CUDA miner

kps

I have updated and built a new ethminer, the hashrate didn't change for gtx 1080, but i've tried to build it now using -cudaminer argument but it won't let me says cuda is missing, is this an issue with driver or maybe i'm missing something ? I did install the cuda package mentioned in step by step guide on github. @Genoil do you think cuda would improve performance over opencl mining on ubuntu on GTX 1080s right now ?

JPS

@kps You did install 8.0 RC right.... not 7.5 ?

Genoil

@kps no, you'll have to wait for the Windows Update. Not sure when it will roll out, either next week or in July.

foobar31

Just a small and fast comment to say thanks to your work, @Genoil.
I stopped GPU mining 3 years go, but seeing what the eth community is still able to do is pretty exciting.
I've almost read all the 59 pages of this thread and have seen the good support you provide.

So, thanks and keep up that great work !
A bunch of eths coming to you as soon as I get my first reward in ~ 2 days

kps

@JPS it looks like i installed 7.5, that was the version in the guide, but i can't seem to find 8.0 version, the one on nvidia website requires membership of the the Accelerated Computing Developer Program, which i obviously don't have, and not sure if registering process requires me to do anything special but i'm gonna try that for now i guess, unless you have an alternative source of a package for ubuntu, then please share.

Genoil

@kps no that membership is free. but you can save yourself the hassle of going through that. you won't get any better performance until WDDM 2.1 is out and new drivers are shipped.

Bensam123

#3 sounds great dude. Being able to combine it with other miners or using them as drop ins and having ethminer take precendence and tweak them, would be great.

The 1xxx issues need to be dealt with as well, but right now I think that the GPUs are just too new and next month will be more telling as Nvidias own driver stability fixes and Windows updates come out for them.

JPS

Genoil said:

@kps no, you'll have to wait for the Windows Update. Not sure when it will roll out, either next week or in July.

Forgive my ignorance, but what does the windows update have to do with running the miner on Ubuntu?

kps

So i got some time, and i installed the CUDA 8 toolkit from nvidia website, but when trying to compile the ethminer with DBUNDLE=cudaminer make -j8 commands, i get error

"[ 35%] Building NVCC (Device) object libethash-cuda/CMakeFiles/ethash-cuda.dir/ethash-cuda_generated_ethash_cuda_miner_kernel.cu.o
nvcc fatal : redefinition of argument 'std'
CMake Error at ethash-cuda_generated_ethash_cuda_miner_kernel.cu.o.cmake:207 (message):
Error generating
/home/kps/Downloads/cpp-ethereum/build/libethash-cuda/CMakeFiles/ethash-cuda.dir//./ethash-cuda_generated_ethash_cuda_miner_kernel.cu.o


libethash-cuda/CMakeFiles/ethash-cuda.dir/build.make:63: recipe for target 'libethash-cuda/CMakeFiles/ethash-cuda.dir/ethash-cuda_generated_ethash_cuda_miner_kernel.cu.o' failed
make[2]: *** [libethash-cuda/CMakeFiles/ethash-cuda.dir/ethash-cuda_generated_ethash_cuda_miner_kernel.cu.o] Error 1
CMakeFiles/Makefile2:320: recipe for target 'libethash-cuda/CMakeFiles/ethash-cuda.dir/all' failed
make[1]: *** [libethash-cuda/CMakeFiles/ethash-cuda.dir/all] Error 2
make[1]: *** Waiting for unfinished jobs....
Makefile:127: recipe for target 'all' failed
make: *** [all] Error 2

"

Clearly looks like some cuda related problem, used default settings for everything in cuda installer, normal opencl miner compiles without any problems, and i have also updated driver to 367.27

edit: looks like 367.27 driver is reducing hashrate, dropped to about 40 mhash/s from 45, revert time...

rdp

Hi guys, I'm having a bit of trouble with a 2x 980 Ti system and I'm hoping someone can tell me what I'm doing wrong.

