MLPerf v6.1 Atlas_Inference agentic SingleStream: Strix Halo + DGX Spark GB10 on Qwen3.6-27B-NVFP4

Fields

id

sources/mlperf-v6-1-atlas-inference-strix-halo-dgx-spark-qwen36-27b

URL

https://github.com/mlcommons/inference_results_v6.1/tree/main/closed/Atlas_Inference/results

Publisher

MLCommons / Atlas_Inference

Kind

benchmark

Accessed

2026-10-14

notes

Agentic-edge submission (BFCL function-calling harness): Strix Halo (gfx1151, ROCm core-7.13) wall 7108.6s, TTFT med 2713ms, TPOT med 47.3ms, 10.2 TPS; DGX Spark GB10 (sm_121a, CUDA 13.0) wall 3834.4s, TTFT med 1153.7ms, TPOT med 32.9ms, 20.08 TPS. Qwen3.6-27B NVFP4, 1007 turns, temp 0.

Notes

Agentic-edge submission (BFCL function-calling harness): Strix Halo (gfx1151, ROCm core-7.13) wall 7108.6s, TTFT med 2713ms, TPOT med 47.3ms, 10.2 TPS; DGX Spark GB10 (sm_121a, CUDA 13.0) wall 3834.4s, TTFT med 1153.7ms, TPOT med 32.9ms, 20.08 TPS. Qwen3.6-27B NVFP4, 1007 turns, temp 0.

Cross-references

This record references no other records.