Understanding How it all Works
Although many in the industry refer to this as GPU based deconvolution out of necessity, certain portions of the deconvolution process still occur on the CPU, notably the initial setup. Once this is complete, AutoQuant accesses the many GPU processor cores that carry out the actual iterative deconvolution. The final results are then transferred back to CPU and saved.
Graphics Processing Units (GPUs) have evolved from simple display adapters into computational powerhouses for applications like image processing. As rendering demands have increased, specialized processors on the GPU became more generalized, faster, and more numerous. These changes culminated in the rise of general-purpose computation being performed on devices that had nominally been designed to make special effects in games and movies look as realistic as possible.
The individual processor cores on a GPU are considerably slower than most CPU processor cores, but what they lack in power, they more than make up in numbers. The core count on a high-end workstation’s CPU(s) will be typically be around 32. In contrast, the core counts reported on even consumer-grade GPUs number into the thousands.
THOUSANDS OF CORES
For computationally-intensive applications like deconvolution, in which a series of mathematical calculations must be performed on each one of millions of individual voxels, being able to process thousands at once is preferable to processing a few at a time. In short, you have more bandwidth available when you utilize the GPU rather the CPU allowing for faster processing of the image.