I have tried Genoil's ethminer both in CUDA mode and in OpenCL mode, and in either case I see the same behavior. I've tried 1.1.5 and 1.0.8. Both show the same error.

Generally speaking, I can mine on gpu0 but not on gpu1. I get all kinds of errors from gpu1 depending on what kind of build I'm running.

This happens both from Win 10 and in Ubuntu 14.04.

Significantly, on Win 10 if I plug the monitor into gpu1, then gpu1 works for mining but gpu0 is the one that starts throwing errors.

So it seems like only the GPU that has a monitor connected can mine, and the one without a monitor cannot. There is nothing wrong with either GPU, they are both working fine in Win 10 SLI mode and during benchmarks they are performing as expected.

OpenCL results (cmake -DBUNDLE=miner ..)

Here is sample ethminer output where I run ONLY gpu0 in OpenCL mode and it is successful:

[email protected]:/usr/local/src/cpp-ethereum/build$ ./ethminer/ethminer -G -F http://127.0.0.1:8080/$(hostname) --farm-recheck 200 --opencl-device 0
Genoil's ethminer 0.9.41-genoil-1.0.8
=====================================================================
Forked from github.com/ethereum/cpp-ethereum
CUDA kernel ported from Tim Hughes' OpenCL kernel
With contributions from nerdralph, RoBiK, tpruvot and sp_ 

Please consider a donation to:
ETH: 0xeb9310b185455f863f526dab3d245809f6854b4d

[OPENCL]:Found suitable OpenCL device [GeForce GTX 980 Ti] with 6441730048 bytes of GPU memory
  m  18:35:10|ethminer  Getting work package...
  m  18:35:10|ethminer  Grabbing DAG for #a1cf1239…
  m  18:35:16|ethminer  Got work package: #de595158
  ℹ  18:35:16|ethminer  Loading full DAG of seedhash: #8f602dc7…
  ℹ  18:35:16|gpuminer0  set work to: #de595158, target  #0000000112e0be82
  ℹ  18:35:16|gpuminer0  Initialising miner...
  m  18:35:16|ethminer  Mining on PoWhash #de595158 : 0.00MH/s [A0+0:R0+0:F0]
[OPENCL]:Using platform: NVIDIA CUDA
[OPENCL]:Using device: GeForce GTX 980 Ti(OpenCL 1.2 CUDA)
[OPENCL]:Printing program log
[OPENCL]:

[OPENCL]:Creating one big buffer for the DAG
[OPENCL]:Loading single big chunk kernels
[OPENCL]:Mapping one big chunk.
  m  18:35:21|ethminer  Mining on PoWhash #de595158 : 0.00MH/s [A0+0:R0+0:F0]
[OPENCL]:Creating buffer for header.
[OPENCL]:Creating mining buffer 0
[OPENCL]:Creating mining buffer 1
  m  18:35:21|ethminer  Mining on PoWhash #de595158 : 5.22MH/s [A0+0:R0+0:F0]
  m  18:35:21|ethminer  Mining on PoWhash #de595158 : 20.87MH/s [A0+0:R0+0:F0]
  m  18:35:22|ethminer  Mining on PoWhash #de595158 : 20.87MH/s [A0+0:R0+0:F0]
  m  18:35:22|ethminer  Mining on PoWhash #de595158 : 20.87MH/s [A0+0:R0+0:F0]
  m  18:35:22|ethminer  Mining on PoWhash #de595158 : 22.17MH/s [A0+0:R0+0:F0]

Here is exactly the same command, but trying to use gpu1 which gives an error:

(Error is: [OPENCL]:clEnqueueMapBuffer(-36))

[email protected]:/usr/local/src/cpp-ethereum/build$ ./ethminer/ethminer -G -F http://127.0.0.1:8080/$(hostname) --farm-recheck 200 --opencl-device 1
Genoil's ethminer 0.9.41-genoil-1.0.8
=====================================================================
Forked from github.com/ethereum/cpp-ethereum
CUDA kernel ported from Tim Hughes' OpenCL kernel
With contributions from nerdralph, RoBiK, tpruvot and sp_ 

