Conference Presentation, Fireside Chat, Interview
Naveen Rao: 4D Computing, AI's Energy Wall & Beating Biology
- Unconventional AI targets a 1,000x power efficiency goal for a new machine by rethinking computer foundations, with the timeline shortened from five to three and a half years due to accelerated progress.
- The company anticipates the AI market reaching a trillion-dollar valuation by 2030, potentially exceeding that figure, and plans to monetize by solving the primary data center constraint of energy scarcity.
- Forecasts indicate the world may exhaust energy supplies for AI services within three years if current model growth and demand trends continue unabated.
- A full product, defined as a virtual machine managing a rack system for model execution, is expected to be ready within two years.
- The new "dynamical computer" architecture replaces standard matrix multiplication with time-varying behavior analyzed via transition matrices, enabling compute and memory to be combined into single elements for "compute everywhere" capabilities.
- To support this architecture, the firm plans to provide Python libraries for expressing stochastic, time-varying elements rather than developing a CUDA-like equivalent.
- Model transitions will occur at the model layer rather than the operations layer, requiring significant compute resources to port existing families given the absence of fundamental operations like matmul in their traditional forms.
- Existing model families are expected to function on the new architecture, though the integration of theorists, physicists, and chip builders is identified as a primary coordination challenge.
- Dropping compute costs by 1,000x is predicted to trigger Jevons Paradox, generating the largest market in history by driving consumption to far exceed the rate of cost reduction.
- Data center operations are projected to shift from large, gigawatt-scale facilities to numerous small, local centers to enhance environmental friendliness and locality.
- The technology aims to enable the construction of "billions of robots" within the next decade capable of dynamic assembly to solve problems.
- The firm expects to surpass biological energy efficiency limits within three and a half years by reaching the physical boundaries of 2D lithography.