You know this is quite similar to what I experience on my home pc with Win 8.1 pro / GTX750Ti. it does have a bit of hashrate, but unhealthy low, including crashes and gpu mining errors. I haven't been able to find the cause for this, but perhaps I'm at least not alone anymore
You know this is quite similar to what I experience on my home pc with Win 8.1 pro / GTX750Ti. it does have a bit of hashrate, but unhealthy low, including crashes and gpu mining errors. I haven't been able to find the cause for this, but perhaps I'm at least not alone anymore
Its kinda strange , yea this rig got 6 750ti cards. I can switch @ linux if its Win 8.1 pro problem.
Edit: Tested 1 single GFX on win7 and it works fine.
Can someone provide binares for Winx64? I want to check my results against ordinary ethminer, but i'm not well-versed on compiling from source, I'm skipping that and downloading the automated builds. Thanks in advance!
I compiled actual github version on Ubuntu 14.04 x64 ( cuda-repo-ubuntu1404_7.0-28_amd64.deb )
when i use "cmake -DETHASHCL=1 -DGUI=0",build is successfull. but when i use the "-U" Switch, only CPU is used. so Cuda is possibly not included.
then i tried your hint "cmake -DBUNDLE=miner -DETHASHCU=1". but then it will not compile:
[ 87%] Building CXX object libethcore/CMakeFiles/ethcore.dir/Ethash.cpp.o In file included from /home/kabini/cpp1/cpp-ethereum/libethcore/EthashAux.h:27:0, from /home/kabini/cpp1/cpp-ethereum/libethcore/BlockInfo.cpp:28: /home/kabini/cpp1/cpp-ethereum/libethcore/Ethash.h: In static member function ‘static void dev::eth::Ethash::CUDAMiner::setDevices(unsigned int*, unsigned int)’: /home/kabini/cpp1/cpp-ethereum/libethcore/Ethash.h:177:24: error: comparison between signed and unsigned integer expressions [-Werror=sign-compare] for (int i = 0; i < gpuDeviceCount; i++) { ^ In file included from /usr/local/cuda/include/CL/opencl.h:44:0, from /home/kabini/cpp1/cpp-ethereum/libethcore/../libethash-cl/cl.hpp:163, from /home/kabini/cpp1/cpp-ethereum/libethcore/../libethash-cl/ethash_cl_miner.h:12, from /home/kabini/cpp1/cpp-ethereum/libethcore/Ethash.cpp:41: /usr/local/cuda/include/CL/cl_gl_ext.h:44:4: error: "/*" within comment [-Werror=comment] * /* cl_VEN_extname extension */ ^ [ 89%] Building CXX object libethcore/CMakeFiles/ethcore.dir/Miner.cpp.o [ 90%] Building CXX object libethcore/CMakeFiles/ethcore.dir/KeyManager.cpp.o cc1plus: all warnings being treated as errors make[2]: *** [libethcore/CMakeFiles/ethcore.dir/BlockInfo.cpp.o] Fehler 1 make[2]: *** Auf noch nicht beendete Prozesse wird gewartet … In file included from /home/kabini/cpp1/cpp-ethereum/libethcore/Ethash.cpp:22:0: /home/kabini/cpp1/cpp-ethereum/libethcore/Ethash.h: In static member function ‘static void dev::eth::Ethash::CUDAMiner::setDevices(unsigned int*, unsigned int)’: /home/kabini/cpp1/cpp-ethereum/libethcore/Ethash.h:177:24: error: comparison between signed and unsigned integer expressions [-Werror=sign-compare] for (int i = 0; i < gpuDeviceCount; i++) { ^ cc1plus: all warnings being treated as errors make[2]: *** [libethcore/CMakeFiles/ethcore.dir/Ethash.cpp.o] Fehler 1 make[1]: *** [libethcore/CMakeFiles/ethcore.dir/all] Fehler 2 make: *** [all] Fehler 2
edit: you can "solve" this by removing the "-Werror" parameter from cmake/EthCompilerSettings.cmake ( line 6 )
then eth is not built. but ethminer is built!
Here the Result for GTX 960 2GB // 8 GB Ram // Energy Saving CPU
as a next step i will play with the parameters.. you also mentioned to upgrade OpenCL to 1.2 .. how is that possible? i already have the newest Nvidia Drivers.. or do i need some ATI packages?
@Messiano84 , attached is the most recent Win64 build. It has a new feature, similar to the -d switch in ccminer:
ethminer -U --gpu-devices 0 2 3
will not use GPU 1.
@buechling , thank for the notice, i will fix the warnings where i can (I develop on windows and therefore am never blocked by get these warnings). Currently the CUDA miner only is compatible with ethminer. You can use it in combination with eth/geth.
most likely --gpu-workgroup-size 128 will give you somewhat better results. Sometimes --gpu-batch-size 20 will do a little bit, as will --gpu-mining-buffers 3 or 4.
