Open Convolver Architecture
For Deep Convolutional Neural Networks
While this technology growth had permitted to DCNNs to evolve rapidly,
the problem continued to be present in embedded systems and non general
purpose computers that, due to the lack of graphics cards and/or other
powerful parallel subsystems, kept relaying into a client-server model
to incorporate this type of technology. While this model could work to
some application, the presence of a good connection to a server is not
always guaranteed and even if it is present the bandwidth needed by the
server for such data transfer can easily grow up.
In this work i tried to implement an open convolver architecture. It can
easily be embedded in existing systems and/or operate as standalone
processing unit. The architecture was tested using the Nexys 4 DDR
development board based on Xilinx Artix-7 FPGA chip. The code is written
in VHDL on Vivado IDE which is the official tool released by Xilinx for
the FPGA Development.
Features:
- Adaptive and reconfigurable for variuos kernel/image-slices
- Have own custom package for code readability
- Highly parallel architeture for each individual sub-system
- Resource selection feature to choose between cost and preformance
Future Development
- Run-time configurability
- Self-improve using a dedicate DCNN in the CHIP COOL !
- You suggest to me
License
Free Software, Hell Yeah!