Feiyang Kang, Newsha Ardalani, Michael Kuchnik, Youssef Emad, Mostafa Elhoushi, Shubhabrata Sengupta, Shang-Wen Li, Ramya Raghavendra, Ruoxi Jia, Carole-Jean Wu. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Machine learning solutions are rapidly adopted to enable a variety of key use cases, from conversational AI assistants to scientific discovery. As the adoption of machine learning models becomes increasingly prevalent, the associated lifecycle carbon footprint is expected to increase, including both *operational carbon* from training and inference and *embodied carbon* from AI hardware manufacturing. We introduce CATransformers, the first carbon-aware co-optimization framework for Transformer-based models and hardware accelerators. By integrating both operational and embodied carbon into early-stage design space exploration, CATransformers enables sustainability-driven model architecture and hardware accelerator co-design that reveals fundamentally different trade-offs than latency- or energy-centric approaches. Evaluated across a range of Transformer models, CATransformers consistently demonstrates the potential to reduce total carbon emissions --by up to 30\% -- while maintaining accuracy and latency. We further highlight its extensibility through a focused case study on multi-modal models. Our results emphasize the need for holistic optimization methods that prioritize carbon efficiency without compromising model capability and execution time performance. The source code of CATransformers is available at https://github.com/facebookresearch/CATransformers.

Vision-Language Models (VLMs) are pretrained on large, diverse, and noisy web-crawled datasets. This underscores the critical need for dataset pruning, as the quality of these datasets is strongly correlated with the performance of VLMs on downstream tasks. Using CLIPScore from a pretrained model to only train models using highly-aligned samples is one of the most successful methods for pruning. We argue that this approach suffers from multiple limitations including: false positives and negatives due to CLIP's pretraining on noisy labels. We propose a pruning signal, Sieve, that employs synthetic captions generated by image-captioning models pretrained on small, diverse, and well-aligned image-text pairs to evaluate the alignment of noisy image-text pairs. To bridge the gap between the limited diversity of generated captions and the high diversity of alternative text (alt-text), we estimate the semantic textual similarity in the embedding space of a language model pretrained on unlabeled text corpus. Using DataComp, a multimodal dataset filtering benchmark, when evaluating on 38 downstream tasks, our pruning approach, surpasses CLIPScore by 2.6% and 1.7% on medium and large scale respectively. In addition, on retrieval tasks, Sieve leads to a significant improvement of 2.7% and 4.5% on medium and large scale respectively.

Training and deploying large-scale machine learning models is time-consuming, requires significant distributed computing infrastructures, and incurs high operational costs. Our analysis, grounded in real-world large model training on datacenter-scale infrastructures, reveals that 14~32% of all GPU hours are spent on communication with no overlapping computation. To minimize this outstanding communication latency and other inherent at-scale inefficiencies, we introduce an agile performance modeling framework, MAD-Max. This framework is designed to optimize parallelization strategies and facilitate hardware-software co-design opportunities. Through the application of MAD-Max to a suite of real-world large-scale ML models on state-of-the-art GPU clusters, we showcase potential throughput enhancements of up to 2.24 × for pretraining and up to 5.27 × for inference scenarios, respectively.

Mixture-of-Experts (MoE) models have recently gained steam in achieving the state-of-the-art performance in a wide range of tasks in computer vision and natural language processing. They effectively expand the model capacity while incurring a minimal increase in computation cost during training. However, deploying such models for inference is difficult due to their large model size and complex communication pattern. In this work, we provide a characterization of two MoE workloads, namely Language Modeling (LM) and Machine Translation (MT) and identify their sources of inefficiencies at deployment. We propose three optimization techniques to mitigate sources of inefficiencies, namely (1) Dynamic gating, (2) Expert Buffering, and (3) Expert load balancing. We show that dynamic gating improves maximum throughput by 6.21-11.55$\times$ for LM, 5.75-10.98$\times$ for MT Encoder and 2.58-5.71$\times$ for MT Decoder. It also reduces memory usage by up to 1.36$\times$ for LM and up to 1.1$\times$ for MT. We further propose Expert Buffering, a new caching mechanism that only keeps hot, active experts in GPU memory while buffering the rest in CPU memory. This reduces static memory allocation by 1.47$\times$. Finally, we propose a load balancing methodology that provides additional robustness to the workload. Our code is available at https://github.com/hyhuang00/moe_inference.

