Since 2018, the consortium MLCommons has been running a sort of Olympics for AI training. The competition, called MLPerf, consists of a set of tasks for training specific AI models, on predefined datasets, to a certain accuracy. Essentially, these tasks, called benchmarks, test how well a hardware and low-level software configuration is set up to train a particular AI model. Twice a year, companies put together their submissions—usually, clusters of CPUs and GPUs and software optimized for them—and compete to see whose submission can train the models fastest. There is no question that since MLPerf’s inception, the cutting-edge hardware for AI training has improved dramatically. Over the years, Nvidia has released four new generations of GPUs that have since become the industry standard (the latest, Nvidia’s Blackwell GPU, is not yet standard but growing in popularity). The companies competing in MLPerf have also been using larger clusters of GPUs to tackle the training tasks. However, the MLPerf benchmarks have also gotten tougher. And this increased rigor is by design—the benchmarks are trying to keep pace with the industry, says David Kanter, head of MLPerf. “The benchmarks are meant to be representative,” he says. Intriguingly, the data show that the large language models and their precursors have been increasing in size faster than the hardware has kept up. So each time a new benchmark is introduced, the fastest training time gets longer. Then, hardware improvements gradually bring the execution time down, only to get thwarted again by the next benchmark. Then the cycle repeats itself. _This article appears in the November 2025 print issue._

Welcome to your weekly drop of developer news. Subscribe for the latest technical deep dives, resources, trainings, and more.

Follow @NVIDIANewsroom for real-time updates from NVIDIA. NVIDIA Wins Every MLPerf Training v5.

NVIDIA、Llama 3.1 405Bモデルの学習を約10分で完了 – MLPerf Training v5.1テストで記録更新 #ITニュース

NVIDIA baru saja memamerkan pencapaiannya dalam hal komputasi AI, lewat hasil terbaru dari MLPerf Training v5.1. Dalam pengujian benchmark industri untuk pelatihan model AI itu, NVIDIA berhasil menyabet posisi tertinggi  di semua kategori,mulai dari large language model (LLM), image generation, hingga graph neural network. Menariknya, NVIDIA juga menjadi satu-satunya platform yang ikut di seluruh tes. […]

Introduction: The Race to Smarter AI Infrastructure In the ever-accelerating era of artificial intelligence, the race is no longer about who builds the largest model — it’s about who trains it fastest, most efficiently, and at scale. The demand for high-speed, high-efficiency computing has exploded as AI reasoning grows more complex and data-hungry. Every layer of technology — from GPUs and CPUs to networking and algorithms — is being pushed to its limits.

