High-Resolution AI Training Workload Traces from H100, B200, and RTX 3060 GPUs

Over 1.8 million sub-second samples capture system-level metrics from 72 AI training sessions on NVIDIA H100, B200, and GeForce RTX 3060 GPUs. The dataset provides 15-minute time-series recordings at 20 ms resolution for multi-GPU nodes and 100 ms for single-machine setups. It covers diverse AI applications including image generation, LLM text generation, and feature forecasting across different computational scales.

Use Cases

• Modeling power consumption and thermal dynamics based on high-resolution GPU and CPU traces.
• Benchmarking AI training efficiency across different hardware platforms (H100, B200, RTX 3060).
• Analyzing workload scheduling and resource allocation for tasks like LLM training and image generation.
• Developing predictive models for data center energy usage from time-series system metrics.

Strengths

• Contains over 1.8 million high-resolution samples, providing granular detail.
• Captures data from both high-performance (H100/B200) and consumer-grade (RTX 3060) GPU systems, enabling comparative analysis.
• Explicitly structured around three principal aspects: platform scale, applications, and AI architectures.

Limitations

• Column names and a precise row count are not provided, limiting immediate understanding of the data structure.
• The temporal coverage is limited to 15-minute sessions, which may not capture longer-term training cycle patterns.
• Metadata completeness is medium; specific AI model architectures and hyperparameter details are not enumerated.

Provenance

Source

Ahmed Abdelaziz Elsayed

Collection Method

High-resolution, sub-second measurements collected from executed AI training workloads on specified GPU systems.

Time Range

Each session is a 15-minute recording.

Freshness

Last updated on 2026-03-19.

Geography

null