forecast
100x architecture gains possible by cutting data movement
Because the vast majority of an H100's power goes to moving data (networking and memory) rather than compute, minimizing data movement and maximizing compute could yield 100x efficiency gains from architecture alone - even if process shrinking stopped entirely - via more efficient ALUs, better memory technologies, power delivery, and chip-to-chip networking.
forecastpending
Speaker
Dylan PatelEvidence Quote
“The compute can get like 100x more efficient just with design changes, and then you could minimize that data movement massively.”
Source
@Asianometry & Dylan Patel — How the semiconductor industry actually works— Dwarkesh Patel PodcastCreated: 6/13/2026, 3:26:32 AM
Your notes
Loading notes...