If I understand you correctly, you have a list of 500 data sets, where each set comprises 1 to 7 matrices of 1K x 1K values (bytes? ints?); your test run of 13 "observation numbers" yielded 51 query executions, so maybe there was an average of about 3 matrices per observation? (13 queries to get the lists of detectors per observation, plus 13 * 3 queries to get the matrices for three detectors, would come to 52 queries -- am I on the right track here?)

You might want to consider separating the data fetching from the actual computation, especially if you plan to be chewing over this particular set of 500 observations for a while (e.g. trying different statistical summaries, grouping or sorting things in different ways, etc).

Local disk i/o on whatever machine is running the perl script will be a lot faster, and a lot less wear-and-tear on the DB server, than (re)fetching ... what is it? 500 * (3 or 4) * 1K * 1K * (whatever byte count per CCD pixel), so somewhere between 1.5 GB and 4 GB, or something like that?

Try just storing all those matrices as local data files, and just do that once. Then write a perl script to do stuff with the local data files. You'll have a lot more flexibility in terms of playing with different strategies for computation and organization that way, and completion time for any given approach will be faster.