Energaizer Gpu Power Eval

energaizer-gpu-power-eval

EnergAIzer: Fast and Accurate GPU Power Estimation Framework for AI Workloads — Lee et al. (2026) (arXiv:2604.20105, 2026)

What this evaluates

Evaluates the accuracy of a lightweight analytical framework in predicting GPU latency and dynamic power consumption for AI workloads across different hardware architectures, operating frequencies, and algorithm configurations.

Datasets

• EnergAIzer Kernel Database & AI Workloads — total ?; splits: offline_database (-1), held_out_test (-1); repo https://github.com/kyungmi-lee/energaizer-ispass26-artifact

Metrics

• MAPE (primary) — range: percent
  • Mean Absolute Percentage Error: (1/N) * Σ |(Actual - Predicted) / Actual| * 100%. Applied independently to latency and power predictions across workloads or kernels.

Input / output format

Input: Kernel specifications (type, library, tensor shape, precision, operating frequency) or end-to-end workload configurations (model architecture, batch size, sequence length, precision, backend settings).

Output: Estimated latency (seconds) and estimated power consumption (watts).

Scoring recipe

def compute_mape(actual, predicted):
    mask = actual != 0
    return np.mean(np.abs((actual[mask] - predicted[mask]) / actual[mask])) * 100

latency_mape = compute_mape(measured_latency, estimated_latency)
power_mape = compute_mape(measured_power, estimated_power)
avg_latency_mape = np.mean(latency_mapes)
avg_power_mape = np.mean(power_mapes)

Common pitfalls

• CPU kernel launch overhead significantly impacts latency estimation for small kernels, inflating end-to-end errors.
• The model assumes sequential kernel execution and consistent energy efficiency across architectures, failing for overlapping kernels or different memory technologies (e.g., GDDR6 vs HBM).
• Errors exceeding 40% are clipped in reported figures, potentially masking worst-case performance on specific configurations.

Evidence (verbatim from paper)

EnergAIzer achieves 11.0% latency and 8.0% power error on the server-grade A100-40GB-PCIE, and 8.8% latency and 8.2% power error on the workstation-grade A10, averaged across all workloads. Our framework captures distinct scaling behaviors of low-utilization (small batch/sequence, left panel) and power-capped (large batch/sequence, right panel) workloads, achieving 6-9% MAPE across frequencies.

Citation

@misc{lee2026energaizer,
  title={EnergAIzer: Fast and Accurate GPU Power Estimation Framework for AI Workloads},
  author={Lee et al. (2026)},
  year={2026},
  note={arXiv:2604.20105}
}

• arXiv: 2604.20105