This thesis investigates the possibility of porting a neural network model trained and modeled in TensorFlow to a low-power AI inference accelerator for IoT edge computing. A slightly modified LeNet-5 neural network model is presented and implemented such that an input frequency of 10 frames per second is possible while consuming 4mW of power. The system is simulated in software and synthesized using the FreePDK45 technology library. The simulation result shows no loss of accuracy, but the synthesis results do not show the same positive results for the area and power. The default version of the accelerator uses single-precision floating-point format, float32, while a modified accelerator using the bfloat16 number representation shows significant improvements in area and power with almost no additional loss of accuracy.
| Identifer | oai:union.ndltd.org:UPSALLA1/oai:DiVA.org:liu-181711 |
| Date | January 2021 |
| Creators | Hansson, Olle |
| Publisher | Linköpings universitet, Datorteknik |
| Source Sets | DiVA Archive at Upsalla University |
| Language | English |
| Detected Language | English |
| Type | Student thesis, info:eu-repo/semantics/bachelorThesis, text |
| Format | application/pdf |
| Rights | info:eu-repo/semantics/openAccess |
Page generated in 0.0024 seconds