Soc Cluster Transcoding Eval

---
name: soc-cluster-transcoding-eval
description: Evaluates the energy efficiency, throughput, and output quality of a custom edge server built from 60 mobile SoCs against traditional CPU and GPU servers for video transcoding workloads. Use when the user wants to benchmark on vbench, or asks about evaluating this task. Reports streams/W.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2212.12842
  bibtex_key: zhang2022more
  confidence: high
---

# soc-cluster-transcoding-eval

> More is Different: Prototyping and Analyzing a New Form of Edge Server with Massive Mobile SoCs — Zhang et al. (2022) (arXiv:2212.12842, 2022)

## What this evaluates

Evaluates the energy efficiency, throughput, and output quality of a custom edge server built from 60 mobile SoCs against traditional CPU and GPU servers for video transcoding workloads.

## Datasets

- **vbench** — total 6; splits: test (6)

## Metrics

- `streams/W` **(primary)** — range: other
  - Number of video streams supported per watt of power consumed. Calculated as total throughput (streams) divided by average power draw (W) during the workload.
- `PSNR` — range: other
  - Peak Signal-to-Noise Ratio measured in dB between the original and transcoded video frames. Higher values indicate better perceptual quality.

## Input / output format

**Input**: Video clips with specified resolution, FPS, source bitrate, and target bitrate, processed through software (libx264) or hardware (MediaCodec, NVENC) encoders on different hardware platforms.

**Output**: Transcoded video streams/files, power consumption readings (Watts), and network traffic metrics (Mbps).

## Scoring recipe

```python
def calc_streams_per_watt(streams, power_w):
    return streams / power_w

def calc_psnr(original, transcoded):
    mse = mean_squared_error(original, transcoded)
    if mse == 0: return float('inf')
    return 10 * log10((255**2) / mse)
```

## Common pitfalls

- Hardware codecs may fail to meet low target bitrate constraints, producing higher bitrates than the source video.
- PSNR differences stem from encoder implementation quality rather than hardware capability alone.
- Network bottleneck analysis assumes ideal software delegation, but daemon processes consume additional CPU resources in practice.

## Evidence (verbatim from paper)

> For live streaming transcoding, energy efficiency is measured as the number of video streams supported per watt (streams/W). For archive video transcoding, energy efficiency is measured as how many frames can be processed per Joule (frames/J). We unify the above metrics as throughput per energy unit (TpE).

## Citation

```bibtex
@misc{zhang2022more,
  title={More is Different: Prototyping and Analyzing a New Form of Edge Server with Massive Mobile SoCs},
  author={Zhang et al. (2022)},
  year={2022},
  note={arXiv:2212.12842}
}
```

- arXiv: 2212.12842