Intermittent Computing
summary
Claims that Stateful deep neural networks embed inference-progress signals into output features so no separate progress writes are needed. Implemented on Texas Instruments microcontrollers, tested on three TinyML models, and compared to Hawaii and Japari. Results show progress-preservation overhead drops from 14–83% down to 1–21% of latency, overall inference speeds increase 1.3–5× (demo shows ~2× vs Japari) with negligible accuracy loss. Runtime middleware patches layers to split value ranges and assign embedded states to uniquely identify resumed jobs; prototype and middleware are released.
AI-written summary · about
chapters (14)
- 0:00Start
- 3:27Progress Indicators
- 5:17Stateful Deep Neural Networks
- 5:53Design Challenges
- 6:47State Information Embedding
- 9:46Demo
- 18:21Volatile Memory
- 20:14Address Distributed Concurrency Control in Intermittent Networks
- 34:22Shading Effect
- 37:48System Overview
- 40:45Power Imbalance Threshold
- 43:51Conclusion
- 45:41What's the Best Time and Frequency To Measure the Power and Energy for the Configuring of the Pv
- 55:47Latency Optimization
related talks
06mIntermittent Learning: On-Device Machine Learning on Intermittently Powered System
40mHands-on TinyML Deployment on RISC-V: Building an Optimized Edge AI Pipeline - Corrected Audio
01h 53mTinyML: Machine Learning’s Future is Tiny and Bright
40mHands-on TinyML Deployment on RISC-V: Building an Optimized Edge AI Pipeline
22mPreview: Hardware and Offloads
28mML on the Edge for Industry 4.0 with Arm Ethos N-78 Neural Processing Unit
27mWhat’s TinyML good for
16mLightning Talk: Efficient Inference at the Edge: Performance You Need at the Lowest...