Google tweak cuts LLM memory needs
Social posts this week show a new Google algorithm is significantly lowering memory requirements for LLMs, and observers say that could push down PC RAM prices after months of stagnation. The social thread includes images of price trends and debate about compute efficiency gains. (x.com)
Google Research published TurboQuant on March 24, 2026, describing a two‑stage quantization framework that compresses LLM key–value (KV) caches down to about 3 bits and reports a 6× reduction in KV cache memory. (research.google)) The paper’s benchmarks claim up to an 8× speedup on NVIDIA H100 GPUs for attention computation while producing identical outputs with no retraining or accuracy loss. (tomshardware.com)) Within 24 hours developers began porting TurboQuant math into local inference projects such as llama.cpp and MLX, even though Google had not published an official code release at the time. (venturebeat.com)) Financial markets reacted immediately: shares of Samsung Electronics, SK Hynix and Micron fell roughly 4–7% across Seoul and U.S. trading after the TurboQuant announcement. (bloomberg.com)) That market move landed against a tight supply backdrop—TrendForce and industry trackers reported PC DRAM contract prices were forecast to jump about 55–60% quarter‑over‑quarter and blended DDR4/DDR5 contract prices were projected up roughly 105–110% in Q1 2026. (nand-research.com)) Some U.S. retailers showed small, short‑term DDR5 price softening in the day or two after the news, but analysts cautioned to CNBC and Bloomberg that TurboQuant targets inference KV cache compression and may not immediately reduce demand for HBM and other server memory that has driven the 2026 price surge. (notebookcheck.net))