Over the past decade, machine learning model complexity has grown at an extraordinary rate, as has the scale of the systems training such large models. However, there is an alarmingly low hardware utilization (5–20%) in large scale AI systems. The low system utilization is a cumulative effect of minor losses across different layers of the stack, exacerbated by the disconnect between engineers designing different layers spanning across different industries. To address this challenge, in this work we designed a cross-stack performance modelling and design space exploration framework. First, we introduce CrossFlow, a novel framework that enables cross-layer analysis all the way from the technology layer to the algorithmic layer. Next, we introduce DeepFlow (built on top of CrossFlow using machine learning techniques) to automate the design space exploration and co-optimization across different layers of the stack. We have validated CrossFlow’s accuracy with distributed training on real commercial hardware and showcase several DeepFlow case studies demonstrating pitfalls of not optimizing across the technology-hardware-software stack for what is likely the most important workload driving large development investments in all aspects of computing stack.

Deep learning recommendation systems serve personalized content under diverse tail-latency targets and input-query loads. In order to do so, state-of-the-art recommendation models rely on terabyte-scale embedding tables to learn user preferences over large bodies of contents. The reliance on a fixed embedding representation of embedding tables not only imposes significant memory capacity and bandwidth requirements but also limits the scope of compatible system solutions. This paper challenges the assumption of fixed embedding representations by showing how synergies between embedding representations and hardware platforms can lead to improvements in both algorithmic- and system performance. Based on our characterization of various embedding representations, we propose a hybrid embedding representation that achieves higher quality embeddings at the cost of increased memory and compute requirements. To address the system performance challenges of the hybrid representation, we propose MP-Rec — a co-design technique that exploits heterogeneity and dynamic selection of embedding representations and underlying hardware platforms.

Machine learning research has long focused on models rather than datasets, and prominent datasets are used for common ML tasks without regard to the breadth, difficulty, and faithfulness of the underlying problems. Neglecting the fundamental importance of data has given rise to inaccuracy, bias, and fragility in real-world applications, and research is hindered by saturation across existing dataset benchmarks. In response, we present DataPerf, a community-led benchmark suite for evaluating ML datasets and data-centric algorithms. We aim to foster innovation in data-centric AI through competition, comparability, and reproducibility. We enable the ML community to iterate on datasets, instead of just architectures, and we provide an open, online platform with multiple rounds of challenges to support this iterative development. The first iteration of DataPerf contains five benchmarks covering a wide spectrum of data-centric techniques, tasks, and modalities in vision, speech, acquisition, debugging, and diffusion prompting, and we support hosting new contributed benchmarks from the community. The benchmarks, online evaluation platform, and baseline implementations are open source, and the MLCommons Association will maintain DataPerf to ensure long-term benefits to academia and industry.

Deep learning (DL) research yields accuracy and product improvements from both model architecture changes and scale: larger data sets and models, and more computation. For hardware design, it is difficult to predict DL model changes. However, recent prior work shows that as dataset sizes grow, DL model accuracy and model size grow predictably. This paper leverages the prior work to project the dataset and model size growth required to advance DL accuracy beyond human-level, to frontier targets defined by machine learning experts. Datasets will need to grow 33--971×, while models will need to grow 6.6--456× to achieve target accuracies.

Recent programmable accelerators are faster and more energy efficient than general purpose processors, but expose complex hardware/software abstractions for compilers. A key problem is instruction scheduling, which requires sophisticated algorithms for mapping instructions to distributed processing elements, routing of operand dependences, and timing the arrival of operands to enable high throughput.

Demand for low-power data processing hardware continues to rise inexorably. Existing programmable and "general purpose" solutions (eg. SIMD, GPGPUs) are insufficient, as evidenced by the order-of-magnitude improvements and industry adoption of application and domain-specific accelerators in important areas like machine learning, computer vision and big data. The stark tradeoffs between efficiency and generality at these two extremes poses a difficult question: how could domain-specific hardware efficiency be achieved without domain-specific hardware solutions?

GPUs have become prevalent and more general purpose, but GPU programming remains challenging and time consuming for the majority of programmers. In addition, it is not always clear which codes will benefit from getting ported to GPU. Therefore, having a tool to estimate GPU performance for a piece of code before writing a GPU implementation is highly desirable. To this end, we propose Cross-Architecture Performance Prediction (XAPP), a machine-learning based technique that uses only single-threaded CPU implementation to predict GPU performance.