Industry’s first MLPerf™ Inference benchmarking across multi-vendor accelerators

Cisco reports MLPerf Inference v6.1 results for a distributed service connecting NVIDIA H200 and AMD MI350X accelerators over an Ethernet fabric. The submission claims the mixed-vendor setup achieved 104.7% of the standalone performance sum, indicating no net penalty for cross-architecture orchestration. This demonstrates a method for scaling inference across heterogeneous hardware fleets without vendor lock-in.