The key point is average fps, let's say you are playing 30fps video and O3DS decodedWell, the fake model option is definitely working, because performance is terrible in O3DS mode.
I am a little confused though... I tried XviD 240x144p @ 24 fps, and the average fps is 51, and the recent average is always above 40, but the video is still playing in slow motion with skipping audio. If the decoding speed is > 40 fps on a 24 fps clip, shouldn't it play at full speed? Based on your benchmarks, H.261 is only ~6 fps faster than H.263+ (not quite the same as XviD, but close enough), so it seems like you'd need to use 160x96 or even 80x48 to get full speed on O3DS...
On the N3DS front, the larger buffer for MVD_Serivces is working. Clips with 5-9 reference frames no longer crash the system.
frame at 20ms
frame at 25ms
frame at 16ms
frame at 38ms
frame at 22ms
The average decoding time is 24.2ms and fps is 41.3, but you'll likely to hear sound stutter at frame because it over 33.3ms(1000ms / 30fps = 33.3ms per frame)
Although average fps is 53, You'll hear sound stutter at black arrow
OLD3DS uses same core for decoding and drawing so using full screen mode may increase decoding speed a little(because it won't draw bottom screen during full screen mode).