NVIDIA Blackwell Tops First AI Hardware Benchmark: Energy Efficiency Over H200x, Outperforming AMD

According to Dynamic Insight Beating monitoring, the evaluation agency Artificial Analysis has released the industry's first Intelligent Agent Hardware Benchmark, AA-AgentPerf. Traditional evaluations are like a single-question "sprint", only focusing on response speed; whereas intelligent agent tasks are like a "relay race", where AI needs to autonomously break down the target, repeatedly circulating in tasks such as reading and writing files, rewriting code, and running tests. The frequent interactions pose a significant challenge to server memory capacity and scheduling efficiency. The benchmark, by replaying real programming traces, uses "supporting concurrent intelligent agent scale per megawatt power consumption" as the core energy efficiency metric, directly addressing data center power and funding bottlenecks.

In the initial test run, a 1.6 trillion-parameter open-source model DeepSeek V4 Pro was used. The results show that the NVIDIA Blackwell liquid-cooled full rack system GB300 NVL72 can support 61,400 concurrent intelligent agents per megawatt of power consumption, while the previous generation Hopper HGX H200 could only support 2,600, achieving an energy efficiency improvement of over 20 times. The concurrent capacity per single GPU also increased by 41 times. This allows data centers to support a 20 times larger concurrent intelligent agent scale within the same power budget, significantly reducing the landing costs of applications such as automatic programming and customer service.

In the initial results, the AMD Instinct MI355X is temporarily lagging behind. The evaluation agency pointed out that both AMD and H200 configurations were built using a general-purpose open-source vLLM framework without deep optimization; as service frameworks and kernel operators are updated, AMD's performance still has room for improvement. Currently, inference vendors like Together AI have already deployed DeepSeek V4 Pro on Blackwell, providing real-time inference support for intelligent agent programming tools like Cursor.