Hi guys, I have a question. The background is rather theoretical, but maybe
you Burst-miners can help.
I have a dataset of 32 byte words and want to apply Shabal to each word.
To be clear, there are no "Shabal-rounds" anywhere, I want to apply Shabal
to independent pieces of data:
(x_1, x_2, ....) |-> (Shabal(x_1), Shabal(x_2), ....)
(Each x_i is a 32 byte word).
If I have a GPU at hand, I guess one could parallelize this: Divide the dataset
into N words, load them into GPU memory and let some opencl implementation
of Shabal do the work.
My question:
1. What is the optimal N? (I guess, depends on GPU.. But just some number of
some reasonable GPU would be great) Can such an optimal N be easily calculated
from some GPU data (i.e. N = # of shaders / memory, or another formula, I'm
clueless sorry
)
2. What speed can I achieve? I.e., how many segments (I call such a cluster
consisting of N 32-byte-words "a segment", hope that's not conflicting notation).
Or, put differently, if I optimize my parallelization, and given that my dataset is
quickly available, how many Shabal-computations can I expect.
I turn to this thread because this is, if I'm not wrong, basically the task
a GPU Burst miner performs. So, put again differently, my question is:
Assume your plots lie in main memory or SSD (I just want to take HDD
speed out of the equation), how many nonces can a GPU process per
second?
Thanks!