Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Performance landscape of CNN acceleration tools and resource constrained platforms

Miliadis Panagiotis

Full record

---

Summary

Over the last years, a rapid growth in the development of applications that are based on Convolutional Neural Networks is observed.Despite of the large advances in processor units, the use of computervision tasks is still challenging in resource constrained platforms. Thisthesis will present four toolkits, that accelerate the performance ofinference applications by targeting the processor units from the tophardware vendors; Intel, Nvidia, Arm and Xilinx. In order to achieveoptimal execution, the toolkits exploit the hardware acceleration thatprocessors provide, as well as special processor units and platforms,which are specially developed for deep learning inference tasks. Themost well-known models for each task are described, alongside withthe frameworks that the toolkits support and are used for model representation. Last but not least, real-world performance results arecollected for different batches of images, in order to achieve a performance landscape of the existing tools.

Available Files

Services