1 min read
Add as a preferred source on Google

AMD MLPerf 6.1 Submissions Show 16% Lower Training Times

The 16% average spans four comparable eight-GPU configurations. AMD’s 64-GPU FLUX.1 run took 29.58 minutes on MI355X.

An accelerator card rests beside a precision stopwatch / TokenPost.ai
An accelerator card rests beside a precision stopwatch / TokenPost.ai

Advanced Micro Devices expanded its MLPerf Training v6.1 submissions to five workloads, with an average 16% reduction in training times across four comparable eight-GPU configurations versus v6.0.

The v6.1 results were released Nov. 16. AMD submitted five benchmarks, up from three in v6.0. The entries covered GPT-OSS-20B and DLRMv4, as well as Llama 2 70B LoRA, Llama 3.1 8B and FLUX.1.

On eight MI355X GPUs, Llama 2 70B LoRA fine-tuning took 6.98 minutes, while Llama 3.1 8B pretraining took 72.94 minutes. Those times were 11% and 9.8% lower than the respective B200 results, and 7.4% and 4.7% higher than the respective B300 results, at the same GPU count.

AMD’s 64-GPU FLUX.1 run took 29.58 minutes on MI355X GPUs. The v6.0 result took 92.36 minutes in a 64-GPU configuration, compared with 29.58 minutes on MI355X in v6.1. AMD also publicly demonstrated FP6 training precision with FLUX.1 for the first time.

In GPT-OSS-20B training across 16- and 32-GPU configurations, AMD’s MI355X training times were 24.8% lower than B300 results and 28.4% lower than B200 results. Partner results across server and cloud systems had normalized performance averaging about 96% of AMD’s reference results, with a range from 91% to 101%.

MLPerf Training compares how long systems take to train models to a defined quality target. In the same v6.1 round, Nvidia’s Vera Rubin NVL72 recorded the fastest times in six of seven tests.

Simon Yoon

Reporter

Simon Yoon reports on blockchain technology for TokenPost. Send corrections or tips to info@tokenpost.com.

Loading…