Etched Unveils Chip Innovations for Scalable AI Inference.

Etched has developed two significant chip-level innovations aimed at resolving fundamental physics challenges that impede the scalability of AI inference. The first, "Low-Voltage Inference," tackles thermal throttling. Traditional AI chips reduce clock speeds under heavy load to prevent overheating, but Etched's redesigned voltage delivery system allows chips to maintain peak performance during sustained inference by running cooler. The second innovation is "Cluster-Scale Memory." While High Bandwidth Memory (HBM) offers capacity, it suffers from speed bottlenecks compared to faster on-chip SRAM. Etched's hybrid approach pools memory across multiple chips within a rack, providing SRAM-level access speeds while retaining the high capacity and compute density typically associated with HBM systems. These breakthroughs are expected to unlock previously impractical AI workloads, including multi-trillion parameter models, long-context inference, low-latency agentic systems, and sustained high-throughput serving.