Cross-Hardware Energy Benchmarking of 16 Vision Architectures

16 deep learning architectures, including CNNs, Vision Transformers, and Hybrid models, are benchmarked for energy consumption. The dataset contains empirical measurements of fine-tuning training energy and inference energy on NVIDIA A100 GPU and AMD EPYC CPU platforms. It was created by Abderrahmen Jedidi and last updated in May 2026.

Use Cases

• Compare accuracy-energy trade-offs across architecture families based on the described benchmarking.
• Select hardware-aware models for energy-efficient deployment based on the reported cross-hardware ranking inconsistencies.
• Validate theoretical compute complexity metrics against practical energy consumption using the provided empirical measurements.
• Benchmark lightweight CNNs like MobileNetV2 and EfficientNet-B0 for CPU inference based on the described results.

Strengths

• Benchmarks 16 distinct deep learning architectures.
• Measures energy consumption on two hardware platforms: NVIDIA A100 GPU and AMD EPYC CPU.
• Employs synthetic data to isolate architectural and hardware factors, as described.
• Reports results with 95% confidence intervals using adaptive multi-run sampling.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Uses synthetic data, which may not reflect performance on real-world datasets.

Provenance

Source

figshare

Collection Method

Empirical measurement using CodeCarbon, with adaptive multi-run sampling.

Freshness

Last updated 2026-05-29 18:31:54; freshness should be verified.