[ANN] cudaMiner & ccMiner CUDA based mining applications [Windows/Linux/MacOSX]

[Number6]

My launch config is -m0 -l T5x24 -I 0 -C 1 -H 1

upper case I is not legal, or not the intended option. -C 1 is ignored by the T kernel.

does it crash when going above +50 MHz overclock?

Christian

Thanks, I'll fix those settings. No it does not crash, and it really seems to have little to no impact on my hash rate. Using Precision-X and I see the frequencies go up, but the hash rates seem to stay consistent at 280 kHash even when going to +100 on core and +500 on memory.

Ok, I found out why the overclock settings were not having any effect, seems I was hitting the Power Target Limit as explained here: http://cryptomining-blog.com/tag/gtx-750-ti-cudaminer/

I think I read that before I got my cards, but in the excitement of building and all I forgot about it.

Now I can see my hash rates go up a little bit when adjusting the overclocking settings, but still no significant improvement. I am getting into the mid 290's, not quite hitting 300 kHash mark I would like to see as a baseline, and still hoping 320 is possible.

[cbuchner1]

Now I can see my hash rates go up a little bit when adjusting the overclocking settings, but still no significant improvement. I am getting into the mid 290's, not quite hitting 300 kHash mark I would like to see as a baseline, and still hoping 320 is possible.

I can't get near 300 kHash/s either, and worse: I get bluescreens when I dial up the overclocking.

Christian

[WattBurner]

Should I try to go back to the December version? Am I missing a prerequisite software to make this work with the latest version?

Maybe try running 4 individual cudaminer instances, each with individual -d flags from 0 to 3

You could also try passing the -m 1 flag to use a single memory block on each card.

I have no idea what is causing the cards to act so strangely.

Christian

Running T14x24 -m 1 flag, 1 GPU per instance works but with the pitiful hashrate of 17-20 kh/s.

I was getting 500 kh/s per Titan Gpu on the December cudaminer build.



 

[trell0z]

Should I try to go back to the December version? Am I missing a prerequisite software to make this work with the latest version?

Maybe try running 4 individual cudaminer instances, each with individual -d flags from 0 to 3

You could also try passing the -m 1 flag to use a single memory block on each card.

I have no idea what is causing the cards to act so strangely.

Christian

Running T14x24 -m 1 flag, 1 GPU per instance works but with the pitiful hashrate of 17-20 kh/s.

I was getting 500 kh/s per Titan Gpu on the December cudaminer build.



 

Try going downwards from 24 till you find one that gives you good khash

[WattBurner]

Should I try to go back to the December version? Am I missing a prerequisite software to make this work with the latest version?

Maybe try running 4 individual cudaminer instances, each with individual -d flags from 0 to 3

You could also try passing the -m 1 flag to use a single memory block on each card.

I have no idea what is causing the cards to act so strangely.

Christian

Running T14x24 -m 1 flag, 1 GPU per instance works but with the pitiful hashrate of 17-20 kh/s.

I was getting 500 kh/s per Titan Gpu on the December cudaminer build.



 

Try going downwards from 24 till you find one that gives you good khash

Retested 12.18 cudaminer release. Confirmed 500 kh/s per Titan running T14x32, all 4 gpus in one instance of cudaminer.

So your recommending I start at T14x24 and go decrease until a good config is found? Why is the auto config failing me? Are there any known working settings for GTX titans using the latest version of cudaminer? Should I expect a better hashrate in the latest version of Cudaminer?

[cbuchner1]

Yes the nVidia submitted kernel is expected to run 10-20% faster than Dave Andersen's kernels.

[WattBurner]

Yes the nVidia submitted kernel is expected to run 10-20% faster than Dave Andersen's kernels.

*sigh* I was almost hoping that was not the case. Now In good consonance I can't leave it be and only hash @ 500 kh/s a card.

Back to fucking this some more.

[Number6]

Now I can see my hash rates go up a little bit when adjusting the overclocking settings, but still no significant improvement. I am getting into the mid 290's, not quite hitting 300 kHash mark I would like to see as a baseline, and still hoping 320 is possible.

