BREAKING
AA-AgentPerf Benchmark Launches
What AgentPerf Measures
1
Chain model calls
↓
2
Use tools & context
↓
3
Iterate to finish
↓
4
Count agents/MW
Real Coding Agent Trajectories
Agents per MW by System
GB300 NVL72
61354
AMD MI355X
3551
H200
2594
0
x
more agents/MW vs Hopper
0
GB300 agents at 60 tok/s
Rolling Results to Come
AI NEWS BLITZ
Artificial Analysis just released the first hardware benchmark built for AI agents.