Machine Learning (AI)

With the recent acknowledgements from Nature Magazine and the broader machine learning community about the effectiveness of the “Von Neumann” architecture for Machine Learning applications, this has opened the door to other potential solutions.

The limitations of the traditional "Von Neumann" CPU architechture has led to the exploration of different architecture, including GPUs dedicated ASICs and FPGAs. While GPUs deliver significant performance enhancements over CPUs, they are quite power hungry and not the optimum architectures for Machine Learning / AI architecture and algorithms. These can be implemented very efficiently on dedicated ASICs, however, it is generally not possible to maintain this efficiency over all classes of Machine Learning algorithms with a single architecture. Furthermore ASICs can not easily be adapted to take advantage of the latest innovation in neural network architectures.

 

FPGAs, however, offer the best of all worlds. It is possible to implement optimum architectures for a given Machine Learning algorithm. They can then be dynamically reconfigured for a different application. Omnitek have an extensive R&D program on the design of optimum FPGA machine learning architectures and will be announcing more details during 2018.

FPGA technology has become established as providing the most flexible architecture for machine learning. The most significant advantage of FPGAs over dedicated ASIC and ASSP devices is that they are potentially future proof in a rapidly evolving arena.

Image courtesy of YOLO V1 Paper



Omnitek have established themselves as a world leading development partner for high performance, responsive and reconfigurable video and vision guided intelligent systems. By leveraging the Xilinx Zynq All Programmable SoCs and MPSoCs with their unique IP, platforms and design services, they have assisted many of Xilinx’s leading customers to get to market faster with highly differentiated solutions.

Steve Glaser, Sr. Vice President, Corporate Strategy & Marketing at Xilinx

As with all reconfigurable technology, compared to ASIC and ASSP technologies, it is possible to continue to enhance designs, both dramatically and subtly as ideas change and new topologies evolve. This is a highly desirable attribute in an area of rapidly evolving algorithms.

As a Xilinx Premiere Partner, Omnitek are an established world leader in FPGA / MPSoC vision and video processing IP design and optimisation so it has become a natural progression for the business to apply its knowledge, skills and expertise finding a solution to the needs of Machine Learning. Omnitek is working with industry experts and academia to provide such a solution which will be announced later in the year.

If you are interested in Machine Learning, please contact us.