I can't get near 300 kHash/s either, and worse: I get bluescreens when I dial up the overclocking.

Christian

Yeah, I get constant crashing at the 135 core 700 memory in the cryptomining blog article. I had to back it down to 100 core, 400 memory to get it stable, and would like to see it run overnight before even making that call.

I have tried almost every configuration I have seen posted and about the best I can get is ~295 from each. I haven't even bothered to measure power consumption yet, but am thinking these are not going to work out quite as well as I hoped.

The Chrome trick also doesn't seem to do anything for my particular setup.

[Morgahl]

I don't have a 750 TI BUT i have noticed almost across the board that those breaking 300 kh/s are doing so at +50 Core +700 Mem. This tells me possibly two things. That the new architecture is very memory dependent (as opposed to Core dependent in Kepler and lower) OR that the memory timing in the card's bios is very relaxed and a higher clock rate is needed.

I'm actually in favor of the memory timings issue as these are initial releases and mid-line cards and the scrypt algo while not reliant on memory throughput is VERY reliant on memory latency higher Mem clocks = higher scrypt hash.

This is similar to what The Stilt did for 7xxx AMDs earlier this year.

TL;DR Use +50 Core +700 (or higher) Memory and tell me what you see. You need to favor Mem over Core to stay under the power limit.

[Morgahl]

Yes the nVidia submitted kernel is expected to run 10-20% faster than Dave Andersen's kernels.

*sigh* I was almost hoping that was not the case. Now In good consonance I can't leave it be and only hash @ 500 kh/s a card.

Back to fucking this some more.

500 is fairly good on a Titan without overclocking. I've got my solo titan at 620 with +155 Core -502 Mem, try it.

[Number6]

Also, the cards are a lot smaller than I thought, I use these frames for my Radeon rigs and I will need to reinvent something else for a frame if I can get these 750TI'sto work

I cannot use risers yet as the screw probably would not be enough to hold the card by itself.

[Number6]

I don't have a 750 TI BUT i have noticed almost across the board that those breaking 300 kh/s are doing so at +50 Core +700 Mem. This tells me possibly two things. That the new architecture is very memory dependent (as opposed to Core dependent in Kepler and lower) OR that the memory timing in the card's bios is very relaxed and a higher clock rate is needed.

I'm actually in favor of the memory timings issue as these are initial releases and mid-line cards and the scrypt algo while not reliant on memory throughput is VERY reliant on memory latency higher Mem clocks = higher scrypt hash.

This is similar to what The Stilt did for 7xxx AMDs earlier this year.

TL;DR Use +50 Core +700 (or higher) Memory and tell me what you see. You need to favor Mem over Core to stay under the power limit.

For me +50 core and +700 memory caused a black screen and driver reset message almost instantly. I backed it down to +50 core +600 memory and got the wild out of control error scrolling in cudaminer. I got in on +50, +500 now, but still back in the 285 range...

[WattBurner]

Yes the nVidia submitted kernel is expected to run 10-20% faster than Dave Andersen's kernels.

*sigh* I was almost hoping that was not the case. Now In good consonance I can't leave it be and only hash @ 500 kh/s a card.

Back to fucking this some more.

500 is fairly good on a Titan without overclocking. I've got my solo titan at 620 with +155 Core -502 Mem, try it.

Is that using the latest cudaminer? What kernal settings are you using?

[Number6]

Yes the nVidia submitted kernel is expected to run 10-20% faster than Dave Andersen's kernels.

*sigh* I was almost hoping that was not the case. Now In good consonance I can't leave it be and only hash @ 500 kh/s a card.

Back to fucking this some more.

500 is fairly good on a Titan without overclocking. I've got my solo titan at 620 with +155 Core -502 Mem, try it.

Is that using the latest cudaminer? What kernal settings are you using?

Yes, latest version cudaminer, tried both x64 and x86 versions, same within a few khash of each other.

Settings are: -m0 -l T5x24 -i 0 -H 1 currently, although I have tried different values and none seem to make much, if any, difference.

[WattBurner]

Yes the nVidia submitted kernel is expected to run 10-20% faster than Dave Andersen's kernels.

