Google's TurboQuant Algorithm Launches: AI Memory Usage Cut to 1/6, Inference Speed Up 8x, Storage Chip Stocks Plummet

The Key-Value Cache (KV Cache) of large AI models often consumes massive memory due to high-dimensional vectors, leading to performance bottlenecks. Google Research launched the new compression algorithm TurboQuant, which uses two core technologies—PolarQuant and QJL—to significantly reduce memory usage without sacrificing model performance.

Tests show that TurboQuant can compress KV Cache to 3 bits, cutting memory usage to 1/6 of the original and increasing inference speed by 8 times. This technology has impacted the storage chip market, with stocks of giants like Micron Technology, Western Digital falling; SanDisk even dropped by 6.5%.