Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
CATransformers: Carbon Aware Transformers Through Joint Model-Hardware Optimization
Authors: Irene Wang, Mostafa Elhoushi, H Ekin Sumbul, Samuel Hsia, Daniel Jiang, Newsha Ardalani, Divya Mahajan, Carole-Jean Wu, Bilge Acun
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 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. (Abstract) and 4 Evaluation (Section Title) |
| Researcher Affiliation | Collaboration | Irene Wang1,2, , Mostafa Elhoushi2, H. Ekin Sumbul3, Samuel Hsia2, Daniel Jiang4, Newsha Ardalani2, Divya Mahajan1, Carole-Jean Wu2, Bilge Acun2 1Georgia Institute of Technology, 2FAIR at Meta, 3Reality Labs at Meta, 4Meta |
| Pseudocode | No | The paper describes methods in prose and diagrams (e.g., Figure 2: Overview of the CATransformers framework) but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The source code of CATransformers is available at https://github.com/facebookresearch/CATransformers. |
| Open Datasets | Yes | For language models we use the MRPC task from the GLUE dataset [WSM+18] for semantic similarity detection, for vision models, the CIFAR-10 dataset [Kri09] for image classification; and for CLIP models, the MSCOCO dataset [LMB+15] for retrieval... and adopt Open CLIP s [IWW+21] Data Comp-1B [GIF+23] models for CLIP. |
| Dataset Splits | Yes | For language models we use the MRPC task from the GLUE dataset [WSM+18] for semantic similarity detection, for vision models, the CIFAR-10 dataset [Kri09] for image classification; and for CLIP models, the MSCOCO dataset [LMB+15] for retrieval. |
| Hardware Specification | Yes | Bayesian optimization is performed on a single node (8 V100 GPUs, 80 CPUs) over 100 trials... We further validate energy and latency predictions against real GPU hardware (V100, A100, H100) by modeling GPU-like architectures. |
| Software Dependencies | No | The paper mentions several software tools and frameworks used (e.g., Hugging Face, Open CLIP, Ax, BoTorch, Accelergy, Sunstone, ACT, Electricity Maps), but does not provide specific version numbers for these software dependencies as required for reproducible description. |
| Experiment Setup | Yes | Bayesian optimization is performed on a single node (8 V100 GPUs, 80 CPUs) over 100 trials, taking 5-20 hours depending on the model. Post-pruning training of Carbon CLIP uses 224 GPUs on the Meta CLIP-2.5B dataset [XXT+24], with a batch size of 128, learning rate of 5 10^4, and 2 distillation epochs... We evaluate different optimization modes under a 20 TOPS compute budget... When latency is not an optimization target, a maximum latency constraint of 50ms is enforced... |