OpenCL 1.2 should come with latest NVidia drivers, but may be Windows only still.
@Messiano84 , attached is the most recent Win64 build. It has a new feature, similar to the -d switch in ccminer:
ethminer -U --gpu-devices 0 2 3
will not use GPU 1.
@buechling , thank for the notice, i will fix the warnings where i can (I develop on windows and therefore am never blocked by get these warnings). Currently the CUDA miner only is compatible with ethminer. You can use it in combination with eth/geth.
most likely --gpu-workgroup-size 128 will give you somewhat better results. Sometimes --gpu-batch-size 20 will do a little bit, as will --gpu-mining-buffers 3 or 4.
OpenCL 1.2 should come with latest NVidia drivers, but may be Windows only still.
slight increase with the three parameters.. 10,8 Mhash..
as a next step i wanted to give windows a try.. i can use your uploaded x64-exe with -G and get the same 10,4 MHash.. but when i use -U, the programm crashes ( "Windows detected an error" ) do i need any additional software on windows? i have a fresh windows 10, Nvidia Drivers, Cuda Drivers.. something missing?
11 over 7 is a bit much indeed. Are you using similar workgroup size on both?
Or it's a driver thing. What card are you using?
I'm going to connect to ethpool to see for certain later today. Ethpool will calculate my hash rate and display it I think? So far i was using private geth server which doesn't report, however I did mine a block or two.
Also your miner has crashed hardly once for me. The only crash happened when the DAG was not generated. I had to start with cpu mining, until the first DAG was written.
Iuse default parameters.. but maybe I will try setting the workgroup.
Ive been trying to get my 750ti short card(no 6pin) working right compared to other benchs. With the opencl miner i get about 3.4Mh with the cuda and running -U I get about 5.3Mh. Improvement but quite a bit lower compared to others with the 750ti OC(im assuming 6pin). There shouldnt be that much of a difference so it must be something on my end. I used the ethereum(++)win64 latest.exe and i5 4690k, 16gb ram, win7 is the specs. Any ideas?
Oh yea funny thing too, Im also using a 290x with the cuda ethminer from ur zip. Gives me a slightly higher hash with -G than the one that came with the above exe. So good job on that side dunno what is different but its different.
@cidmo@antonio8 do you get similarly bad OpenCL performance with the official ethminer?
the older binaries here were built with CUDA 7.0, hence its dependency on cudart64_70.dll. i don't have these any longer either, but I wouldn't bother to try if they're faster, they're only marginally different.
i don't know what it is with these unwilling 750s, have yet to figure it out.
@Genoil I do get the same bad performance on the ethminer but can't say for sure as it is the official.
The only thing that I found out is for some reason your miner locks one of my cards up and I have to reboot where that does not happen on the regular etherminer.
Maybe it's a driver issue on my end and I am downloading the newest one for Windows 7 now and retrying. Also not sure if I have your latest miner but using the 20150806.
I also tried to recompile your source but can't find the .snl file anywhere.
- download/git clone source - cd cpp-ethereum - mkdir build - cd build - cmake -DBUNDLE=miner -DETHASHCU=1 [-DCOMPUTE=XX] -G "Visual Studio 12 2013 Win64" .. - launch generated .sln - set ethminer as starup project - build/run
[-DCOMPUTE=XX] is optional. You can use it to build only for specific version of CUDA Compute. i.e. for GTX 9x0 (Compute 5.2) specificy -DCOMPUTE52. It won't change performance, only decrease binary size and compile time.
Note for people that have hung around for a while. I removed the -DMAXREGCOUNT build parameter. Instead register usage can be tuned with __launch_bounds(max_threads, min_blocks). As I only have 1 working GTX780, I haven't been able to find optimal settings for other devices. My guess is that it won't be much different between GTX7x0 and GTX9x0 cards, as they are quite similar in this respect.
Something I just noticed on the 750ti computer. I am just using this to pool mine: -U --gpu-devices 0 1 2 3 4 but for some reason all of mu cpu cores are at 100%. Even if I just use device 0 That is not the case with my 960 rig though.
@antonio8, You should only need -t 5 for all GPU's instead of --gpu-devices. Can you see the GPU's being picked up when ethminer ramps up? Also, what CPU are you using? I have a Phenom II x4 925 and curiously it's jumped from 200% (2 of 4 cores) with 4 GPU's to 390% usage after I put in a 5th 750Ti yesterday.
@o0ragman0o , I will try the -t. I was just using the --gpu-devices from a post I read a few pages back that Genoil (I believe) posted.
