Distributed Random Brute Force

I don't have the GPU power to make any progress with sequential brute force.
I also found, by experiment, that guessing random numbers can take much longer.

So, to maximize the fun, I am doing both. 
I distribute the scan over the 20-3F keyspace, pick 3 random bytes, and brute force the rest.

My ranges look like this: (where XXXXXX is 3 random bytes)

[20-3F][XXXXXX]00000000 - [20-3F][XXXXXX]FFFFFFFF

My old card can try fifteen 3-byte randoms per scan, every 13 hours, at 44Mkey/s.  Plus about 2 million really random randoms with the leftover starting points.

What does that get me?

15 random blocks of 4.3 billion keys in each of 32 sub-ranges [20-3F] per scan = 2 trillion.  4T/day.  Pffft.

so, every 26-hour day, scanning the following:
128 B keys in 20XXXXXX00000000-20XXXXXXFFFFFFFF
128 B keys in 21XXXXXX00000000-21XXXXXXFFFFFFFF
128 B keys in 22XXXXXX00000000-22XXXXXXFFFFFFFF
.
.
.
128 B keys in 3DXXXXXX00000000-3DXXXXXXFFFFFFFF
128 B keys in 3EXXXXXX00000000-3EXXXXXXFFFFFFFF
128 B keys in 3FXXXXXX00000000-3FXXXXXXFFFFFFFF


i just need to get lucky with 3 bytes.  how hard can that be?
or, get lucky with 2 bytes, but wait a week to find out.  that shouldn't take much more than 65535 weeks.