Charlene Yang

Felix Abecassis, Anjulie Agrusa, Dong Ahn et al. · nvidia

Large Language Models (LLMs) today are powerful problem solvers across many domains, and they continue to get stronger as they scale in model size, training set size, and training set quality, as shown by extensive research and experimentation across the industry. Training a frontier model today requires on the order of tens to hundreds of yottaflops, which is a massive investment of time, compute, and energy. Improving pretraining efficiency is therefore essential to enable the next generation of even more capable LLMs. While 8-bit floating point (FP8) training is now widely adopted, transitioning to even narrower precision, such as 4-bit floating point (FP4), could unlock additional improvements in computational speed and resource utilization. However, quantization at this level poses challenges to training stability, convergence, and implementation, notably for large-scale models trained on long token horizons. In this study, we introduce a novel approach for stable and accurate training of large language models (LLMs) using the NVFP4 format. Our method integrates Random Hadamard transforms (RHT) to bound block-level outliers, employs a two-dimensional quantization scheme for consistent representations across both the forward and backward passes, utilizes stochastic rounding for unbiased gradient estimation, and incorporates selective high-precision layers. We validate our approach by training a 12-billion-parameter model on 10 trillion tokens -- the longest publicly documented training run in 4-bit precision to date. Our results show that the model trained with our NVFP4-based pretraining technique achieves training loss and downstream task accuracies comparable to an FP8 baseline. These findings highlight that NVFP4, when combined with our training approach, represents a major step forward in narrow-precision LLM training algorithms.

Charlene Yang, Yunsong Wang, Steven Farrell et al.

This paper presents a practical methodology for collecting performance data necessary to conduct hierarchical Roofline analysis on NVIDIA GPUs. It discusses the extension of the Empirical Roofline Toolkit for broader support of a range of data precisions and Tensor Core support and introduces a Nsight Compute based method to accurately collect application performance information. This methodology allows for automated machine characterization and application characterization for Roofline analysis across the entire memory hierarchy on NVIDIA GPUs, and it is validated by a complex deep learning application used for climate image segmentation. We use two versions of the code, in TensorFlow and PyTorch respectively, to demonstrate the use and effectiveness of this methodology. We highlight how the application utilizes the compute and memory capabilities on the GPU and how the implementation and performance differ in two deep learning frameworks.

Yunsong Wang, Charlene Yang, Steven Farrell et al.

Deep learning applications are usually very compute-intensive and require a long run time for training and inference. This has been tackled by researchers from both hardware and software sides, and in this paper, we propose a Roofline-based approach to performance analysis to facilitate the optimization of these applications. This approach is an extension of the Roofline model widely used in traditional high-performance computing applications, and it incorporates both compute/bandwidth complexity and run time in its formulae to provide insights into deep learning-specific characteristics. We take two sets of representative kernels, 2D convolution and long short-term memory, to validate and demonstrate the use of this new approach, and investigate how arithmetic intensity, cache locality, auto-tuning, kernel launch overhead, and Tensor Core usage can affect performance. Compared to the common ad-hoc approach, this study helps form a more systematic way to analyze code performance and identify optimization opportunities for deep learning applications.