I decided I would graph the probability of
NOT FINDING a hash after searching the NONCE space using three different amounts of memory... 100% 50% and 25%
Both use cases search the same number of nonces and thus do the same number of SHA512... but you take a performance hit if you reduce your memory...
Obviously, slight reductions in memory don't hurt too much, but if you attempted to reduce your memory from 768 MB down to 1 MB so that you could operate 1000 runs in parallel then your probability of NOT finding a match would be:
0.368196 WITH MEMORY,
0.997402 WITHOUT MEMORY
So your probability of finding a match is 243x smaller for the same number of computations.
So I guess this means that you can trade 765 MB of RAM for 243 parallel processes each with 1 MB of RAM and have the SAME hash rate. Now you only actually decreased your RAM requirements by 33%... while increasing the number of computations required by 243x...