Thanks, wqweto and Olaf. Nice tests.

The RC5-SB (using Mid$) being about factor 4 faster on my machine (for such larger outputs).
It's because of too frequent reallocations. For such large data, initial Chunk increasing size in VolteFace's SB should be adjusted to ~ like 10 MB (Const CHUNK_SIZE = 10485760).
In such case above test returns on my machine in compiled mode:
SB1 Dragokas 568,67msec 49845014
SB2 RichClient 610,79msec 49845014
SB3 Collection 407,87msec 49645012