Re:Sorting with GPUs: A Survey
by dragontamer on Monday September 11, 2017 @11:45AM
(#15219367)
> I would think the breakeven point for using the GPU (assuming inputs and results are on the CPU) is several megabytes of millions or elements at least.
Well, the idea of GPU computing is to keep the data on the GPU as much as possible.
The CPU <---> GPU link is at best, 16x PCIe. Which is fast, but not anywhere close to the speed of HBM, GDDR5 or DDR4 RAM. (Exception: AMD's A10 operators, which have a GPU / CPU which shares on-chip cache memory. As well as the "Crystalwell" Intel chips IIRC. But these are rare chips that I doubt most people use)
So if you happen to be running a major algorithm on the GPU, you'll probably want to sort it as well, before bringing it back to the CPU. Under no circumstances should you be introducing a PCIe delay for a simple operation like sorting.
Well, the idea of GPU computing is to keep the data on the GPU as much as possible.
The CPU <---> GPU link is at best, 16x PCIe. Which is fast, but not anywhere close to the speed of HBM, GDDR5 or DDR4 RAM. (Exception: AMD's A10 operators, which have a GPU / CPU which shares on-chip cache memory. As well as the "Crystalwell" Intel chips IIRC. But these are rare chips that I doubt most people use)
So if you happen to be running a major algorithm on the GPU, you'll probably want to sort it as well, before bringing it back to the CPU. Under no circumstances should you be introducing a PCIe delay for a simple operation like sorting.