The Future of Frontier Model Architectures with Walter Goodwin, Fractile Founder and CEO

AI inference requires a fundamental shift in chip architecture, moving beyond traditional flop-centric designs to prioritize extreme memory bandwidth. Fractile, led by CEO Walter Goodwin, addresses this by developing chips that enable multi-trillion parameter models to run at thousands of tokens per second. By moving away from SRAM toward high-bandwidth DRAM, these accelerators accommodate the growing context lengths and sparsity requirements of modern foundation models. This full-stack approach—integrating front-end design, physical layout, and deep workload analysis—allows for more agile responses to the rapid evolution of AI architectures. While hyperscalers and frontier labs experiment with proprietary silicon, the need for platform diversity remains critical; over-committing to a single hardware bet risks obsolescence if new computational breakthroughs emerge, making flexible, high-performance accelerators essential for maintaining competitive advantage at the frontier.