# Proai Hardware Benchmark Eval

> Evaluates the power efficiency, throughput, and real-time inference performance of embedded AI hardware platforms running multitask and single-task deep neural networks for automotive vision tasks. Use when the user wants to benchmark on COCO test2017, or asks about evaluating this task. Reports FPS, inference time, memory usage, energy efficiency (Wtotal, W/fps).

- Skill: `qhjqhj00/proai-hardware-benchmark-eval` (Agent Skill)
- Install (CLI): `npx skillmds add qhjqhj00/proai-hardware-benchmark-eval`
- Raw SKILL.md: https://api.skillmd.com/api/skills/qhjqhj00/proai-hardware-benchmark-eval/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: qhjqhj00 (https://skillmd.com/u/qhjqhj00)
- Updated: 2026-09-08
- Page: https://skillmd.com/skills/qhjqhj00/proai-hardware-benchmark-eval

---


# proai-hardware-benchmark-eval

> ProAI: An Efficient Embedded AI Hardware for Automotive Applications -- a Benchmark Study — Mantowsky et al. (2021) (arXiv:2108.05170, 2021)

## What this evaluates

Evaluates the power efficiency, throughput, and real-time inference performance of embedded AI hardware platforms running multitask and single-task deep neural networks for automotive vision tasks.

## Datasets

- **COCO test2017** — total 41000; splits: test (5000)

## Metrics

- `FPS, inference time, memory usage, energy efficiency (Wtotal, W/fps)` **(primary)** — range: other
  - FPS = total_images / total_inference_time. Inference time = total_inference_time / total_images. Memory usage = peak logged RAM/VRAM. Energy efficiency = Wtotal (total power) and W/fps (power per frame).

## Input / output format

**Input**: RGB images from the COCO test2017 dataset (5,000 images), passed through standard pre-processing pipelines for object detection, semantic segmentation, and human pose estimation.

**Output**: Model predictions (bounding boxes, segmentation masks, pose keypoints) and system telemetry logs (CPU/GPU utilization, memory, temperature, power, latency) recorded at 0.5 Hz.

## Scoring recipe

```python
def compute_metrics(logs, num_images=5000):
    total_time = logs['inference_time']
    fps = num_images / total_time
    inference_time = total_time / num_images
    memory_usage = max(logs['memory'])
    w_total = max(logs['power'])
    w_per_fps = w_total / fps
    return {'FPS': fps, 'inference_time': inference_time, 'memory_usage': memory_usage, 'Wtotal': w_total, 'W/fps': w_per_fps}
```

## Common pitfalls

- Comparing pre/post-processing times across different CPU architectures (ARM vs x86) without accounting for architectural differences.
- Using the full 41k COCO test set instead of the specified 5k subset, which alters throughput and memory profiling results.
- Measuring training metrics instead of inference-only metrics, as the benchmark explicitly focuses on deployment and real-time inference.

## Evidence (verbatim from paper)

> The most essential part for mobile applications is efficiency and the amount of throughput per second of the hardware. Therefore, the metrics frames-per-second (FPS), inference time, memory usage and energy efficiency (W) (Wtotal and W/fps) were used for the benchmark. For the benchmark evaluation, we use the specifications given in Table 1 as the maximum GPU power consumption.

## Citation

```bibtex
@misc{mantowsky2021proai,
  title={ProAI: An Efficient Embedded AI Hardware for Automotive Applications -- a Benchmark Study},
  author={Mantowsky et al. (2021)},
  year={2021},
  note={arXiv:2108.05170}
}
```

- arXiv: 2108.05170