A new research breakthrough from Google is putting pressure on RAM manufacturers. The development, which has not been fully detailed publicly, appears to involve a novel memory architecture or compute-in-memory technique that could reduce reliance on traditional DRAM.

The AI industry has been pushing against memory bandwidth limits for years. LLM inference and training are increasingly bottlenecked by how fast data can be moved between memory and compute, rather than by compute speed itself. Any breakthrough that changes this calculus would have significant implications.

If Google’s research leads to practical memory innovations, it could reshape the hardware landscape for AI workloads, potentially reducing the dominance of traditional RAM suppliers and changing how AI accelerators are designed.

Source