More Performance From Every Byte: How Full-Stack Codesign Solves AI’s Memory Challenge
Reported by NVIDIA and Blockchain News2 sources ↓
NVIDIA describes managing memory across GPUs, CPUs, networking, storage and software to ease pressure from larger models and longer contexts. It reports up to 30x more agentic throughput per dollar on AgentX, though the articles cite different platform comparisons.
Key points
- NVIDIA says it coordinates memory use across the AI system rather than relying only on more GPU memory
- NVIDIA reports up to 30x higher agentic throughput per dollar on SemiAnalysis AgentX
- NVIDIA compares Blackwell Ultra with Vera Rubin; Blockchain News cites Vera Rubin versus H200
- Blockchain News reports GB300 NVL72 delivers up to 10x higher throughput per dollar than H200
Sources
Details
NVIDIA says its approach coordinates GPUs, CPUs, networking, storage and software to manage data across memory tiers. The company says this is intended to address growing demand from larger models, longer contexts and agentic applications, while GPU memory supply is tight and prices are rising.
The company reports that HBM capacity per GPU rose roughly 3% from Blackwell Ultra to Vera Rubin, while agentic throughput per dollar rose by up to 30x on SemiAnalysis AgentX. Blockchain News also reports a gain of up to 30x for Vera Rubin NVL72 on a specific AgentX workload, but describes the comparison as against H200; the NVIDIA article compares Blackwell Ultra with Vera Rubin. Blockchain News separately reports up to 10x higher throughput per dollar for GB300 NVL72 than H200.
NVIDIA says its methods include reducing model-weight and KV-cache storage with NVFP4, distributing model weights across GPUs in an NVL72 system, and moving inactive context into lower-cost memory and storage. It also describes reusing KV cache across GPU memory, CPU memory and storage. The articles do not provide independent measurements of these techniques or the reported performance claims.
Written from coverage as of
Images
- NVIDIA
- Blockchain News
- NVIDIA
- NVIDIA
Stocks in this story
Change since open: 20 Nov to 21 Nov 2026
Last refresh