Google's TurboQuant Algorithm Cuts AI Memory Usage Sharply, Hitting Storage Chip Sector

KV Cache in AI large models often causes performance bottlenecks due to high-dimensional vectors occupying large memory. Google Research launched the new compression algorithm TurboQuant, which uses two core technologies—PolarQuant and QJL—to reduce KV Cache memory usage to 1/6 of the original without sacrificing model performance.

The algorithm achieved zero precision loss in tests on open-source models like Gemma, and inference speed increased by 8x on H100 GPUs. As a result, the storage chip sector saw stock price drops, with giants like Micron Technology and Western Digital experiencing declines of varying degrees.