NullHop: A Flexible Convolutional Neural Network Accelerator Based on Sparse Representations of Feature Maps.

Area of Science:

• Computer Engineering
• Artificial Intelligence
• Hardware Acceleration

Background:

• Convolutional Neural Networks (CNNs) are state-of-the-art for visual processing but are power-intensive.
• Existing graphical processing units (GPUs) offer limited power efficiency for CNN inference ( < 10 GOp/s/W).
• Low-power, low-latency solutions are crucial for edge computing and real-time visual tasks.

Purpose of the Study:

• To propose NullHop, a flexible and efficient CNN accelerator architecture.
• To leverage neuron activation sparsity for accelerated computation and reduced memory footprint.
• To enable high utilization of computing resources for diverse CNN kernel sizes and feature maps.

Main Methods:

• Designed a flexible CNN accelerator architecture, NullHop.
• Implemented NullHop on a Xilinx Zynq Field-Programmable Gate Array (FPGA).
• Evaluated performance on various CNNs, including VGG16 and VGG19, using postsynthesis simulations.

Main Results:

• Achieved over 450 GOp/s for VGG19 at 500 MHz in a 28-nm process.
• Demonstrated over 368% efficiency by exploiting sparsity.
• Attained a power efficiency exceeding 3 TOp/s/W within a 6.3 mm² core area.
• Maintained over 98% utilization of multiply-accumulate units.

Conclusions:

• NullHop provides a highly efficient and flexible solution for CNN acceleration.
• The architecture significantly reduces power consumption and latency for visual processing tasks.
• Successful FPGA implementation and integration with neuromorphic cameras demonstrate practical usability for real-time interactive systems.