Post
Topic
Board Mining (Altcoins)
Re: CCminer(SP-MOD) Modded NVIDIA Maxwell kernels.
by
djm34
on 18/09/2015, 17:58:28 UTC
Agree, Ethminer is still very lucrative although it looks like Nvidia hardware hit the memory controller limitation and can't go anywhere else, much like Cryptonote.
Look at lyra2v2. It's also a memory hard algorithm.
280x Sgminer opensource(200watt) 4MHASH
750ti Djm-34 (sp-mod) opensource (40watt) 5MHASH
It's like creating a etherum miner that does 35MHASH on the 750ti.
But the opensource is only doing 8MHASH...
(djm34 here...)
The main difference between the algo of ethereum and other mem hard algo, is that you can't rescale mem usage as it always requires the full dag file to run (ie 1.2Gb or so of vram with many random over the full dag file)
Meaning you can't really improve passed what has already been done... (yeah, I tried already  Grin ).

But you should be able to make the kernal run at copyspeed. (memory bandwidth limit) The gpu can do register operations while writing to memory. The keccak passes should be integrated and more than one hash per run. Then you will get keccak for free. (memory pipelining)


not sure to understand what you mean in the context, there are only 2 keccak rounds, 1 before and 1 after the sequence of random access to the dag file. So in my opinion it is already running (mostly) at mem access speed...
(it is easy to check, actually, just need to remove the 2 keccaks loops and see how much you would get).

actually it could be interesting to use the dag file only over a certain range of mem value, and for the remainder recalculate the missing dag item (but the calculation is rather heavy, so it isn't obvious it would work.)

regarding the advance of ccminer over sgminer this is probably only temporary, Wolf0 has a much faster code which is getting pretty much the same speed as my private kernel (so it can be achieved...)