Please consider a donation to:
ETH: 0xeb9310b185455f863f526dab3d245809f6854b4d

[OPENCL]:Found suitable OpenCL device [GeForce GTX 980 Ti] with 6441730048 bytes of GPU memory
  m  18:35:30|ethminer  Getting work package...
  m  18:35:30|ethminer  Grabbing DAG for #a1cf1239…
  m  18:35:35|ethminer  Got work package: #de595158
  ℹ  18:35:35|ethminer  Loading full DAG of seedhash: #8f602dc7…
  ℹ  18:35:35|gpuminer0  set work to: #de595158, target  #0000000112e0be82
  ℹ  18:35:35|gpuminer0  Initialising miner...
  m  18:35:35|ethminer  Mining on PoWhash #de595158 : 0.00MH/s [A0+0:R0+0:F0]
[OPENCL]:Using platform: NVIDIA CUDA
[OPENCL]:Using device: GeForce GTX 980 Ti(OpenCL 1.2 CUDA)
[OPENCL]:Printing program log
[OPENCL]:

[OPENCL]:Creating one big buffer for the DAG
[OPENCL]:Loading single big chunk kernels
[OPENCL]:Mapping one big chunk.
  m  18:35:41|ethminer  Mining on PoWhash #de595158 : 0.00MH/s [A0+0:R0+0:F0]
[OPENCL]:Creating buffer for header.
[OPENCL]:Creating mining buffer 0
[OPENCL]:Creating mining buffer 1
[OPENCL]:clEnqueueMapBuffer(-36)
  m  18:35:41|ethminer  Mining on PoWhash #de595158 : 0.00MH/s [A0+0:R0+0:F0]
  m  18:35:41|ethminer  Mining on PoWhash #de595158 : 0.00MH/s [A0+0:R0+0:F0]
  m  18:35:41|ethminer  Mining on PoWhash #de595158 : 0.00MH/s [A0+0:R0+0:F0]
  m  18:35:41|ethminer  Mining on PoWhash #de595158 : 0.00MH/s [A0+0:R0+0:F0]

CUDA Results (cmake -DBUNDLE=miner -DETHASHCUDA=on ..)

[email protected]:/usr/local/src/cpp-ethereum/build$ ./ethminer/ethminer -U --list-devices
Genoil's ethminer 0.9.41-genoil-1.0.8
=====================================================================
Forked from github.com/ethereum/cpp-ethereum
CUDA kernel ported from Tim Hughes' OpenCL kernel
With contributions from nerdralph, RoBiK, tpruvot and sp_ 

Please consider a donation to:
ETH: 0xeb9310b185455f863f526dab3d245809f6854b4d

[CUDA]:
Listing CUDA devices.
FORMAT: [deviceID] deviceName
[0] GeForce GTX 980 Ti
        Compute version: 5.2
        cudaDeviceProp::totalGlobalMem: 6441730048
[1] GeForce GTX 980 Ti
        Compute version: 5.2
        cudaDeviceProp::totalGlobalMem: 6442254336

Again gpu0 is working fine:

[email protected]:/usr/local/src/cpp-ethereum/build$ ./ethminer/ethminer -U -F http://127.0.0.1:8080/$(hostname) --farm-recheck 200 --cuda-devices 0
Genoil's ethminer 0.9.41-genoil-1.0.8
=====================================================================
Forked from github.com/ethereum/cpp-ethereum
CUDA kernel ported from Tim Hughes' OpenCL kernel
With contributions from nerdralph, RoBiK, tpruvot and sp_ 

