[NVIDIA Proposes AI Memory Full-Stack Co-design Solution, Boosting Inference Throughput by Up to 30x] Odaily Planet Daily reports that NVIDIA announced that as AI model sizes increase, context windows lengthen, and agent applications become more widespread, GPU memory pressure continues to rise, leading to tight HBM supply and price increases. NVIDIA proposes a solution to improve memory utilization efficiency through a co-design of GPUs, CPUs, networks, storage, and software, rather than simply increasing GPU memory capacity. Specifically, NVIDIA uses NVFP4 low-precision format compression for model weights and KV Cache to move temporarily inactive contexts to a lower-cost storage layer and leverages multi-GPU distributed deployment to free up HBM space. NVIDIA disclosed that in the SemiAnalysis AgentX test, the GB300 NVL72 achieved up to 10 times the unit cost inference throughput of the H200 platform under certain workloads; the Vera Rubin NVL72 achieved up to 30 times the unit cost inference throughput compared to the GB300 NVL72, while the single GPU HBM capacity only increased by about 3%.