MLPerf v6.1 edge: NVIDIA DGX Spark (GB10) agentic-edge SingleStream 3807.66 ms
3807.66351136147 ms SingleStream latency for the agentic-edge benchmark (lower is better; scenario-completion latency)
Fields
id
benchmarks/mlperf-v6-1-edge-dgx-spark-gb10-agentic-singlestream-latMetric
itl_ms
Value
3,807.6635
Unit
ms SingleStream latency for the agentic-edge benchmark (lower is better; scenario-completion latency)
Model
MLPerf 'agentic edge' workload
Methodology
MLPerf Inference v6.1 EDGE closed submission (Atlas_Inference, NVIDIA DGX Spark GB10, NVFP4; software: Atlas (spark), CUDA NVFP4). SingleStream latency in ms for the 'agentic edge' benchmark - this is a LATENCY result, not throughput; recorded under itl_ms as the only ms-latency member of the metric enum that fits a per-request latency measure. NVFP4-quantized run. ~1.85x faster than the same-submitter Strix Halo row (7058.99 ms).
Measured by
vendor
Reproducible
yes