On the gpu ramp up, when I start it and the cpu usage kicks to 100% it bogs the system down a lot and takes a while. Nvidia Inspector does show the load at 90-100% on some but some show 0% but the system is so bogged down I don't trust that they are not mining.
I am using a Phenom II X6 1100T with the 750ti's (bogs down) and in the 960 rig an I7 4790K (running smooth).
Also both rigs over 12 GB of memory so I ruled that out as an issue, or hopefully so. 750ti rig has 12 Gb of memory and the 960 rig has 16 GB of memory.
EDIT: The -t made no difference in the 750ti rig. Still bogs down a lot and cpu 100% on all cores. I also have 5 750ti's in that rig and only 3 960's with 1 750ti.
@o0ragman0o "--gpu-devices" is the new and improved "-t", allowing you to exactly specify which devices you want instead of from "0 to t"
@ptk are you using the exact same binary or did you try one of my newer builds?
Attached is a zip with 2 binaries. One uses shared memory to share data between GPU threads, the other uses warp shuffles. Could be that 750Ti has issues with the latter. On Kepler I get only .5% speed increase between the two.
I don't think the issue is lack of RAM or CPU, or number of cards. I have GTX750Ti on i5-4570 with 8GB that just crawls (ethminer only). When I take that 750 and plug it alongside my 780 in a 16GB Xeon E5, it easily does 8MH/s. Must be some driver or Windows thing...
I have also problem with 750ti and G1820T CPU. I get around 2.3Mh/s and afterburner shows 46% GPU usage. Task manager reports that ethminer has 46 to 50% CPU usage. I'm running on win10.
@Genoil I am not a dev by any means at all and can't code to save my life. With that said though I do have a question.
Was this build on compute_50 or compute_52?
The cuda-share did seem better as the hash rate for the 750ti's improved to about 3 Mh/s but did spike for a little while at about 8-9 Mh/s. The rig was still bogged down a lot but no issues like that with the 960.
I could not follow your instructions earlier to create the .sln file as when I compile they are already created.
the binaries provided have support from Compute 2.0 up to 5.2.
if you get only 3MH/s per 750Ti, there is a problem somewhere else than in shared vs shuffled. it should do 8-9MH/s. i mean, i get reports from others with the same problem that the shuffled version works better
Comments
Edit: Tested 1 single GFX on win7 and it works fine.
I compiled actual github version on Ubuntu 14.04 x64 ( cuda-repo-ubuntu1404_7.0-28_amd64.deb )
when i use "cmake -DETHASHCL=1 -DGUI=0",build is successfull. but when i use the "-U" Switch, only CPU is used. so Cuda is possibly not included.
then i tried your hint "cmake -DBUNDLE=miner -DETHASHCU=1". but then it will not compile: edit: you can "solve" this by removing the "-Werror" parameter from
cmake/EthCompilerSettings.cmake ( line 6 )
then eth is not built. but ethminer is built!
Here the Result for GTX 960 2GB // 8 GB Ram // Energy Saving CPU
Benchmarking on platform: { "platform": "NVIDIA CUDA", "device": "GeForce GTX 960", "version": "OpenCL 1.1 CUDA" }
Using device: GeForce GTX 960(OpenCL 1.1 CUDA)
Trial 1... 10398378
Trial 2... 10485760
Trial 3... 10398378
Trial 4... 10398378
Trial 5... 10485760
min/mean/max: 10398378/10433330/10485760 H/s
inner mean: 10427505 H/s
Benchmarking on platform: { "platform": "CUDA 7.0", "device": "GeForce GTX 960", "version": "Compute 5.2" }
Using device: GeForce GTX 960(5.2)
Trial 1... 10660522
Trial 2... 10660522
Trial 3... 10660522
Trial 4... 10660522
Trial 5... 10660522
min/mean/max: 10660522/10660522/10660522 H/s
inner mean: 10660522 H/s
so we increased from 10,4 to 10,6 Mhash
as a next step i will play with the parameters.. you also mentioned to upgrade OpenCL to 1.2 .. how is that possible? i already have the newest Nvidia Drivers.. or do i need some ATI packages?
ethminer -U --gpu-devices 0 2 3
will not use GPU 1.
@buechling , thank for the notice, i will fix the warnings where i can (I develop on windows and therefore am never blocked by get these warnings). Currently the CUDA miner only is compatible with ethminer. You can use it in combination with eth/geth.
most likely --gpu-workgroup-size 128 will give you somewhat better results. Sometimes --gpu-batch-size 20 will do a little bit, as will --gpu-mining-buffers 3 or 4.
OpenCL 1.2 should come with latest NVidia drivers, but may be Windows only still.
as a next step i wanted to give windows a try.. i can use your uploaded x64-exe with -G and get the same 10,4 MHash.. but when i use -U, the programm crashes ( "Windows detected an error" )
do i need any additional software on windows? i have a fresh windows 10, Nvidia Drivers, Cuda Drivers.. something missing?
error may be windows 10 related. the binary was built on windows 7 with cuda 7.5RC.
