News summarySina FinanceNov 5, 2026, 08:17 PM
Nvidia Details AI Memory Co-Design, Citing Up to 30x Agent Throughput per Dollar
Nvidia published a company blog on November 5 U.S. time detailing its full-stack AI memory co-design strategy. [[fact-82hyix9iv]] The company says HBM capacity per GPU rises only about 3% from Blackwell Ultra to Vera Rubin, while throughput per dollar improves by up to 30x on specific agent workloads. [[fact-vcina8cik]] Nvidia says processing contexts in some agent task trajectories have reached 3 million tokens. [[fact-76nlumbkr]] NVFP4 can compress model weights to 4-bit precision, reducing storage needs further compared with FP8. [[fact-21bqylfax]] Nvidia's Dynamo inference software can offload temporarily inactive KV Cache from GPU memory to its CMX context-memory storage platform. [[fact-azp4mh4br]]