News

AMD Highlights 16% Faster AI Training in MLPerf 6.1 Submissions

AMD's MLPerf Training v6.1 results show 16% faster training times and expanded benchmark coverage with Instinct MI355X GPUs.

AMD Highlights 16% Faster AI Training in MLPerf 6.1 Submissions

AMD has announced significant advancements in AI training performance and workload coverage with its submissions to MLPerf Training v6.1, published on November 16, 2026. The results highlight a 16% average reduction in training times across four comparable eight-GPU configurations using AMD Instinct™ MI350X and MI355X GPUs, compared to MLPerf Training v6.0. These improvements allow teams to iterate faster without requiring hardware changes.

Expanded Benchmark Coverage

In MLPerf Training v6.1, AMD expanded its submissions from three benchmarks in v6.0 to five, the most comprehensive coverage for AMD Instinct GPUs to date. The new benchmarks include GPT-OSS-20B and DLRMv4, alongside Llama 2 70B LoRA, Llama 3.1 8B, and FLUX.1. Notably, AMD also achieved its first public demonstration of FP6 training precision with FLUX.1, marking a significant milestone in training precision options.

Performance Improvements

AMD's submissions demonstrated competitive performance against NVIDIA's Blackwell-generation GPUs. For example, in Llama 2 70B LoRA fine-tuning and Llama 3.1 8B pre-training, AMD's MI355X GPUs achieved training times of 6.98 minutes and 72.94 minutes, respectively. These times were 11.0% and 9.8% faster than NVIDIA B200 GPUs but 7.4% and 4.7% slower than NVIDIA B300 GPUs.

In the FLUX.1 benchmark, AMD's MI355X GPUs showed a 3.1× performance improvement over its own v6.0 results, reducing training time from 92.36 minutes to 29.58 minutes on a 64-GPU configuration. Similarly, in GPT-OSS-20B training, AMD's MI355X GPUs achieved 24.8% and 28.4% faster training times than NVIDIA B300 and B200 GPUs, respectively, across 16- and 32-GPU configurations.

Reproducibility and Partner Results

AMD's reference submissions were closely matched by partner results across server and cloud systems, with an average normalized performance of approximately 96% compared to AMD's own training times. The range of partner results spanned from 91% to 101% of AMD's reference performance, demonstrating consistency across different deployment environments.

These results underscore the growing competitiveness of AMD Instinct GPUs in AI training workloads, offering customers broader workload coverage, faster training cycles, and diverse precision options for evaluating training infrastructure.

Sign in to save this article Free account: build your research collection and return to read it anytime.