More is Different: Prototyping and Analyzing a New Form of Edge Server with...
summary
A commercial SoC Cluster with 60 Snapdragon 865 SoCs in a 2U rack and deployed for mobile cloud gaming is measured against servers using a benchmark suite that exercises video transcoding and deep-learning inference with FFmpeg and Android media codec, TVM and TensorFlow/ONNX stacks. Results show mobile SoCs deliver up to 3.2× CPU energy efficiency and 4.5× vs A14 GPU for streaming, hardware codecs give ~2.5–4.7× throughput gains, but video PSNR can drop up to 15% and large-model inference latency (ResNet152) remains hundreds of milliseconds per SoC.
AI-written summary · about
chapters (12)
- 0:00Energy-efficient edge servers
- 1:03Hardware selection: cloud vs edge
- 2:51Smartphones as edge hardware
- 4:55Smartphone farms vs SOC clusters
- 6:01SOC cluster architecture
- 7:45CPU performance benchmarking
- 9:12Video transcoding energy results
- 11:43Hardware codec performance
- 13:54DL inference latency challenges
- 15:20Scaling inference with SOCs
- 16:24Mobile processor performance study
- 18:08Conclusion and future directions