NVIDIA describes how full-stack codesign reduces AI memory pressure

In a blog post, NVIDIA described how full-stack codesign is helping AI systems get more useful work out of memory as larger models, longer contexts and agentic workloads increase GPU memory pressure. The company said HBM per GPU rose by about 3% from Blackwell Ultra to the Vera Rubin platform, while agentic throughput per dollar increased by up to 30x on SemiAnalysis AgentX. NVIDIA also highlighted NVFP4, Dynamo, CMX and cuVS as ways to offload model weights, KV cache and RAG data across the memory hierarchy.