|
FPGA and GPU be designed and implemented?Bog'liq Alireza FasihFPGA and GPU be designed and implemented?
CNN is a complex design in terms of implementation and performance on
traditional architectures of the von Neumann type (such as CPU and sequential
processors). Therefore, we did a survey about CNN implementations on hardware
and later on we have proposed our implementation on FPGA. Due to the limitation
of resources in FPGA, we have implemented a fix-point DT-CNN with high accuracy
in results. Due to the offered flexibility of design by GPUs, we have also
implemented a CNN universal machine on GPU image processing. We have used
OpenCL which is a very strong framework for developing parallel algorithms on
GPUs. The results obtained have shown to be at least 100 times faster than on
normal computer/CPU for processing images.
Research question 7: How far can CNN be used/involved in an evolutionary
computing/control context example (for illustration)?
118
CNN has a great potential for signal processing tasks and it can generate very
complex nonlinear waves and oscillation patterns at the output of CNN cells.
Controlling both the kinematic and the inverse-kinematic of complex robots with
high degree of freedom (DOF) is a very complex scenario whereby classical
solutions fail to solve it easily. In the frame of this research question we have
integrated a CNN processors system to the leg-robot for high level inverse
kinematic controlling. For calculating templates we used GA and a very simple
objective function without considering any inverse or direct kinematic. Different
types of robots are able to move in an optimum way between two points or any
kind of other scenarios and trajectories.
As outlook and future work we do see the integration of all functionalities developed in
this thesis in a working prototype and improve it progressively according to test results in
the real environment.
|
| |