a significant increase in performance requirements, where even a 1% improvement on chips running frontier or large-scale AI workloads yields massive gains, prompting plans to build chips highly customized to specific user and company workload needs.