Please consider a donation to:
ETH: 0xeb9310b185455f863f526dab3d245809f6854b4d

[CUDA]:Using grid size 8192, block size 128

[CUDA]:Found suitable CUDA device [GeForce GTX 980 Ti] with 6441730048 bytes of GPU memory
  m  18:48:02|ethminer  Getting work package...
  m  18:48:02|ethminer  Grabbing DAG for #a1cf1239…
  m  18:48:07|ethminer  Got work package: #4c9398e7
  ℹ  18:48:07|ethminer  Loading full DAG of seedhash: #8f602dc7…
  ℹ  18:48:07|cudaminer0  set work to #4c9398e7, target  #0000000112e0be82
  ℹ  18:48:07|cudaminer0  Initialising miner...
  m  18:48:07|ethminer  Mining on PoWhash #4c9398e7 : 0.00MH/s [A0+0:R0+0:F0]
  m  18:48:12|ethminer  Got work package: #c12a4904
Using device: GeForce GTX 980 Ti (Compute 5.2)
  ℹ  18:48:13|cudaminer0  set work to #c12a4904, target  #0000000112e0be82
  m  18:48:13|ethminer  Mining on PoWhash #c12a4904 : 10.49MH/s [A0+0:R0+0:F0]
  m  18:48:13|ethminer  Got work package: #af9be788
  ℹ  18:48:13|cudaminer0  set work to #af9be788, target  #0000000112e0be82
  m  18:48:13|ethminer  Mining on PoWhash #af9be788 : 10.49MH/s [A0+0:R0+0:F0]
  m  18:48:13|ethminer  Mining on PoWhash #af9be788 : 20.87MH/s [A0+0:R0+0:F0]
  m  18:48:14|ethminer  Mining on PoWhash #af9be788 : 20.87MH/s [A0+0:R0+0:F0]
  m  18:48:14|ethminer  Mining on PoWhash #af9be788 : 20.87MH/s [A0+0:R0+0:F0]
  m  18:48:14|ethminer  Mining on PoWhash #af9be788 : 26.08MH/s [A0+0:R0+0:F0]

But gpu1 is not, it has a different error this time though:

[email protected]:/usr/local/src/cpp-ethereum/build$ ./ethminer/ethminer -U -F http://127.0.0.1:8080/$(hostname) --farm-recheck 200 --cuda-devices 1
Genoil's ethminer 0.9.41-genoil-1.0.8
=====================================================================
Forked from github.com/ethereum/cpp-ethereum
CUDA kernel ported from Tim Hughes' OpenCL kernel
With contributions from nerdralph, RoBiK, tpruvot and sp_ 

Please consider a donation to:
ETH: 0xeb9310b185455f863f526dab3d245809f6854b4d

[CUDA]:Using grid size 8192, block size 128

[CUDA]:Found suitable CUDA device [GeForce GTX 980 Ti] with 6442254336 bytes of GPU memory
  m  18:48:38|ethminer  Getting work package...
  m  18:48:38|ethminer  Grabbing DAG for #a1cf1239…
Generally speaking, I can mine on gpu0 but not on gpu1. I get all kinds of errors from gpu1 depending on what kind of build I'm running.

  m  18:48:44|ethminer  Got work package: #1eb68f02
  ℹ  18:48:44|ethminer  Loading full DAG of seedhash: #8f602dc7…
  ℹ  18:48:44|cudaminer0  set work to #1eb68f02, target  #0000000112e0be82
  ℹ  18:48:44|cudaminer0  Initialising miner...
  m  18:48:44|ethminer  Mining on PoWhash #1eb68f02 : 0.00MH/s [A0+0:R0+0:F0]
  m  18:48:49|ethminer  Got work package: #5e691686
Using device: GeForce GTX 980 Ti (Compute 5.2)
CUDA error in func 'search' at line 290 : an illegal memory access was encountered.
  ✘  18:48:49|cudaminer0  Error CUDA mining: an illegal memory access was encountered

Any ideas how I can change this?

I've tried with and without the SLI bridge, that had no impact.

Thanks in advance for any insight you can provide.

Genoil

bitcanuck said:

The master (1.0.8) build works fine as expected. Reverting to off-GPU dag generation is quite painful...

1.1.5 also works fine when I comment out the nicehash stratum checks in EthStratumClient.cpp.

