#AI Memory

This report examines the hardware reality of NVLink versus PCIe, dissects every major parallelism technique under PCIe‑only constraints, quantifies the viability of quantization on a 96 GB card, and provides clear, scenario‑specific guidance.

MSA or memory sparse attention, a new AI system that can remember 100 million tokens with less than 9% accuracy loss, current AI models forget everything after about 128,000 tokens.