Deep Learning Neural Network Acceleration at the Edge
summary
Survey of frameworks and accelerators with concrete pain points and a proposed fix: major OSS frameworks (TensorFlow ~32k commits, 1.6M LOC, ~456 developer‑years; MXNet, Caffe, PyTorch) use different model formats so converters are mandatory; edge silicon ranges from Google Edge TPU and Arm Project Trillium to Qualcomm NPE and HiSilicon NPUs, each with its own runtime and SDK so fragmentation multiplies maintenance cost. Linaro builds an open CI-backed runtime and plugin model, evaluating ONNX as common format and continuous accuracy tests to catch regressions.
AI-written summary · about
chapters (7)
- 0:00Deep learning at the edge
- 4:19Edge computing for neural networks
- 10:50ML frameworks overview
- 20:13Hardware acceleration
- 27:12Framework-accelerator challenges
- 29:03Open collaboration call
- 36:40Closing remarks & Q&A
related talks
30mScalable ML acceleration with ONNX Runtime
14mLightning Talk: Every Millisecond Counts: The Fine-tuning Journey of an Ultra-Eff...
36mApplied AI on the edge
16mLightning Talk: Efficient Inference at the Edge: Performance You Need at the Lowest...
42mAI and Machine Learning BoF
59mBuilding models for TPU devices
33mHigh Performance Deep Learning on Edge Devices With
34mFaster deep learning solutions...