FPGA Routes

Posters :

ENPOWER poster for ZYNQ devices (PDF, 552kB)

Video 1 :

Follow this link to see Adaptive Power adaptation in the ZYNQ board for a video coding application:

The figure below shows preliminary results of how ENPOWER can scale energy and performance in a hybrid FPGA device. The figure shows the energy used by a motion estimation accelerator mapped to the FPGA fabric as the user design. This motion estimation engine is part of a video coding system built around the H.264 video codec standard. In this standard motion estimation can represent 60 percent of total computation so it is good candidate for FPGA acceleration. The red line shows the energy used by the core at nominal voltage which, as expected, is independent of the frequency of operation. ENPOWER enables elastic performance and energy with the measured results shown in blue.

Preliminary performance and energy scaling with ENPOWER in a ZYNQ device

Preliminary performance and energy scaling with ENPOWER in a ZYNQ device

Video 2 :

Follow this link to see a video fusion application in ZYNQ.

Video fusion merges frames from infrared and visible into a single frame where more information is available. In this demonstrator the hardware  engine in the FPGA performs the dual tree complex wavelet transforms both forward and backwards.  The ARM processor and the FPGA work together  performing fusion and sharing data using the coherence ports with a Linux kernel driver used to access reserved memory required for the data sharing. Initial results show that the FPGA+CPU uses 4 times less energy than the CPU on its own. Video fusion also shows the interesting property that if the amount of available energy decreases below a threshold require for correctness degradation is graceful. This is shown in these frames that show that the application supports a certain error  rate in the FPGA before the quality of the output is affected significantly.  This enables over-scaling in the FPGA to save energy or increase performance.

  PSNR degradation in video fusion


Edit this page