The 9x8 looks close to optimal for this case. I've got only 1910 frames with the variable buffer method in simulation (compared to 1565 for 9x8MB).
A question for you: what write speed are you getting if you only write one frame at a time? (that is, split the buffers in very small chunks equal to frame_size - and as many as you can - should be 81 stars for the numbers above). This will be helpful in fine-tuning the speed curve.
Edit: current simulation results (1280x426, 23.976, speed tuned to 21.16MB/s to match 1565 frames):
8MB + 2x32MB - 317 frames:
9x8MB - 1566 frames:
Variable buffering, starting from 8MB + 2x32MB - 1910 frames:
X: time in seconds; Y: number of buffered frames.
All 3 cases with the same write speed and the same amount of total memory.
Warning: this is pure theory, I didn't try this on any camera. But you can see the mathematical background behind raw recording, study the algorithms and suggest improvements.
Note: if write speed wouldn't vary with buffer size, the optimal algorithm would be trivial: just write one frame at a time.
