Ultra fast cnn based Hardware Computing Platform Concepts for adas visual Sensors and Evolutionary Mobile Robots

Download 3,22 Mb. Pdf ko'rish
bet	44/81
Sana	16.05.2024
Hajmi	3,22 Mb.
	#238917

1 ... 40 41 42 43 44 45 46 47 ... 81

Bog'liq
Alireza Fasih

7.

Implementation of CNN on FPGA

In this chapter the focus lies on the following research question:

“How far can an
efficient implementation of CNN on FPGA and on GPU be designed and realized? “

7.1

Introduction
The implementation of CNN on hardware is very similar to the emulation of analog
computing on FPGA. There are some bottlenecks for accessing the memory and updating
state variables for each cell. In the emulation of analog computing in FPGA, we did model
basic elements of system which are necessary for modeling the behavior of the system. The
same method can lead to implementing CNN on FPGA. In contrast to ANN and because of
the local connectivity of CNN cells, one does have the possibility to implement this
architecture on FPGA and also on GPU [50]. FPGA macro cells and logical components can
work in a highly parallel manner. And because of a very flexible routing between them, we
can implement any complex digital circuit or model. To get more advantages from using
FPGA we can use high level behavioral modeling languages such as VHDL, Verilog or
SystemC. FPGA has local and embedded memory, which is very important for storing the
CNN states, otherwise a transceiver of memory between FPGA and an external memory
could be very time consuming and constitute a significant bottleneck. Today, most FPGAs
have an internal dedicated standard CPU that has access to the hardware and logical field
of FPGA through the standard bus controller [117, 118]. There are many high level
compilers based on ANSI-C standard for coding and debugging. This technology increases
the system performance by integrating of hardware and software. Therefore, loading CNN
initial states, templates, and setting time scales and other parameters can be done easily by
CPU. The resources of an FPGA are not endless; thus we have to consider this issue while
designing the CNN architecture.

70
Concerning GPU, using GPU is getting more popular every day. The highly parallel
structure of GPU makes it more efficient for image processing and for processing large
blocks of data. The high memory bandwidth between CPU and GPU, the integration of GPU
and CPU through the standard protocols and the running multi-kernels scripts on GPU
make it a very efficient technology for the implementation of CNN [119, 120]. Since 2003
GPU technology is growing up dramatically and we can implement very complex models
and systems by using flexible and robust tools and high-level software development
instruments/tools.

Download 3,22 Mb.

1 ... 40 41 42 43 44 45 46 47 ... 81

Download 3,22 Mb.

Pdf ko'rish