Local AI & Edge

AMD expands MLPerf Training v6.1 to five benchmarks, MI355X matches NVIDIA B300

Key Takeaway

AMD's most complete MLPerf Training submission to date demonstrates that MI355X GPUs with ROCm software improvements achieve competitive time-to-train against NVIDIA's Blackwell generation, including the first-ever MXFP6 training submission.

Summary

On November 16, 2026, AMD submitted results for five of seven MLPerf Training v6.1 benchmarks, its broadest coverage yet on AMD Instinct GPUs. The submission added GPT-OSS-20B MoE pretraining and DLRM v4 recommendation model to returning workloads including Llama 2 70B LoRA, Llama 3.1 8B, and Flux.1 image generation. All first-party results ran on CDNA 4-based MI355X and MI350X GPUs.

The results show AMD matching or beating comparable NVIDIA B200 submissions and coming within a few points of B300 on identical benchmarks. On repeated workloads from v6.0, identical MI355X hardware achieved approximately 1.19x faster time-to-train through software improvements alone in the Primus framework and training recipes. Nine total submitters including Dell, HPE, Supermicro, Oracle, and Crusoe published AMD Instinct results, with Crusoe scaling to 128 GPUs.

Primus framework and MI355X optimization details

  • Primus architecture: AMD's unified training framework consists of three components: Primus-LM (training framework with primus-cli, configuration templates, and ROCm-native Megatron-LM/TorchTitan backends), Primus-Turbo (optimized FlashAttention, GEMM, and collective communication kernels via AITER, Composable Kernel, and ROCm Triton), and Primus-SaFE (cluster management and fault tolerance layer).
  • MXFP4 training recipe for GPT-OSS-20B: Weights use 32x32 blocks with shared scale, activations use 1x32 blocks, paired Hadamard transforms redistribute outliers, output gradients use stochastic rounding, and weight deoscillation prevents repeated quantization flips; only the Universal Optimal Scaling method from MXAttention is adopted for scale selection.
  • GPT-OSS-20B MoE optimizations: Primus-Turbo provides FlyDSL grouped and dense GEMMs, specialized BF16 attention, SDMA collective transfers, and fused routing, normalization, optimizer, and loss operations to reduce memory traffic.
  • First MXFP6 MLPerf Training submission: This round marks the first use of six-bit floating-point precision in MLPerf Training history, pushing low-precision training while maintaining benchmark convergence targets.
  • Flux.1 scaling improvement: Training throughput improved 3.1x on 64 GPUs compared to previous results, demonstrating effective scaling for image generation workloads.

Why it matters

The breadth of partner and independent submissions indicates ROCm has matured from AMD-only reference implementations to a production ecosystem that OEMs and cloud providers will deploy at scale. For Linux-based AI infrastructure operators, the Primus framework's availability on GitHub under open-source terms provides a reproducible path to match published benchmark results without proprietary training stacks, while the MXFP4/MXFP6 precision advances reduce memory bandwidth requirements that often bottleneck large-model training on commodity clusters.

Original reference: rocm.blogs.amd.com →