ehyo/GPU-Resources-Estimation-for-Deep-Learning-Training-Tasks

GPUMemNet and GPUUtilNet Dataset

This dataset accompanies the paper “GPU Memory and Utilization Estimation for Training-Aware Resource Management: Opportunities and Limitations.”

It contains synthetic deep learning training configurations and their measured GPU memory consumption and utilization characteristics.

Dataset configurations

The dataset is divided into separate configurations because MLP, CNN, and Transformer workloads use different feature schemas.

Configuration	Rows	Description
mlp-memory-step1	3,000	Initial MLP memory and average-utilization measurements
mlp-memory-step2	3,000	MLP measurements with batch-normalization and dropout features
mlp-utilization	3,000	MLP average and maximum utilization measurements
mlp-memory-legacy	1,091	Earlier fully connected network memory dataset
cnn-memory-step1	9,000	CNN measurements including architecture identifiers
cnn-memory-revised	9,000	Revised CNN representation
cnn-utilization	9,000	CNN average and maximum utilization measurements
transformer-memory	5,011	Transformer memory measurements
transformer-utilization	5,011	Transformer average and maximum utilization measurements

Prediction targets

The primary GPU-memory prediction target is:

• Max GPU Memory (MiB)

The legacy MLP configuration uses:

• gpumemory_max

The utilization configurations contain average and maximum values for:

• GPUTL
• GRACT
• SMACT
• SMOCC
• FP32A
• DRAMA

Loading

Install the Hugging Face datasets package:

pip install datasets

Load a specific configuration:

from datasets import load_dataset

dataset = load_dataset(
    "ehyo/GPU-Resources-Estimation-for-Deep-Learning-Training-Tasks",
    "cnn-utilization",
)

Each configuration currently provides a train split containing the complete corresponding table.

Dataset characteristics

The workload families contain different features:

• MLP: depth, activation function, activation counts, parameter counts, batch size, batch-normalization layers, and dropout layers.
• CNN: depth, activation function, layer-type counts, batch size, parameter counts, activation counts, and analytical memory estimates.
• Transformer: sequence length, embedding size, number of layers, number of attention heads, parameter counts, activation counts, and layer-type counts.

The configurations should be loaded independently because their schemas are workload-family specific.

Data collection

The measurements were collected from generated deep learning training workloads under controlled execution conditions. The original column names and units are preserved for compatibility with the accompanying code and published experiments.

Further implementation and experimental details are available in the paper and GitHub repository.

Repository

Code, models, and reproducibility artifacts are available at:

https://github.com/itu-rad/GPUMemNet

Citation

@inproceedings{yousefzadehaslmiandoab2026gpumemory,
  author    = {Ehsan Yousefzadeh-Asl-Miandoab and
               Reza Karimzadeh and
               Danyal Yorulmaz and
               Bulat Ibragimov and
               Pınar Tözün},
  title     = {GPU Memory and Utilization Estimation for Training-Aware
               Resource Management: Opportunities and Limitations},
  booktitle = {Proceedings of the Sixth European Workshop on Machine
               Learning and Systems},
  series    = {EuroMLSys '26},
  pages     = {127--138},
  publisher = {Association for Computing Machinery},
  year      = {2026},
  doi       = {10.1145/3805621.3807621}
}

License

This dataset is licensed under the Creative Commons Attribution 4.0 International License.