*sigh* I was almost hoping that was not the case. Now In good consonance I can't leave it be and only hash @ 500 kh/s a card.

Back to fucking this some more.

500 is fairly good on a Titan without overclocking. I've got my solo titan at 620 with +155 Core -502 Mem, try it.

Is that using the latest cudaminer? What kernal settings are you using?

Yes, latest version cudaminer, tried both x64 and x86 versions, same within a few khash of each other.

Settings are: -m0 -l T5x24 -i 0 -H 1 currently, although I have tried different values and none seem to make much, if any, difference.

Seems like T10x24 gets me about the same hashrate as the 12.18 cudaminer version. I'll play around some more and see what happens.  Trying to do this without overclocking.

[jk_14]

I have strange "configuration" problem under linux:

root@live-X9SCM:~/CudaMiner# ./configure

<cut>

checking whether we can compile XOP code... yes
checking whether we can compile AVX2 code... yes
checking for json_loads in -ljansson... no
checking for pthread_create in -lpthread... yes
checking for SSL_library_init in -lssl... no
configure: error: OpenSSL library required
root@live-X9SCM:~/CudaMiner#
root@live-X9SCM:~/CudaMiner#
root@live-X9SCM:~/CudaMiner# find / | grep libssl
/var/cache/apt/archives/libssl1.0.0_1.0.1e-3ubuntu1.1_amd64.deb
/var/lib/dpkg/info/libssl1.0.0:amd64.postrm
/var/lib/dpkg/info/libssl1.0.0:amd64.postinst
/var/lib/dpkg/info/libssl1.0.0:amd64.md5sums
/var/lib/dpkg/info/libssl1.0.0:amd64.templates
/var/lib/dpkg/info/libssl1.0.0:amd64.symbols
/var/lib/dpkg/info/libssl1.0.0:amd64.shlibs
/var/lib/dpkg/info/libssl1.0.0:amd64.list
/lib/x86_64-linux-gnu/libssl.so.1.0.0
/root/openssl-1.0.1f/libssl.a
/root/openssl-1.0.1f/libssl.pc
/usr/lib/libssl.a
/usr/lib/pkgconfig/libssl.pc
/usr/lib/x86_64-linux-gnu/libssl3.so
/usr/local/ssl/lib/pkgconfig/libssl.pc
/usr/local/ssl/lib/libssl.a
/usr/share/doc/libssl1.0.0
/usr/share/doc/libssl1.0.0/changelog.Debian.gz
/usr/share/doc/libssl1.0.0/copyright
root@live-X9SCM:~/CudaMiner#
root@live-X9SCM:~/CudaMiner#
root@live-X9SCM:~/CudaMiner# nm /usr/lib/libssl.a | grep SSL_library_init
0000000000000000 T SSL_library_init
root@live-X9SCM:~/CudaMiner#
root@live-X9SCM:~/CudaMiner#


any hints?

[Number6]

Now I am thinking a 7x 750TI rig is possible, using the  ASRock card you listed.

I also think a 600 watt PS would be enough? I calculate 420 watts (7 * 60 per card) + 100 (MB, CPU, etc), or around 520 watts load. A quality 600 watt such as http://www.newegg.com/Product/Product.aspx?Item=N82E16817139028 should be enough I hope.

The motherboard you mention http://www.newegg.com/Product/Product.aspx?Item=N82E16813157460 is about double of the famed http://www.newegg.com/Product/Product.aspx?Item=N82E16813157471, but what are you going to do? I was trying to keep the ancillary costs to around $200, but now closer to $300.

How much system memory do you think a 7x 750TI rig will need? 4 gig enough or go with 8 gig?

So I am coming up with:

$   300 - MB, CPU, Memory, PS
$1,120 - 7 x $160 GTX750ti cards
$     30 - shipping
so around $1,450 for a 7x rig capable of anywhere between 1960 kHash (280*7) to 2100 kHash (300*7) at only 550 watts draw.

Payback will be a bit longer than usual, but I think the key here is the lower overall power draw. A similar hashrate 3 x R9 290x (edit should have been 280x) system would draw nearly 820 watts.