It seems kind of weird that not everybody complains about this. Or did everybody stop mining until the DAO storm is over? Haven't been able to reproduce this, perhaps it's Linux only?

ethfan

I have been running older versions (1.1.0, when first in beta or thereabouts) for quite a while. Eg. one rig has submitted >40,000 shares (at ethermine). Just one test rig on the latest version which also has not caused any problems. No problems due to the miner anyway. Issues related to under-volting and temp control, yes. But nothing due to the miner itself that I can see. All under Linux.

Genoil

string bug is now fixed in 1.1.6

bbcoin

Hey boys looks who's back

Just fired up a bunch of brand spanking new r7 370s with 1.1.3 + proxy.
Everything worked on the first start.

@Genoil do you think running stratum inside your client will improve performance over dwarfs proxy? No difference? How to point your client to run proxy on a pool with no registration/just wallet address? Seems like it constantly waits for work.

PS. DAG generation is very fast good job.

PS2. those 370's even the cheapo single fan cards OC very well and run at 64C' max... Considering it's summer, I'm loving it

Genoil

@bbcoin can you give an example of what you are trying to achieve?

I'm planning to support dwarf stratum native but haven't got much time lately.

bbcoin

@genoil simply trying to point your client at -S eth.f2pool.com:8008 but when I put my wallet as the -O it never gets any work. How would I set this up without dwarfs proxy?

donaldtrump2016

Here are my findings regarding the 750ti under windows:

Running in benchmark mode:
- one 750ti - Memory controller load at 100%, ethminer uses 90MB of RAM, speed = 8.8Mh;
- two 750ti's - MCU at 100% for both cards, ethminer uses 146MB of RAM, speed = 17.5Mh;
- three 750ti's - MCU at 100%, 10%, 100% for the three cards, ethminer uses 1232MB of RAM!, speed = 18.3Mh;
- four 750ti's - MCU at 100%, 5%, 7%, 100% for the four cards, ethminer uses 1282MB of RAM, speed = 18.5Mh;
- five 750ti's - MCU at 100% for all five cards, ethminer uses 1335MB of RAM, speed = 43.5Mh!;

I have no skills in coding. Maybe this helps in resolving the windows bug...

kps

I have compiled cudaminer in 16.04 with compute 61 on GTX 1080, and i've been trying to tweak settings, however any change from default 128 to --cuda-block-size causes dag to fail loading resulting in error,

Cuda error in func 'ethash_generate_dag' at line 133 : invalid argument.

When changing --cuda-block-size to 256, it loads from 0 to roughly 90% within a second and crashes, i also use --farm-recheck 200, are there any options i could use to optimize mining speed ?

Genoil

Block size should be a power of 2, so 256 should work.

But I should make it possible to use different sizes for generating/mining. Will add this soon.

IronFists

eth100tx said:

looks like it's the device driver crashing...

Eth100tx set a start bat with a timer that starts the ethminer bat .. I set mine to 300 seems to be the sweet spot form driver crashing. and if it does it will restart the miner on its own after the driver refreshes it self. keeps needless idling from happening.

Genoil

1.1.7 is out. Adds support for pools using eth-proxy-style stratum. Main reason to support it is because of the upcoming choice between soft-fork or not. Supporting more pools will allow users to choose between more pools that choose the path they support, without having to change to a different mining client.

Removed flags:
-SV (replaced by -SC)
-ES (superseded by -SP)

Added flags
-SC [n] Where n is 1 (async, default) or 2 (sync) stratum client. async uses a bit more CPU but has working --work-timeout, wheras sync client doesn't have that (yet).
-SP [n] Where n is 0 (stratum, default), 1 (eth-proxy) or 2 (nicehash). Stratum protocol to use.
-SE [e] Where e is email address. Used by eth-proxy. In some pools (like nano) you can also add it to the worker name using /.

Example dwarfpool:

ethminer.exe -U -S eth-eu.dwarfpool.com:8008 -SP 1 -O 0xeb9310b185455f863f526dab3d245809f6854b4d.worker -SE [email protected]

https://github.com/Genoil/cpp-ethereum/tree/110/releases

Yes it could still have bugs, tested major pools (dwarf, f2, nano) with success, but something may have slipped through. Thanks for testing!

