Azure and AMD report new AI performance results in MLPerf Inference 6.1

Major technology providers Azure and AMD have released performance results from the MLPerf Inference 6.1 benchmarks, highlighting advancements in AI hardware and software integration.

Azure reported strong performance for the DeepSeek-R1 model using NVIDIA GB300 NVL72 and GB200 NVL72 platforms. The results included Azure’s first rack-scale and cross-rack MLPerf Inference submissions. On the GB300 NVL72 platform, Azure achieved the highest reported interactive score, while the GB200 NVL72 platform saw high throughput in both server and offline scenarios at 18-node and 72-node scales.

AMD’s MLPerf Inference 6.1 results focused on the velocity of its ROCm software. Using unchanged MI355X hardware, AMD demonstrated significant throughput improvements, including a 38% increase for GPT-OSS-120B and a 70% improvement for Wan 2.2 single-stream performance. In certain closed division tests, an 8-GPU MI355X system outperformed NVIDIA B200 and B300 submissions on the GPT-OSS-120B model by margins ranging from 13% to 33%.