Series forces 100% of the flow through each card, and amount of restriction is combined.
Parallel splits the flow between cards, and the amount of restriction to the loop is also split.
Which is better is a 'tastes great' vs 'less filling' argument. Your results will vary from system to system depending on what pump(s) you have set to what speeds and amount of restriction you have elsewhere, etc.
FWIW, Swiftech did a study and found the differences between running GPUs in series vs parallel is "nonimal". In their case they did manage to get a fraction of a degree benefit to running GPUs in parallel vs in series.
A few more thoughts about this regarding the two most popular types of pumps ..
(aka MCP35x, PMP400) pump has higher head pressure and runs better in loops with a least a few waterblocks of restriction - actually runs hotter and are known to be more likely to overheat if there isn't enough restriction - so a DDC pump would probably tilt the scales a little towards running GPUs in serial.D5
s (aka MCP-655, PMP450, etc) have higher flow and less head pressure, more suitable for loops with less restriction, so would probably lean a little more towards running in parallel.
Do note that I have no empirical evidence that those last two observations are certainties. They are assumptions I have made regarding what are well-known characteristics of the two pumps and how I assume that would extrapolate with regards to running GPUs in series/parallel. If anyone has any evidence that either backs them up or shows to the contrary I would be very interested to see it. Edited by Unicr0nhunter - 11/29/13 at 10:01pm