Print Email Facebook Twitter SASCNN: A Systolic Array Simulator for CNN Title SASCNN: A Systolic Array Simulator for CNN Author Marigi Rajanarayana, Shashanka (TU Delft Electrical Engineering, Mathematics and Computer Science) Contributor van Leuken, Rene (mentor) Degree granting institution Delft University of Technology Programme Electrical Engineering | Circuits and Systems Project PRYSTINE Date 2019-11-05 Abstract Convolution Neural Networks (CNN) are used in many applications ranging from real-time object detection to robot-motion planning. CNNs are implemented on high-performance systems like multi-core CPU and GPU, these are of high power in nature and thus cannot be deployed in edge devices due to their limited battery power. The edge device has to provide real-time performance along with being low power, this prompts for an exploration of novel architectures catered towards the processing of CNNs. The recent works towards this goal have been the development of CNN accelerators using systolic array spatial architectures. The row-column stationary data-flow approach maximizes the reuse of weights, input feature maps and output feature maps across the array. Different applications require different performance, area and energy needs, and this makes it imperative to quickly prototype the architectural ideas and perform design space exploration. The challenging part is the non-trivial interactions between different architectural design parameters, as they play an important part in the complex design decisions. Hence, a hardware simulator to accelerate CNN is designed in this work. It is based on systolic array and uses row-column stationary data-flow with a near memory computing approach. The simulator supports different numerical precision such as 16-bit and 8-bit floating-point along with numerous design parameters such as the size of the systolic array, latency of MAC operation, PE local memory size, PE local memory latency and external memory latency. The functionality of the proposed design is verified on AlexNet. The Destiny memory modelling tool, along with energy and area estimation model, is used to perform a system study to investigate the trade-offs between different architectural design parameters. Subject Deep neural networks, CNN, Accelerator, Simulator, systolic array, Near-memory computing. To reference this document use: http://resolver.tudelft.nl/uuid:5266a567-9864-4ffd-8e25-0d4d0e5f322a Part of collection Student theses Document type master thesis Rights © 2019 Shashanka Marigi Rajanarayana Files PDF SASCNN_A_Systolic_Array_S ... 45698_.pdf 4.13 MB Close viewer /islandora/object/uuid:5266a567-9864-4ffd-8e25-0d4d0e5f322a/datastream/OBJ/view