In the age of AI reasoning, training smarter, more capable models is critical to scaling intelligence. Delivering the massive performance to meet this new age requires breakthroughs across GPUs, CPUs, NICs, scale-up and scale-out networking, system architectures, and mountains of software and algorithms. In MLPerf Training v5.1 — the latest round in a long-running series of industry-standard tests of AI training performance — NVIDIA swept all seven tests, delivering the fastest time to train across large language models (LLMs), image generation, recommender systems, computer vision and graph neural networks. NVIDIA was also the only platform to submit results on every test, underscoring the rich programmability of NVIDIA GPUs, and the maturity and versatility of its CUDA software stack. ## **NVIDIA Blackwell Ultra Doubles Down** The GB300 NVL72 rack-scale system, powered by the NVIDIA Blackwell Ultra GPU architecture, made its debut in MLPerf Training this round, following a record-setting showing in the most recent MLPerf Inference round. Compared with the prior-generation Hopper architecture, the Blackwell Ultra-based GB300 NVL72 delivered more than 4x the Llama 3.1 405B pretraining and nearly 5x the Llama 2 70B LoRA fine-tuning performance using the same number of GPUs. These gains were fueled by Blackwell Ultra’s architectural improvements — including new Tensor Cores that offer 15 petaflops of NVFP4 AI compute, twice the attention-layer compute and 279GB of HBM3e memory — as well as new training methods that tapped into the architecture’s enormous NVFP4 compute performance. Connecting multiple GB300 NVL72 systems, the NVIDIA Quantum-X800 InfiniBand platform — the industry’s first end-to-end 800 Gb/s scale-up networking platform — also made its MLPerf debut, doubling scale-out networking bandwidth compared with the prior generation. ## **Performance Unlocked: NVFP4 Accelerates LLM Training** Key to the outstanding results this round was performing calculations using NVFP4 precision — a first in the history of MLPerf Training. One way to increase compute performance is to build an architecture capable of performing computations on data represented with fewer bits, and then to perform those calculations at a faster rate. However, lower precision means less information is available in each calculation. This means using low-precision calculations in the training process calls for careful design decisions to keep results accurate. NVIDIA teams innovated at every layer of the stack to adopt FP4 precision for LLM training. The NVIDIA Blackwell GPU can perform FP4 calculations — including the NVIDIA-designed NVFP4 format as well as other FP4 variants — at double the rate of FP8. Blackwell Ultra boosts that to 3x, enabling the GPUs to deliver substantially greater AI compute performance. NVIDIA is the only platform to date that has submitted MLPerf Training results with calculations performed using FP4 precision while meeting the benchmark’s strict accuracy requirements. ## **NVIDIA Blackwell Scales to New Heights** NVIDIA set a new Llama 3.1 405B time-to-train record of just 10 minutes, powered by more than 5,000 Blackwell GPUs working together efficiently. This entry was 2.7x faster than the best Blackwell-based result submitted in the prior round, resulting from efficient scaling to more than twice the number of GPUs, as well as the use of NVFP4 precision to dramatically increase the effective performance of each Blackwell GPU. To illustrate the performance increase per GPU, NVIDIA submitted results this round using 2,560 Blackwell GPUs, achieving a time to train of 18.79 minutes — 45% faster than the submission last round using 2,496 GPUs. ## **New Benchmarks, New Records** NVIDIA also set performance records on the two new benchmarks added this round: Llama 3.1 8B and FLUX.1. Llama 3.1 8B — a compact yet highly capable LLM — replaced the long-running BERT-large model, adding a modern, smaller LLM to the benchmark suite. NVIDIA submitted results with up to 512 Blackwell Ultra GPUs, setting the bar at 5.2 minutes to train. In addition, FLUX.1 — a state-of-the-art image generation model — replaced Stable Diffusion v2, with only the NVIDIA platform submitting results on the benchmark. NVIDIA submitted results using 1,152 Blackwell GPUs, setting a record time to train of 12.5 minutes. NVIDIA continued to hold records on the existing graph neural network, object detection and recommender system tests. ## **A Broad and Deep Partner Ecosystem** The NVIDIA ecosystem participated extensively this round, with compelling submissions from 15 organizations including ASUSTeK, Dell Technologies, Giga Computing, Hewlett Packard Enterprise, Krai, Lambda, Lenovo, Nebius, Quanta Cloud Technology, Supermicro, University of Florida, Verda (formerly DataCrunch) and Wiwynn. NVIDIA is innovating at a one-year rhythm, driving significant and rapid performance increases across pretraining, post-training and inference — paving the way to new levels of intelligence and accelerating AI adoption. _See more NVIDIA performance data on the_ _Data Center Deep Learning Product Performance Hub_ _and_ _Performance Explorer_ _pages._ Categories: Data Center | Hardware | Networking | Software Tags: Artificial Intelligence

Two new genAI tests (Llama 3.1 8B, Flux.1) align with production stacks as multi-node results climb. NVIDIA posts many fastest times; University of Florida, Wiwynn, and Datacrunch expand the ecosystem...

In MLPerf Training v5.1, NVIDIA swept all seven tests, delivering the fastest time to train across LLMs, image generation, recommender systems, computer vision and graph neural networks.

Während Nvidia einmal mehr versuchte, die neuen MLPerf-Training-Ergebnisse für sich zu nutzen, glänzt auch AMD stetig mehr.

AI training races are heating up as benchmarks get tougher.

AI training races are heating up as benchmarks get tougher.