The high CPU load also seem to stem from the cuda_sync blocking routines. I usually work around this by tying all my threads to a single CPU. This reduces power and cpu load and has no effect on hashing since the CPU cycles are all busy waiting.
@ennui, I'm all ears here... How does one go about tying my threads to a single CPU (single core?)
I can't check right now as my rig's 700W PSU just inexplicably died while only drawing 480W . Worse, I replaced it with a brand new one, fresh out of the box and it sh!t itself too! So I'm out of the game for a few days. I've got a 1200W PSU coming in but that was going to be for my HD7990 rig...
I can't check right now as my rig's 700W PSU just inexplicably died while only drawing 480W . Worse, I replaced it with a brand new one, fresh out of the box and it sh!t itself too! So I'm out of the game for a few days. I've got a 1200W PSU coming in but that was going to be for my HD7990 rig...
@Genoil : Hey , mate , did you try compile the miner withh cuda 6.5. It might get descent performance , more that 7 and 7.5. Unfortunately my compilation skills are week to try that...
/path/to/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++) {
I do see : 'cudaminer0' rather than 'gpuminer0' reported on workloop, but avg is pretty much 17.1MH/s, so not much diff. Launched thusly: ./ethminer -U --gpu-workgroup-size 128 --gpu-batch-size 20
I've altered the settings, doesn't seem to make much diff so far. --gpu-workgroup-size maybe adds 1MH/s, not sure. Q: where to read about the 'compute=35' and so on, what doc(s) cover that?
Anyone have a compiled source for Windows? I can't figure out how to convert it from the source for the .snl
Genoil showed me the steps but that is like a foreign language for me.
This version was compiled on Win10 with CUDA 6.5. However it doesn't run at full speed for me. See if it works for you. This has a mod so cpu utilization is zero.
Anyone have a compiled source for Windows? I can't figure out how to convert it from the source for the .snl
Genoil showed me the steps but that is like a foreign language for me.
This version was compiled on Win10 with CUDA 6.5. However it doesn't run at full speed for me. See if it works for you. This has a mod so cpu utilization is zero.
using the --gpu-devices, No matter what number I type in it uses the same gpu. I have 4 cards in my system so I wanted to test what that 750ti does mixed with my 960's. I even used --gpu-devices 12 and it usess the same 960 card. So no matter if I do 0 or 1 or 2 or 3 it chooses My gpu 2 which is a 960.
My total hash is is about 36 Mh/s but I wanted to try and choose each card to see what the difference is with cards in the pci-e slot versus being in risers.
EDIT: Was able to use --gpu-devices 0 2 3 and stop my 750ti and hash is at 33.8 Mh/s so my 750ti is doing about 2 Mh/s on that system.
I have a bunch of 750Tis running on Ubuntu; noticed speeds around 8.8Mhs to 9.8Mhs (PNY stock OC). Was able to use nvidia-settings to OC it to 1.28GHz/6.6Ghz for about 10.6Mhs. I used CUDA7.0.
The high CPU load also seem to stem from the cuda_sync blocking routines. I usually work around this by tying all my threads to a single CPU. This reduces power and cpu load and has no effect on hashing since the CPU cycles are all busy waiting.
Also, one interesting power saving trick I noticed was that if you set the power limit of the 750Ti to 30W from 38.5W, the hash rate does not drop significantly (~5%?), but the power is massively reduced. For my rig of 6 750Tis, I observed a Kill-A-Watt drop 520W to 430W, and it's still currently running at a hash rate of about 51MH/s (all stock cards) and 56MH/s (all OC cards).
@ennui, I got my system stable again last night and was playing with the power settings as you suggested. I recorded a 0.2% drop in hashrate (43.0-> 42.9) for a 6.1% drop in power consumption. Then my PSU inexplicably died as did the replacement as soon as I started hashing (!!!???). Admittedly they were cheap ebay buys but WTF. 2x 12V rails at 25A and 28A. At 300W, even the lesser rail has enough power for 5 GPU's. It advertises current and over volt protection so I'm pretty pissed at it. The second PSU was only 550W and it went in the same way.
Can I ask you what PSU's and connectors you are using on your 750 rigs?
@Genoil Just tried out the binary you posted back on page 5, on Windows 10 with a 660GTX 2GB. Using -U I'm seeing hashrates of about 9.2MH/s compared to 7.4MH/s with stock ethminer, and verified with ethpool.org that the hashing is correct. Nice work!
I am another unlucky 750ti user. I have tried all of the versions and cannot get it work over 2.xMH. It is in win10 pc together with 970. The 970 makes 18+MH, both togehter 20.5MH and I noticed also a weird thing with --gpu-devices parameter. It just cannot make to work only 750ti. If i use --gpu-devices 0 or --gpu-devices 1 it makes no difference and runs only 970.
