Re: lolMiner 0.5 Equihash Miner, Update: Sep 25th: More Speed & AION Mining
Ok then,
first of all thanks to all testing this new stuff so early after release and commenting here.
I was thinking a lot about consequences from what I read and I think this will be the way it goes:
1) I will stop advertising this miner to be usable for NVidias (at least for the moment)
It seems it is ok for AMD. Some times more then 4 or 5 GPUs in one instance make it non-stable but thats an issue one can deal with.
On Nvidia it is completely different somehow. For MNX and AION it is an easy pick: use ewbf, its just much quicker. For 144/5 and 192/7 the picture is very mixed. Performance for 1070 is ok, for 1080 is good, but below it does not scale. And this in a way I can not explain because well... Ideally the 1060 has 60% of bandwidth of a 1080 and exactly 50% of the shader clusters, so one expect the performance to be between 50 and 60% ... but here its at only 40% which just does not make sense. I expect that this is due to some strange things the OpenCL compiler does on this cards...
2) The current code / situation seems to be a trap.
The host side mining code I have is still similar to the one I used for MNX back when I did 0.21. Not bad, but I am hitting the limitations of it now. Still supporting Nvidia ends up that my scheduling for AMD is non optimal (I could use way less CPU load and also get rid of problems with more then 4 GPUs when I would not have to support NV). Also supporting Nvidias causes that I can not use more modern OpenCL 2.0 or 2.1 language features. Other way around sticking to OpenCL makes the results on Nvidia unpredictable.
So I decided that I will change some things for future versions.
First I will design a more flexible mining back end that schedules work better then now. This means that 0.51 will likely be AMD only. Then I will see if I finally learn some CUDA to add support for Nvidia back in, algorithm by algorithm. Seems like a lot of work? It is! But I got the feeling that the algorithms are not bad in itself. In fact on AMD they work pretty well on on Nvidia you can see here and there its potential. So I think it is worth doing the effort.
Stay tuned.
Ps: will later reply to the concrete support requests here. Right now busy with typing a new front page entry. Keep mining
first of all thanks to all testing this new stuff so early after release and commenting here.
I was thinking a lot about consequences from what I read and I think this will be the way it goes:
1) I will stop advertising this miner to be usable for NVidias (at least for the moment)
It seems it is ok for AMD. Some times more then 4 or 5 GPUs in one instance make it non-stable but thats an issue one can deal with.
On Nvidia it is completely different somehow. For MNX and AION it is an easy pick: use ewbf, its just much quicker. For 144/5 and 192/7 the picture is very mixed. Performance for 1070 is ok, for 1080 is good, but below it does not scale. And this in a way I can not explain because well... Ideally the 1060 has 60% of bandwidth of a 1080 and exactly 50% of the shader clusters, so one expect the performance to be between 50 and 60% ... but here its at only 40% which just does not make sense. I expect that this is due to some strange things the OpenCL compiler does on this cards...
2) The current code / situation seems to be a trap.
The host side mining code I have is still similar to the one I used for MNX back when I did 0.21. Not bad, but I am hitting the limitations of it now. Still supporting Nvidia ends up that my scheduling for AMD is non optimal (I could use way less CPU load and also get rid of problems with more then 4 GPUs when I would not have to support NV). Also supporting Nvidias causes that I can not use more modern OpenCL 2.0 or 2.1 language features. Other way around sticking to OpenCL makes the results on Nvidia unpredictable.
So I decided that I will change some things for future versions.
First I will design a more flexible mining back end that schedules work better then now. This means that 0.51 will likely be AMD only. Then I will see if I finally learn some CUDA to add support for Nvidia back in, algorithm by algorithm. Seems like a lot of work? It is! But I got the feeling that the algorithms are not bad in itself. In fact on AMD they work pretty well on on Nvidia you can see here and there its potential. So I think it is worth doing the effort.
Stay tuned.
Ps: will later reply to the concrete support requests here. Right now busy with typing a new front page entry. Keep mining