NVIDIA launches AIPerf for 15-plus endpoint benchmarking

NVIDIA released AIPerf, a multiprocess successor to GenAI-Perf that supports 15-plus endpoint types—including chat, responses, image generation, and NIM rankings—and captures latency, throughput, and GPU telemetry in one run, according to NVIDIA’s launch guide. Its arrival matters because a single-process load script can become the bottleneck under concurrency; the open-source repository provides the migration path and reproducible exports. Inference teams should use it to compare serving stacks under their own traffic shape, then keep application-quality gates separate as the archive’s MLPerf analysis shows.