Cost-effective Hardware Accelerator Recommendation for Edge Computing
summary
A recommender prototype collects per-device execution traces (time and power) for ras950 and tiny uno yellow models, fits random-variable latency and power profiles, then estimates system cost (hardware + electricity) over a 24-month deployment across a three-level cloud-fog-edge topology with three groups of four cameras (need ≥2/4 per group). Results show general CPUs are cheapest-performing, FPGAs and embedded GPUs give best long-term cost-power tradeoffs, NCS ease-of-use trades off with low compute, and FPGA usability overhead is high.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 0:27Outline
- 0:52What are hardware accelerators
- 1:47Why do we need hardware accelerators
- 2:29Current challenges
- 3:10Metrics
- 3:31Hardware acceleration strategy
- 6:40Per device validation
- 7:15Design application topology
- 8:45Design latency
- 9:30Conclusions
related talks
28mML on the Edge for Industry 4.0 with Arm Ethos N-78 Neural Processing Unit
43mSoftware Development Trends for FPGAs
22mPreview: Hardware and Offloads
01mAAEON's new ARM based solution.
50mAI acceleration at the edge
22mExtending European Strength with Embedded AI
06mEmbedded Executive LIVE
14mEmbedded Executive LIVE