Also your miner has crashed hardly once for me. The only crash happened when the DAG was not generated. I had to start with cpu mining, until the first DAG was written.
Iuse default parameters.. but maybe I will try setting the workgroup.
Using Windows 7 64bit.
Any ideas?
@Genoil
I tried running some of the other versions and it says missing "cudart64_70.dll" and I tried googling but could not find it.
Oh yea funny thing too, Im also using a 290x with the cuda ethminer from ur zip. Gives me a slightly higher hash with -G than the one that came with the above exe. So good job on that side dunno what is different but its different.
the older binaries here were built with CUDA 7.0, hence its dependency on cudart64_70.dll. i don't have these any longer either, but I wouldn't bother to try if they're faster, they're only marginally different.
i don't know what it is with these unwilling 750s, have yet to figure it out.
I do get the same bad performance on the ethminer but can't say for sure as it is the official.
The only thing that I found out is for some reason your miner locks one of my cards up and I have to reboot where that does not happen on the regular etherminer.
Maybe it's a driver issue on my end and I am downloading the newest one for Windows 7 now and retrying. Also not sure if I have your latest miner but using the 20150806.
I also tried to recompile your source but can't find the .snl file anywhere.
- download/git clone source
- cd cpp-ethereum
- mkdir build
- cd build
- cmake -DBUNDLE=miner -DETHASHCU=1 [-DCOMPUTE=XX] -G "Visual Studio 12 2013 Win64" ..
- launch generated .sln
- set ethminer as starup project
- build/run
[-DCOMPUTE=XX] is optional. You can use it to build only for specific version of CUDA Compute. i.e. for GTX 9x0 (Compute 5.2) specificy -DCOMPUTE52. It won't change performance, only decrease binary size and compile time.
Note for people that have hung around for a while. I removed the -DMAXREGCOUNT build parameter. Instead register usage can be tuned with __launch_bounds(max_threads, min_blocks). As I only have 1 working GTX780, I haven't been able to find optimal settings for other devices. My guess is that it won't be much different between GTX7x0 and GTX9x0 cards, as they are quite similar in this respect.
One thing I do notice is when I find a "share" my hash rate is super high, around 50-84 Mh/s but drops fast to 9-12 Mh/s.
-t 5 for all GPU's
instead of --gpu-devices. Can you see the GPU's being picked up when ethminer ramps up? Also, what CPU are you using? I have a Phenom II x4 925 and curiously it's jumped from 200% (2 of 4 cores) with 4 GPU's to 390% usage after I put in a 5th 750Ti yesterday.On the gpu ramp up, when I start it and the cpu usage kicks to 100% it bogs the system down a lot and takes a while. Nvidia Inspector does show the load at 90-100% on some but some show 0% but the system is so bogged down I don't trust that they are not mining.
I am using a Phenom II X6 1100T with the 750ti's (bogs down) and in the 960 rig an I7 4790K (running smooth).
Also both rigs over 12 GB of memory so I ruled that out as an issue, or hopefully so. 750ti rig has 12 Gb of memory and the 960 rig has 16 GB of memory.
EDIT: The -t made no difference in the 750ti rig. Still bogs down a lot and cpu 100% on all cores. I also have 5 750ti's in that rig and only 3 960's with 1 750ti.
Maybe it is the amount of cards holding it back.
@ptk are you using the exact same binary or did you try one of my newer builds?
Attached is a zip with 2 binaries. One uses shared memory to share data between GPU threads, the other uses warp shuffles. Could be that 750Ti has issues with the latter. On Kepler I get only .5% speed increase between the two.
I don't think the issue is lack of RAM or CPU, or number of cards. I have GTX750Ti on i5-4570 with 8GB that just crawls (ethminer only). When I take that 750 and plug it alongside my 780 in a 16GB Xeon E5, it easily does 8MH/s. Must be some driver or Windows thing...
I am not a dev by any means at all and can't code to save my life. With that said though I do have a question.
Was this build on compute_50 or compute_52?
The cuda-share did seem better as the hash rate for the 750ti's improved to about 3 Mh/s but did spike for a little while at about 8-9 Mh/s. The rig was still bogged down a lot but no issues like that with the 960.
I could not follow your instructions earlier to create the .sln file as when I compile they are already created.
if you get only 3MH/s per 750Ti, there is a problem somewhere else than in shared vs shuffled. it should do 8-9MH/s. i mean, i get reports from others with the same problem that the shuffled version works better
it remains a mystery...
the problem with utilize videocard memory controller - only 10-30% of max
but for gtx970 memory controller utilized fully