Let's look deeper into how the GPU accelerates certain processing operations. As we know, CPUs are designed for the sequential execution of data that results in significant running time for certain classes of applications. Let's look into the example of processing an image of a size of 1,920 x 1,200. It can be calculated that there are 2,204,000 pixels to process. Sequential processing means that it will take a long time to process them on a traditional CPU. Modern GPUs such as Nvidia's Tesla are capable of spawning this unbelievable amount of 2,204,000 parallel threads to process the pixels. For most multimedia applications, the pixels can be processed independently of each other and will achieve a significant speedup. If we map each pixel with a thread, they can all be processed in O(1) constant time.

But image processing is not the only application where we can use data parallelism to speed up the process. Data parallelism can be used in preparing data for machine learning libraries. In fact, the GPU can massively reduce the execution time of parallelizable algorithms, which include the following:

• Mining money for bitcoins
• Large-scale simulations
• DNA analysis
• Video and photos analysis

GPUs are not designed for Single Program, Multiple Data (SPMD). For example, if we want to calculate the hash for a block of data, it is a single program that cannot run in parallel. GPUs will perform slower in such scenarios.