Intel / article

NVIDIA Vera Rubin and Blackwell Set a New Standard for Agentic AI Performance per Watt

Source: developer.nvidia.com
Author: Elizabeth Goodman
Date: Published Aug 24, 2026

Context

Hardware efficiency for coding agents depends on repeated requests, growing and pauses for tools. NVIDIA tested these patterns with SemiAnalysis's AgentX benchmark and reported up to 30 times the throughput per provisioned megawatt for Vera Rubin NVL72 versus GB300 NVL72, running DeepSeek V4-Pro at 160 output tokens per second per user. Its August 24 report labels these NVIDIA measurements as pending SemiAnalysis review.

AgentX's recorded-session approach replays anonymized recordings of real coding- sessions on their original request schedule, preserving context reuse and traffic timing. It measures serving capacity under those conditions, rather than whether an agent solves the original programming problems correctly. The benchmark methodology describes the traces and the performance metrics.

Terms in this piece · Glossary

context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
subagent — A separate agent spawned by another to do one scoped piece of work in its own context, returning only the result.
mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.

Read the source developer.nvidia.com