Chris Walker, Untether AI | AI Infrastructure Silicon Valley - Executive Series
summary
Claims a reduction of the memory wall by embedding custom SRAM next to compute to avoid the 80–90% energy lost moving data, delivering at-memory computing with chips that reach two petops of inferencing and a focus on tokens per second per watt. Ships second‑generation SpeedAI silicon and an SDK in standard card and in‑package forms (SpeedAI240 Slim) for heavy edge uses — automotive, satellites, ag — and targets a $100B inference silicon market with production deployments and MLPerf-style benchmarking scrutiny.
AI-written summary · about