Genoil

bitcanuck said:

I don't think there's much left that you can improve, except maybe optimization of the opencl kernel so that it is just as fast as @Wolf0 :-)

If I'm going down that road, I'll do it for the green camp. Highly unlikely though

Mo35

latest 1.1.7 compiled with -DCOMPUTE=52, for Cuda 6.5 , OpenCL and Stratum support included

http://wikisend.com/download/667542/ethminer-0.9.41-genoil-1.1.7-cuda6.5.zip

p.s. downloads TTL 90 days

Genoil

Can anybody with a 1080 or 1070 and Win10 try to get 368.20 drivers running? They are installed by Windows Update and may require a Windows Insider subscription. When running dxdiag with 368.20, WDDM version is upgraded from 2.0 to 2.1. No TLB fix for my 970, but it may solve the issue on Pascal cards.

Genoil

Mo35 said:

latest 1.1.7 compiled with -DCOMPUTE=52, for Cuda 6.5 , OpenCL and Stratum support included

http://wikisend.com/download/667542/ethminer-0.9.41-genoil-1.1.7-cuda6.5.zip

p.s. downloads TTL 90 days

thanks, copied it to the repo

saiphir

@Genoil thanks again for your great CUDA fork.

Running 2x GTX 1080 in ubuntu 14.4 with CUDA 8.0 and 367.27 NVIDIA drivers. Just a few heads up for people with similar setups getting stuck.

When compiling Genoil's fork don't forget to use -DCOMPUTE=61 as Pascal architecture is running compute 6.1

I've seen the highest hashrate results so far messing around these parameters

./ethminer -G --farm-recheck 100 -F (eth-proxy/poolurl without brackets) -U --cuda-devices 0 1 --cuda-grid-size 8192 --cuda-block-size 128 --cuda-schedule sync --cl-global-work 16384

From a fresh install set nomodeset in grub and first install 367 NVIDIA drivers and dependencies, they only work on the most recent drivers I've noticed. After don't forget to run the nvidia-xconfig for xorg.conf. Reboot and run the CUDA 8.0 local .deb. When installing CUDA 8.0 don't forget the environmental variables export LD_LIBRARY_PATH .

For some reason apt-get install CUDA forces older drivers to be installed and messes up your x so make sure to NOT install those display drivers when apt-get or installing through cuda_8.0.27_linux.run (the .run for some reason did not work for me hence using .deb)

Anyone know how to voltage/power OC in ubuntu 14.4 or have any tweaks/driver/fork tweaks to up the MH/s, my 970s we're running similar MH/s rates.

Keep up securing the blockchain fellow miners o7

Oh and there is a CUDA 8.0 man but you have to get accepted for the NVIDIA Accelerated Computing Developer Program and I can't attach it here because of that so apply they have some cool stuff there: https://developer.nvidia.com/accelerated-computing-developer

JPS

How about making it a bit easier and putting the compiled version up on let's say Zippyshare for example ?

Genoil

@saiphir where did you find the CUDA 8 manual/documentation?

Smokyish

saiphir said:

Anyone know how to voltage/power OC in ubuntu 14.4 or have any tweaks/driver/fork tweaks to up the MH/s, my 970s we're running similar MH/s rates.

You need to enable it in the Xorg conf file at /etc/X11/xorg.conf and enable OPTION "Coolbits" "XX" ie. OPTION "Coolbits" "28" to enable manual fan speed, over/underclocking and over/undervoltage.

https://wiki.archlinux.org/index.php/NVIDIA/Tips_and_tricks#Enabling_overclocking
(applies also to Ubuntu even though its arch linux wiki)

saiphir

Genoil said:

@saiphir where did you find the CUDA 8 manual/documentation?

When you go to the download page of CUDA 8.0 and select the .deb or .run file, below it gives you a link to the manual. I know it's kind of hidden away, I'll just attach it here because they did such a bad job at providing it.