Then comes the tweaking. If indeed these can get greater than 300 kHash each and if they can be under-volted for even less power draw, it would swing the balance even more in favor of NVidia.

Ok, based upon my original assumptions from a week ago (above) and now with two cards in hand I can recalculate and see how close those assumptions were.

System power draw, idle with two cards is drawing 45 watts from the wall. Running my two 750TI's at 290 khash and overclocking them is drawing a total of 215 watts from the wall (kill-a-watt reading).

So I have:

 215 system (2 cards hashing)
- 45 system idle
= 170  card consumption when hashing

170/2 = 85 watts per card when hashing.

So if I scale this back up compare with my original projections, I get 7 cards * 85 watts + 45 watts system = 633 watts system draw. This will require a 750 watt PS to be safe, 650 might cut it but why chance it.

My initial results with hash rates of ~290 khash per card would end up making a total of 290*7=2030 khash, lets call it 2k to be conservative.

So figuring on the side of caution I come up with 2 Mhash for 650 watts, or 3.07 khash/watt. Looking at one of my 3x R9 280 rigs, it is getting 2,160 kHash (2.16 Mhash) for 820 watts, or 2.634 khash/watt.

I think getting to a consistent 300-320 would help tip the balance, but I am on the fence right now. Maybe the newer Maxwell cards will perform a bit better, and maybe some further optimizations can be achieved with cudaminer. I think I will finish building this rig out to 6 (hopefully 7) cards, but after which I will go back to the Radeon route for the time being, unless the dualminers would come down in price.

[ivanlabrie]

Why are everyone's 780s or Titans so slow? :p

Mine can do 700kh/s without breaking a sweat...and more.  

[Coinman422]

Why are everyone's 780s or Titans so slow? :p

Mine can do 700kh/s without breaking a sweat...and more.  

what are you settings? scrypt/algo/ etc...

[djm34]

I have strange "configuration" problem under linux:

root@live-X9SCM:~/CudaMiner# ./configure

<cut>

checking whether we can compile XOP code... yes
checking whether we can compile AVX2 code... yes
checking for json_loads in -ljansson... no
checking for pthread_create in -lpthread... yes
checking for SSL_library_init in -lssl... no
configure: error: OpenSSL library required
root@live-X9SCM:~/CudaMiner#
root@live-X9SCM:~/CudaMiner#
root@live-X9SCM:~/CudaMiner# find / | grep libssl
/var/cache/apt/archives/libssl1.0.0_1.0.1e-3ubuntu1.1_amd64.deb
/var/lib/dpkg/info/libssl1.0.0:amd64.postrm
/var/lib/dpkg/info/libssl1.0.0:amd64.postinst
/var/lib/dpkg/info/libssl1.0.0:amd64.md5sums
/var/lib/dpkg/info/libssl1.0.0:amd64.templates
/var/lib/dpkg/info/libssl1.0.0:amd64.symbols
/var/lib/dpkg/info/libssl1.0.0:amd64.shlibs
/var/lib/dpkg/info/libssl1.0.0:amd64.list
/lib/x86_64-linux-gnu/libssl.so.1.0.0
/root/openssl-1.0.1f/libssl.a
/root/openssl-1.0.1f/libssl.pc
/usr/lib/libssl.a
/usr/lib/pkgconfig/libssl.pc
/usr/lib/x86_64-linux-gnu/libssl3.so
/usr/local/ssl/lib/pkgconfig/libssl.pc
/usr/local/ssl/lib/libssl.a
/usr/share/doc/libssl1.0.0
/usr/share/doc/libssl1.0.0/changelog.Debian.gz
/usr/share/doc/libssl1.0.0/copyright
root@live-X9SCM:~/CudaMiner#
root@live-X9SCM:~/CudaMiner#
root@live-X9SCM:~/CudaMiner# nm /usr/lib/libssl.a | grep SSL_library_init
0000000000000000 T SSL_library_init
root@live-X9SCM:~/CudaMiner#
root@live-X9SCM:~/CudaMiner#


any hints?

Personally I would remove entirely openssl and reinstall openssl-dev
What you could try also is to add the path to the openssl library to LD_LIBRARY_PATH
(I didn't run into those problems when I set up linux)