I have a bunch of 750Tis running on Ubuntu; noticed speeds around 8.8Mhs to 9.8Mhs (PNY stock OC). Was able to use nvidia-settings to OC it to 1.28GHz/6.6Ghz for about 10.6Mhs. I used CUDA7.0.
The high CPU load also seem to stem from the cuda_sync blocking routines. I usually work around this by tying all my threads to a single CPU. This reduces power and cpu load and has no effect on hashing since the CPU cycles are all busy waiting.
Also, one interesting power saving trick I noticed was that if you set the power limit of the 750Ti to 30W from 38.5W, the hash rate does not drop significantly (~5%?), but the power is massively reduced. For my rig of 6 750Tis, I observed a Kill-A-Watt drop 520W to 430W, and it's still currently running at a hash rate of about 51MH/s (all stock cards) and 56MH/s (all OC cards).
@ennui, I got my system stable again last night and was playing with the power settings as you suggested. I recorded a 0.2% drop in hashrate (43.0-> 42.9) for a 6.1% drop in power consumption. Then my PSU inexplicably died as did the replacement as soon as I started hashing (!!!???). Admittedly they were cheap ebay buys but WTF. 2x 12V rails at 25A and 28A. At 300W, even the lesser rail has enough power for 5 GPU's. It advertises current and over volt protection so I'm pretty pissed at it. The second PSU was only 550W and it went in the same way.
Can I ask you what PSU's and connectors you are using on your 750 rigs?
Not quite sure, I'm using a mix of cheap PSUs from China as well all around 800W (though they did advertise theirs at 80 plus), your situation could really be just real bad luck.
@mrp@VeritasSapere@antonio8@SuchFakeAccount My Rig is running with 3x GTX 750 Ti ( no risers, just Mainboard with 3 x16-Slots ) Idle: 54 W CPU under load: 110 W 1 GPU ( and 1 core 100% ): 9,3 Mhash @ 174 W 2 GPU ( and 2 cores 100% ): 18,6 Mhash @ 255 W 3 GPU ( and 3 cores 100% ): 27,9 Mhash @ 355 W
so everything seems fine.. as i use ubuntu 14.03 , i can not tell anything about frequencies etc.. i just installed nvidia drivers..
@Genoil, Thanks. I didn't realise the full nature of the --gpu-devices switch.
@antonio8. Can you do a study up on your PSU. I think your cards might be getting current limited. My rig currently has:
3x Galaxy 750Ti slim
1x emTek 750Ti
1x Gigabyte 750 OC Low profile.
pulling 43MH/s
700W Casecom PSU with 2 12v rails (CC-700W-12CM)
485W at the wall
Issues I had were mainly with the Gigabyte 750 dropping out. It doesn't seem to like the powered risers but works fine pulling power from the PCIe slot itself. Not a problem with the other cards on powered risers. Remember there's only a total of about 175W avaliable to the PCIe slots so 3 Ti's at TDP of 60W will get current limited.
I'll also note that the full size card is running 'much' cooler and faster than the low profiles.
As the 750/Ti's generally don't come with a 6/8 PCIe power connector, I'm left just powering them on Molex connectors. The question mark here is that I don't know which 12V rail the Molex connectors are on so there was a bit of trial, error and hope they weren't all on the same one which would have limited the number of cards without hacking together a PCIe -> Molex adapter.
It's been running nice now. Going from 4 to 5 cards jumped CPU usage from 200% to 400% which was unexpected.
I'm still waiting on one more Gigabyte 750Ti Low profile. I'll then rebuild the rig onto a 6 PCIe slot mobo I received yesterday. The current 5 slot mobo will be retasked with 2 (3 if I can land it) HD 7990's
i don't think there is a 175W limit? ( see above ). I think the limit is 75 W per PEG-Port ( special kind of PCI-X-Port, german: http://www.heise.de/ct/hotline/Was-ist-PEG-325734.html ), so we have 3*75W = 225W . i thought about connecting 6 cards per mainboard.. but these seld-built risers seem to be very risky.. so i will just use 3 cards per mainbaord, without risers..
The high CPU load also seem to stem from the cuda_sync blocking routines. I usually work around this by tying all my threads to a single CPU. This reduces power and cpu load and has no effect on hashing since the CPU cycles are all busy waiting.
@ennui, I'm all ears here... How does one go about tying my threads to a single CPU (single core?)
Also, one interesting power saving trick I noticed was that if you set the power limit of the 750Ti to 30W from 38.5W, the hash rate does not drop significantly (~5%?), but the power is massively reduced. For my rig of 6 750Tis, I observed a Kill-A-Watt drop 520W to 430W, and it's still currently running at a hash rate of about 51MH/s (all stock cards) and 56MH/s (all OC cards).
interesting! can you tell us: how can you set a power limit?
