AMD: Instinct MI355X in MLPerf Training v6.0 Within 5% of NVIDIA, 3.5× Faster Than Previous Generation
AMD's MLPerf Training v6.0 results show that the Instinct MI355X is within approximately 5% of an equivalent NVIDIA GPU's performance on LLM benchmarks. MI355X is 3.5× faster than last year's MI300X and 13–19% faster than the previous round. AMD introduced MXFP4 (FP4) training recipes and the Primus unified framework for the first time, alongside a multi-node submission of 512 MI300X GPUs across 64 nodes.
This article was generated using artificial intelligence from primary sources.
AMD published results in MLPerf Training v6.0 showing that its Instinct MI355X has closed the gap with NVIDIA on key large language model training benchmarks.
How close is MI355X to NVIDIA?
According to AMD’s measurements, MI355X is within approximately 5% of an equivalent NVIDIA GPU’s performance on both LLM benchmarks in round v6.0. This is the narrowest gap to date and a signal that AMD is becoming a more serious alternative for training workloads. MLPerf Training is a standardized benchmark suite measuring model training time to a target accuracy.
What is the improvement over predecessors?
MI355X is 3.5× faster than last year’s MI300X on the same benchmarks, and 13–19% faster than the previous round (v5.1) on tasks such as Llama 2 70B LoRA and Llama 3.1 8B. AMD introduced MXFP4 training recipes for the first time — a 4-bit format that reduces memory and compute requirements — along with the new Primus unified training framework.
What does this mean for the AI hardware market?
The multi-node submission covered 512 MI300X GPUs across 64 nodes (with OCI), showing AMD’s coverage at high scale. The same-day publication alongside NVIDIA’s MLPerf sweep intensifies competition: a narrower gap and FP4 training make AMD more competitive in data centers looking for an alternative to the NVIDIA stack.
Frequently Asked Questions
- How close is MI355X to NVIDIA?
- Within approximately 5% of an equivalent NVIDIA GPU's performance on both LLM benchmarks in MLPerf Training v6.0.
- How much faster is MI355X than its predecessor?
- 3.5× faster than last year's MI300X and 13–19% faster than the previous round (v5.1).
- What is MXFP4?
- A 4-bit number format (FP4) that AMD uses for training for the first time, reducing memory and compute requirements.