Why is the machine vision industry not exploiting the potential of GPUs?
Many third party libraries in the machine vision sector are taking advantage of the additional computing power offered by GPUs to speed up the execution of some image processing algorithms. The main issue is to guarantee the lowest possible latency from the moment when the image is acquired to the availability of the image for processing in the GPU memory. This requires the image to be copied from the host PC into the GPU memory, which uses a small amount of CPU time. With AMD's DirectGMA, the Coaxlink cards from Euresys allow graphic memory to be exposed to any device.
The machine vision market is moving towards a completely new approach. The availability of the image directly in the GPU memory in combination with very high-resolution, high-speed cameras will open up new frontiers for high-end applications.
The Euresys CoaXPress frame grabbers include the latest PCIe technology. They are based on the PCIe 3.0 (Gen 3) x4 bus that can achieve a sustained bus bandwidth of 3300 MB/s. These frame grabbers provide the highest data acquisition rate in the industry, which allows for bandwidth of 25 Gbit/s (2500 MB/s) from the camera to the host PC memory and, thanks to AMD's DirectGMA technology, directly to the GPU memory. The accessibility of the GPU memory eliminates unnecessary system memory copies, dramatically lowers CPU overheads and reduces latency, resulting in significant performance improvements in data transfer times for applications that use the AMD FirePro W5x00 (and above) and for all AMD FirePro S series products.
Market trends indicate that customers are only interested in higher resolutions if they are supported by higher speeds. Only a few years ago, 5Mpx cameras were referred to as high resolution, but nowadays many applications require 12, 16 or even 25Mpx and have acquisition frame rates that have risen from a few frames per second to 80 fps with full resolution and a CoaXPress interface. This means that for many applications which require heavy image processing and analysis, the bottleneck has now moved to the host CPU. A single PC, even a high-end or multi-CPU machine, may not be powerful enough to keep pace with these data rates.
The aim of the AMD direct graphic memory access technology is to make image data available for most of the image processing that can be performed in the CPU instead of the graphics card. This is the first step towards the development of a full machine vision system where all the image processing takes place inside the GPU. The result will be a significant improvement in the elaboration time and an increase in the amount of data that can be processed, which will allow tasks to be performed in real time.