SaaS Implementation of Method A vs Method B24 Feb 2024 21:47
In response to a suggestion elsewhere regarding operation on GPUs and translation to CUDA...
It may not work like that in the case of ASICS. A SHA256 ASIC is likely to be a, few, thousand instances of logic blocks specifically designed to perform SHA256. As such all the front end will do is throw a thousand nonces at them and check the outputs for a, correct, solution before throwing another thousand at them until a solution is found
The problem for the front end, within a rig and accross multiple rigs, will be to ensure that no specific SHA256 block is asked to waste its time solving a duplicate nonce. The problem therefore is not converting to parallelism, CPU(Assembler) -> GPU(CUDA) it is getting the right nonces to the right blocks without duplication.
QBT's problem, for Method B which apparently selects nonce ranges, will be to use their SaaS to dole out those nonce ranges across rigs, and farms of rigs, in a way that avoids that duplication. On top of that, if you assume multiple unnassociated customers or clients, they have to do that in a fair and transparent way.
QBT's, similar, problem for Method A will be precalculating interstitial values, possibly for each nonce, and then serving them up to different rigs whilst avoiding similar duplication issues.
On the basis of the above the implementation for Method A, under SaaS conditions, is likely to be orders of magnitude harder than it is for Method B so this is possibly why they are not talking so much about it.