after reading the miner source code, i think we dont need disk space at all
we can create opencl kernel of this code :
https://github.com/BurstProject/pocminer/blob/master/src/pocminer/util/MiningPlot.javanonce argument is any nonce u want, addr is ur account number
we can bruteforce any nonce within 240 seconds each block using GPU
and broadcast the best nonce that produce lowest deadline
assuming GPU can do shabal256 at 1 GH/s , that is equal to 240 Giga Nonce (240 secs blocktime) which is equal to 60 GB of disk space ( 1 nonce = 0.25MB )
and again, getting into 1 TB disk space performance, will require atleast 17 GPU (assuming 1 GH/s is accurate)
conclusion : if it is near 1 GH/s its cheaper to use harddrive than GPU
and loking at a glance of shabal256 it is really cheap in computation, i think we really need replace shabal with more memory intensive algo such as scrypt or something to make harddrive mining is far more efficient than GPU or future ASIC
shabal256 may be computationally cheap, but generating 1 plot does many thousands of rounds of it, instead of the couple rounds you see in most coins. I doubt a gpu could compute with hdd read speed considering how many rounds are run, although there's no current implementation to prove/disprove this. I would consider a gpu miner with bad efficiency a good thing however, since it would still help speed up plot generation for hdds by a significant margin. If I am proven wrong, we can always adjust the miningplot constants to increase the ram. Using a larger buffer, but decreasing the hash cap by the correct amounts could result in roughly the same computation complexity as exists now, but with a much higher ram requirement per thread. This would be negligible for pcs, but could cause gpus to not be able to use all their cores due to lack of ram.