Genoli -- many thanks for this thread. Took me a couple of days of tweaking to get this up and running, but I have one rig with 6 x 750 ti, linux mint 17, 8gb ddr3, 3.1gz haswell cpu running at 54M+ and best of all, it's coming in at 390 watts at the wall... love that 750 ti!
ℹ 23:58:25|ethminer Mining on PoWhash #9e3b5328… : 54143339 H/s = 471859200 hashes / 8.715 s
also interesting! can you tell us how you achieved the low power usage?
The gpu-devices flag was quickly thrown together without actually having multiple cards, so I'll have to look at it again when I got a machine with 2 cards.
Hashrate-wise, on my GTX780 it unfortunately doesn't do any better than what was already there. Actually a tiny bit worse (caused by the cudaDeviceScheduleBlockingSync). But apparently that dramatically lowers CPU usage, which is a good thing! No solution for the GTX750Ti problem yet..
Comments
ps -eL | grep cudaminer | grep -v grep | awk '{ print "taskset -cp 1 " $2 }' | bash
I can't check right now as my rig's 700W PSU just inexplicably died while only drawing 480W
thx, I built as you recommended:
cmake .. -DBUNDLE=miner -DETHASHCU=1 -DCOMPUTE=35
I had a bit of trouble at these (2) points:
/path/to/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++) {
( solved by s/int/unsigned/ )
/usr/local/cuda-7.0/include/CL/cl_gl_ext.h:44:4: error: "/*" within comment [-Werror=comment]
* /* cl_VEN_extname extension */
( solved by * /\* cl_VEN_extname extension */ )
I do see : 'cudaminer0' rather than 'gpuminer0' reported on workloop, but avg is pretty much 17.1MH/s, so not much diff.
Launched thusly: ./ethminer -U --gpu-workgroup-size 128 --gpu-batch-size 20
I've altered the settings, doesn't seem to make much diff so far. --gpu-workgroup-size maybe adds 1MH/s, not sure.
Q: where to read about the 'compute=35' and so on, what doc(s) cover that?
thx!
https://bitcointalk.org/index.php?topic=826901.msg12157142#msg12157142
Genoil showed me the steps but that is like a foreign language for me.
Says no files to extract
Cpu usage at 100% though and hash at about 2 Mh/s
Something weird I just noticed.
using the --gpu-devices, No matter what number I type in it uses the same gpu. I have 4 cards in my system so I wanted to test what that 750ti does mixed with my 960's. I even used --gpu-devices 12 and it usess the same 960 card. So no matter if I do 0 or 1 or 2 or 3 it chooses My gpu 2 which is a 960.
My total hash is is about 36 Mh/s but I wanted to try and choose each card to see what the difference is with cards in the pci-e slot versus being in risers.
EDIT: Was able to use --gpu-devices 0 2 3 and stop my 750ti and hash is at 33.8 Mh/s so my 750ti is doing about 2 Mh/s on that system.
Windows 7 64bit driver 353.30
seems like your ethminer has settled out to run about 17.6MH/s and I did see 18MH/s at least once, looks like
it is improved.
Can I ask you what PSU's and connectors you are using on your 750 rigs?
I am another unlucky 750ti user. I have tried all of the versions and cannot get it work over 2.xMH. It is in win10 pc together with 970. The 970 makes 18+MH, both togehter 20.5MH and I noticed also a weird thing with --gpu-devices parameter. It just cannot make to work only 750ti. If i use --gpu-devices 0 or --gpu-devices 1 it makes no difference and runs only 970.
BTW has anyone built a "sp modded" genoil miner for win?
My Rig is running with 3x GTX 750 Ti ( no risers, just Mainboard with 3 x16-Slots )
Idle: 54 W
CPU under load: 110 W
1 GPU ( and 1 core 100% ): 9,3 Mhash @ 174 W
2 GPU ( and 2 cores 100% ): 18,6 Mhash @ 255 W
3 GPU ( and 3 cores 100% ): 27,9 Mhash @ 355 W
so everything seems fine.. as i use ubuntu 14.03 , i can not tell anything about frequencies etc.. i just installed nvidia drivers..
i don't think there is a 175W limit? ( see above ). I think the limit is 75 W per PEG-Port ( special kind of PCI-X-Port, german: http://www.heise.de/ct/hotline/Was-ist-PEG-325734.html ), so we have 3*75W = 225W . i thought about connecting 6 cards per mainboard.. but these seld-built risers seem to be very risky.. so i will just use 3 cards per mainbaord, without risers.. great! it works. thank you! interesting! can you tell us: how can you set a power limit? also interesting! can you tell us how you achieved the low power usage?
Attached is a Win64 binary with sp_ mods.
Hashrate-wise, on my GTX780 it unfortunately doesn't do any better than what was already there. Actually a tiny bit worse (caused by the cudaDeviceScheduleBlockingSync). But apparently that dramatically lowers CPU usage, which is a good thing! No solution for the GTX750Ti problem yet..