0100010 (noemail@noemail.noemail) on 4/18/12 wrote:
---------------------------
>Noone in the financial industry will ever use a GPU for algorithmic trading. Why?
>Latency on GPU compute sucks.

How about not speaking when you have no clue?

1) Latency requirements vary from miliseconds to overnight, if a high latency product can't be used for one specific thing it still may be used for something else.

2) If GPU happens to be faster than a CPU the latency for the job will be lower, hell, that's the definition of "faster", if there is a task there where GPUs can compute faster than CPUs they may be used to reduce latency.

3) Copying data from main memory to GPU memory, than doing something simple and copying back to main memory still orders of magnitude faster than the most strict latency requirements on the financial market.

4) The architecture of GPUs imply they would be used to analyze a big amount of data wich also imply it is a more elaborated algorithm than just checking the top of the book, for this algorithms trying to predict much ahead of time the latency requirement is just "faster than others", see #2.