Nvidia Blackwell tops the first intelligent agent hardware benchmark: energy efficiency exceeds H200 by 20 times, surpassing AMD According to Beating monitoring, the evaluation agency Artificial Analysis has released the industry's first AI hardware benchmark, AA AgentPerf. Traditional evaluation is like a single question and answer sprint, focusing only on response speed; The task of an intelligent agent is like a relay race, where the AI needs to autonomously disassemble the target, repeatedly flow through reading and writing files, rewriting code, and running tests. Frequent interactions pose a significant challenge to server memory capacity and scheduling efficiency. The benchmark targets the power and funding bottlenecks of data centers by replaying real programming trajectories and using the core energy efficiency indicator of 'supporting concurrent intelligent agent scale per megawatt of power consumption'. The first phase of testing runs the 1.6 trillion parameter open-source model DeepSeek V4 Pro. The results show that the Nvidia Blackwell liquid cooled full cabinet system GB300 NVL72 can support 61400 concurrent agents per megawatt of power consumption, while the previous generation Hopper HGX H200 can only support 2600, with an energy efficiency improvement of over 20 times. The concurrent capacity of a single graphics card has also increased by 41 times. This allows data centers to support 20 times the concurrent scale of intelligent agents under the same power budget, significantly reducing the cost of implementing applications such as automatic programming and customer service. In the first batch of results, AMD Instinct MI355X is temporarily lagging behind. The evaluation agency pointed out that both AMD and H200 configurations are built using the universal open-source vLLM framework without deep optimization; With the adaptation of service frameworks and kernel operators, there is still room for improvement in AMD performance. At present, inference providers such as Together AI have taken the lead in deploying DeepSeek V4 Pro in Blackwell, providing real-time inference support for the intelligent agent programming tool Cursor. [Original link]