« on: September 21, 2010, 06:17:36 AM »
Architecture and balance ratios are key concepts that are essential to understanding OpenCL device performance. In particular, the bandwidth of the link between the host and OpenCL device, along with the memory bandwidth of the device can be key performance metrics. In some cases, either of these hardware characteristics can make it too expensive to move a calculation from the host onto an OpenCL device. However, many problems require enough computation per data item transferred to greatly speed OpenCL applications.
Finally, this tutorial has provided a quick and basic introduction to OpenCL along with an example code that can run on both CPU and GPU device types. Try it and see how it works.
Full story here