Micron Shares Slide 3% as Google's TurboQuant Cuts AI Memory 6x

MUMU

Memory stocks dropped after Google unveiled TurboQuant, with Micron shares sliding 3% as Western Digital fell 4.7% and SanDisk slipped 5.7%. The new algorithm compresses AI key-value cache to 3 bits, achieving a 6x memory reduction while maintaining model accuracy.

1. Google Introduces TurboQuant

Google unveiled TurboQuant compression algorithm designed to reduce AI systems’ key-value cache to 3 bits without fine-tuning, aiming to relieve memory bottlenecks in large language models and vector search engines.

2. Memory Stocks Tumble

Shares of Micron Technology fell 3% on the announcement, while Western Digital dropped 4.7% and SanDisk declined 5.7%, despite broader technology indexes advancing.

3. Algorithm Performance and Methods

Testing on open-source models like Gemma and Mistral demonstrated a 6x reduction in key-value memory size and up to 8x performance improvement on H100 GPUs; TurboQuant combines a vector rotation method and a Quantized Johnson-Lindenstrauss algorithm to preserve accuracy.

4. Implications for Micron

By slashing memory requirements for AI workloads, TurboQuant could curb demand for high-capacity DRAM and NAND products, potentially pressuring Micron’s data center sales growth and pricing power.

Sources

BF