Reproducing AMD MLPerf Training v6.0 Submission Result

2026-06-16 · Source: AMD ROCm Blogs · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cloud Computing & IT Infrastructure · Depth: Intermediate, long

Summary

AMD has released a step-by-step guide for reproducing its MLPerf Training v6.0 submission results, achieved on AMD Instinct MI325X, MI350X, and MI355X GPUs. The guide covers three benchmarks: Llama 2 70B LoRA fine-tuning, Llama 3.1 8B pretraining, and Flux.1-schnell text-to-image training. For LLM benchmarks, AMD's Primus training framework was used, abstracting Megatron-LM and TorchTitan. Reproduction requires ROCm 7.2.2 or later, Docker, and specific disk space, such as 6 TB for Flux.1-schnell. The article details environment setup, dataset preparation, training configuration, execution, and result validation for each benchmark. Expected scores include Llama 2 70B LoRA at 8.27 minutes on MI355X and 10.25 minutes on MI350X, Llama 3.1 8B pretraining at 86.84 minutes on MI355X and 109.76 minutes on MI350X, and Flux.1-schnell at 92.36 minutes on an 8-node MI325X configuration.

Key takeaway

For AI Engineers evaluating AMD Instinct GPUs for large-scale model training, this guide offers a clear path to validate AMD's MLPerf Training v6.0 performance claims. You should follow the detailed steps for environment setup and dataset preparation to reproduce Llama 2 70B LoRA, Llama 3.1 8B, or Flux.1-schnell benchmarks. Your team can utilize the Primus framework for streamlined LLM workflows and ensure MLPerf-compliant result validation by averaging 8 of 10 runs.

Key insights

AMD provides a detailed guide to reproduce its MLPerf Training v6.0 benchmark results on Instinct GPUs using the Primus framework.

Principles

• MLPerf results require rigorous, multi-run validation.
• Unified frameworks simplify large-scale model training.
• Hardware-specific configurations optimize performance.

Method

The reproduction method involves setting up a Docker environment with ROCm 7.2.2+, preparing specific datasets, configuring platform-specific hyperparameters, and executing training runs, followed by MLPerf-compliant result validation.

In practice

• Use primus-cli for LLM pretraining/fine-tuning.
• Download pre-tokenized datasets for efficiency.
• Average 8 of 10 runs for MLPerf scores.

Topics

• MLPerf Training v6.0
• AMD Instinct GPUs
• Primus Training Framework
• LLM Training
• Text-to-Image Models
• Benchmark Reproduction

Code references

• AMD-AGI/Primus
• mlcommons/training
• mlcommons/training_policies
• mlcommons/logging

Best for: Machine Learning Engineer, AI Engineer, AI Hardware Engineer

Related on AIssential

• Reproducing the AMD MLPerf Inference v6.0 Submission Result — AMD provides a guide to reproduce its MLPerf Inference v6.0 results on Instinct MI355X platforms.
• DeepSeek-V3: A Large-Scale MoE Pretraining Benchmark for MLPerf Training v6.0 — MLPerf Training v6.0 now benchmarks MoE LLMs, reflecting industry shifts to sparse computation and specialized architectures.
• Technical Dive into AMD’s MLPerf Training v6.0 Submission — AMD's MLPerf v6.0 submission validates MXFP4 training, Primus framework, and multi-node scalability for competitive AI training on Instinct GPUs.
• Improved performance and model support with GGUF — Ollama 0.30 significantly boosts performance and broadens model/hardware compatibility via GGUF and llama.cpp integration.
• NVIDIA Extreme Co-Design Delivers New MLPerf Inference Records — Co-designed hardware, software, and models are crucial for maximizing AI factory throughput and minimizing token cost.
• A new GPT-OSS benchmark and DeepSeek R1 updates for latency-optimized reasoning — MLPerf Inference v6.0 introduces new benchmarks for open LLMs, separating performance and accuracy, and standardizing speculative decoding.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AMD ROCm Blogs.