PostHeaderSection

Symbol

Global NavBar

PostContentSection

Global Footer

Blog Detail Page

Byron Anderson

AMD’s MLPerf Training 6.0 results highlight MI350 Series performance, MXFP4, ROCm, Primus, and advances in scalable, multi-node AI training.

AMD MLPerf Training 6.0: Breaking Down the MI350 Series Results

AMD’s recent MLPerf Training 6.0 results showcased their MI350 series GPU performance, and the results were staggering. Testing both their MI350X and MI355X GPUs, AMD delivered their most comprehensive training entry to date. As demand grows for larger models that can be trained more quickly, AMD's latest testing demonstrates the value of a platform designed to support them at scale. 

With the introduction of production-ready MXFP4, the debut of its Primus training framework, and its first multi-node MLPerf submission, AMD demonstrated how much it has advanced its AI training platform.


Performance Gains on LLM Workloads

AMD's Instinct MI350 series GPUs offer 288GB of HBM3E memory, up to 10PF of MXFP4 performance, and a peak memory bandwidth of 8TB/s. The MI350 series was built on AMD's CDNA 4 architecture, optimized for exascale AI, machine learning, and high performance computing. 

Offered in the industry standard UBB8 node,  the MI350 series comes in both air cooled and direct to chip water cooled variants. That combination of capacity and  bandwidth allows for larger models and batch sizes before having to be split across multiple GPUs or nodes, while keeping compute units fed and reducing idle GPU cycles. And as the industry shifts toward lower-precision formats like MXFP4 to speed up large-scale model training, the MI350 series' 10PF of throughput translates directly into faster training runs and lower compute costs.

Since their first training submission with the MI300X GPU, AMD’s MI355X saw a 3.5x improvement in performance switching from MXFP8 in MLPerf Training 5.0 to MXFP4 in MLPerf Training 6.0. AMD noted that this doesn’t just reflect the hardware upgrades between the models but the advancements of the whole stack including the ROCm ecosystem and support of the MXFP4 datatype. The MI355X also demonstrated close performance with the NVIDIA B200 on two MLPerf training workloads, coming within 5% on Llama 2-70B fine-tuning and 6% on Llama 3.1-8B pre-training. 


Scaling beyond a Single Node

They trained FLUX.1 , a transformer-based text-to-image model, across both a 64- and a 512-GPU cluster. Being AMD’s first application of a multi-node training format, the results demonstrated performance more relevant to what larger models and higher resolution workloads demand, compared to single-node training benchmarks.

For enterprises, scale is often the difference between a strong benchmark and a deployable training platform. With the MI300X tested at this scale, AMD matched NVIDIA’s largest FLUX.1 / GPT-OSS-scale submission, showing their confidence in the ability of their full stack, not just the hardware. Raw chip performance means little without drivers, interconnects, and distributed training frameworks that scale effectively alongside it. As enterprises increasingly demand jobs requiring large amounts of GPUs, the ability to prove that performance is becoming an expectation, not a differentiator. 


ROCm and Primus

The ROCm software was used to convert the GPU’s capabilities into training throughput and allow these results to be replicable by AMD partners. The training required AMD partners to run ROCm 7.2.2 or later versions to replicate AMD's results as closely as possible. AMD credits ROCm with optimizing the training throughput through precision support and model convergence work, kernel and compiler optimization through tuned GEMM, and memory and communication efficiency through better bandwidth utilization and improved compute/communication overlap. 

The training also utilized AMD’s Primus for the first time ever in an MLPerf training environment. Primus is a large-model training framework supporting both Megatron-LM and TorchTitan backends. Its use in training was threefold: 

 * To orchestrate the workflow, including pre-training and fine-tuning
 * To optimize GEMM/FlashAttention/communication ops with AITER, CK, and ROCm Triton
 * To transform the GPU server fleets into efficient, self-monitoring training environments

The Primus system allows for the same frameworks, ROCm software recipes, and replicable experiences, across the full training pipeline. AMD noted that, “ROCm software tuning contributed to the double-digit uplift on AMD Instinct MI350 Series platforms..”


Reproducibility and Partner Validation

The results that AMD produced across Llama 2-70B LoRA fine-tuning and Llama 3.1-8B pre-training workloads were also replicated within 6% by 10 of their partners - proving that these standings are replicable and not dependent on the internal reference configuration. The advancements of the ROCm ecosystem also contributed to the success of the testing, with partners noting that similar results to AMD’s would have proved difficult without predictable kernels, tuned libraries and a software path that can be followed. 

In a joint submission between Dell and Mangoboost, testing proved a 95.5% scaling efficiency across three different AMD GPU generations. The test used clusters in the U.S. and South Korea, combining the GPUs into a single live pipeline, and still held the efficiency mark despite the expected wide-area latency. This was the first heterogeneous, multi-region submission in MLPerf inference, and it proved that newer generation AMD hardware can be implemented with older models, or even across separate data centers.


Strategic Takeaway

AMD’s MLPerf Training 6.0 results show that its AI strategy is broadening to become less about individual GPU performance and more about delivering a scalable, reproducible, full-stack training platform. With the features that the MI350 series offers and software like ROCm and Primus running it, AMD is demonstrating that it can provide competitive results in full stack applications, not just chip for chip. 

These results matter beyond the benchmark tables, because they speak to AMD's position against NVIDIA, the incumbent leader in AI training infrastructure. The MLPerf Training 6.0 signals a shift in the terms of training, using multiple nodes for the submission and with AMD demonstrating that it can compete on criteria that matter most to enterprise buyers: reproducibility, scalability, and deployment readiness, not just raw hardware specs. 

TensorWave Cloud: Engineered for AI VelocityTensorWave is the industry’s premier AMD-exclusive cloud provider, delivering memory optimized AI infrastructure for training, fine-tuning and inference. 

Our platform combines the raw computational power of AMD's latest GPU architectures with innovative liquid cooling technology and ultra-high-speed networking, creating an optimized environment for the most demanding AI workloads. From large language model training to real-time inference, TensorWave provides the performance, scalability, and cost-efficiency that modern AI initiatives demand.

Ready to get started? Connect with a sales engineer.

Breaking Down AMD's MLPerf Training 6.0 Results

Product

Solutions

Resources

Company

Product

Resources

Solutions

Company