computing HW improvements are foundation of "Moore's Law".
The Comparison: If cars had experienced the same exponential performance and efficiency acceleration as computer chips over the last several decades, you would be able to
drive to the sun in just one hour using only one gallon of gas.
Silicon Horizons | The Moonshot Podcast - YouTube
This episode of The Moonshot Podcast explores the "compute crisis"—a growing gap between the world's exploding demand for computing power and our current ability to design and build efficient chips. Host Astro Teller interviews guests from two X projects, Project Positron and Project Bodger, which seek to fundamentally reinvent how we design and build the computer chips of tomorrow ( - ).
Key Highlights
1. Project Positron: Building Brains for Robots ( - )
- The Problem: Traditional CPUs are inefficient for AI models, leading to slow response times. For robots moving through the world, real-time decision-making is critical ( - ).
- The Approach: Tammo Spalink explains that Positron focused on inference—the "performance" phase where robots act—rather than just training ( - ).
- Innovation: They took the controversial risk of using 4-bit integer representations to make chips faster and more efficient, a practice that is now more common in the industry ( - ).
2. Project Bodger: Automating Chip Design ( - )
- The Problem: Moore's Law—the 50-year trend of chips doubling in power every couple of years—is ending because transistors have hit an atomic size limit ( - ).
- The Approach: Raj Apte describes Bodger as a way to automate chip design. Instead of human-designed circuits, the goal is a "hardware compiler" that designs chips automatically based on specifications ( - ).
- Innovation: They utilize inverse design, where a computer searches the entire space of possible chip layouts to find the best design, ignoring traditional, inefficient constraints like Manhattan geometry ( - ).
No comments:
Post a Comment