I need some help with a little experiment.
Source:
https://bitbucket.org/hudson/magic-lantern/commits/3046238719b25D3 113 binary:
autoexec.binRun it from Debug menu, "Card buffer benchmark (infinite)". It will write a 256MB file with random buffer sizes, which should help us find the optimal ones. The results will be saved in BENCH.LOG on your card - send me these files and I'll post some nice graphs.

You can run the test for as long as you can (1 minute, 1 hour, until the battery runs flat, or leave it during the night if you have an AC adapter).
Be careful:
- BENCH.LOG is overwritten every time you run the test.
- Don't forget about Canon power saving.
- LiveView will shutdown after 30 minutes no matter what, so it's probably best to run it in photo mode.
I know the file size is a bit small, but I'm trying to find the repeatability of this process. Also I want to see if the variations are white noise or there's a clear pattern (e.g. if there are any preferred buffer sizes). Current hypothesis: best buffer sizes are from 16 to 30 MB, with a small speed drop at 32MB.
If you are worried about card wear, don't run the test

You can also post results from other cameras.
Here's an Octave script that you can use to parse BENCH.LOG. It might work in Matlab too.
d = dlmread('BENCH.LOG', ' ', 4, 0);
bufsize = d(:,1);
speed = d(:,2) / 10;
plot(bufsize, speed, '*')