Nvidia's Blackwell Debuts First Agent Hardware Benchmark, 20x More Efficient Than H200

According to Artificial Analysis, Nvidia's Blackwell GB300 NVL72 liquid-cooled cabinet system supports 61,400 concurrent agents per megawatt of power in a newly released agent hardware benchmark AA-AgentPerf, compared to 2,600 for the prior-generation Hopper HGX H200. The benchmark measures efficiency by replaying real programming workflows, where AI agents autonomously decompose tasks and iterate through file operations, code modifications, and testing cycles. Single-GPU concurrent capacity improved by 41 times.

With equivalent power budgets, data centers can now support 20 times more concurrent agents, significantly reducing deployment costs for autonomous programming and customer service applications. AMD's Instinct MI355X currently trails in initial results, though the evaluation firm noted that AMD's performance may improve as service frameworks are optimized.