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

[SubNoize]

Can someone explain this "64x2" "S27x3,28x4" thing to me or point me in the right direction on reading up on it??

I have a 580gtx and i'm trying to figure out the best set up

[1713999073]

1713999073

[SubNoize]

also gettings lots of crashes after like 20 minutes of use.

[hammz]

I stopped using it.  Can't connect to the pools I want to, and kattare connects but keeps giving massive stales.

back to CGminer for me.

[cancrena]

Can someone explain this "64x2" "S27x3,28x4" thing to me or point me in the right direction on reading up on it??

I have a 580gtx and i'm trying to figure out the best set up

well i did not figure out the meaning, but if you run it and let it autotune, it will choose automatically what's the best one (then you can add the flag in the batch file, like -l 112x2 for me)

[dabster]

This is what I am getting with my gtx 680

[2013-04-14 03:56:23] GPU #0:  186.17 khash/s with configuration  80x2
[2013-04-14 03:56:23] GPU #0: using launch configuration  80x2

[2013-04-14 04:06:23] GPU #0: GeForce GTX 680, 1448960 hashes, 181.06 khash/s
[2013-04-14 04:06:23] accepted: 8/11 (72.73%), 181.06 khash/s (yay!!!)
[2013-04-14 04:07:03] GPU #0: GeForce GTX 680, 7424000 hashes, 181.87 khash/s
[2013-04-14 04:07:32] DEBUG: got new work in 131 ms
[2013-04-14 04:07:35] DEBUG: hash <= target


Not bad.

[Notanon]

Trying out the 10/4 release, it's working better than the initial release and am getting about 7.9kh/s on my GT 220M, which is 33% better than using Scryptminer. Wasn't expecting miracles given how old the card in my laptop is, but it is a nice boost all the same.

Edit: and the 14/4 release is giving me 8.54kh/s.

[cbuchner1]

also gettings lots of crashes after like 20 minutes of use.

I've seen one crash after letting it run through the night. This is actually the first time I've had cudaminer seen crashing during regular operation.
Weirds me out.

[cbuchner1]

Can someone explain this "64x2" "S27x3,28x4" thing to me or point me in the right direction on reading up on it??

I have a 580gtx and i'm trying to figure out the best set up

well i did not figure out the meaning, but if you run it and let it autotune, it will choose automatically what's the best one (then you can add the flag in the batch file, like -l 112x2 for me)

112x2 means it throws 112 blocks at CUDA, and each consists of 2 warps. A warp is a group of 32 threads.

So in total it computes 112*2*32 = 7168 hashes in parallel in a single CUDA kernel launch.

And because the scrypt scratchpad is 131072 bytes long, this would consume 7168*131072 bytes of memory
on the card. That's about 917 MB.

[cbuchner1]

Has anyone successfully to connected to the weminteltc.com pool?

I also tried just http://eu.wemineltc.com:3333

I downloaded their stratum proxy executable, it connects, but only solo mines packets.

Not sure what I need to populate for the connection address at this point.

They've just updated their windows stratum binary to default to scrypt. Previously you had to explicitly instruct it to use scrypt for proof of work.
Also note that it opens local port 8332

It's kind of cool to have a stratum proxy that requires zero setup. Just click the EXE and you're good to point cudaminer at it.

So you need to direct cudaminer to -o http://localhost:8332 -O yourworkername.1:yourpassword

If you used port 9332 and your litecoin app ran in -server mode, then yes, you might have been solo mining.

[cbuchner1]

Trust me, you want to compile this for 32 bit.

Compilation instruction for Fedora:
http://www.patrickmin.com/linux/tip.php?name=bitcoin_litecoin

On ubuntu and debian, you need the
g++-multilib and ia32-lib packages installed and possibly the libcurl4-dev:i386 package as well

Christian

[SubNoize]

Can someone explain this "64x2" "S27x3,28x4" thing to me or point me in the right direction on reading up on it??

I have a 580gtx and i'm trying to figure out the best set up

well i did not figure out the meaning, but if you run it and let it autotune, it will choose automatically what's the best one (then you can add the flag in the batch file, like -l 112x2 for me)

112x2 means it throws 112 blocks at CUDA, and each consists of 2 warps. A warp is a group of 32 threads.

So in total it computes 112*2*32 = 7168 hashes in parallel in a single CUDA kernel launch.

And because the scrypt scratchpad is 131072 bytes long, this would consume 7168*131072 bytes of memory
on the card. That's about 917 MB.

cheers! are there guides for these things? or is it just trial and error? Sorry i don't want to annoy you too much but do you know what would be some settings to throw at a 580gtx to see any improvement or should i just let the auto tune take care of it and not worry?

Also is there any information i can give you when it crashes?

[to3m]

also gettings lots of crashes after like 20 minutes of use.

I've seen one crash after letting it run through the night. This is actually the first time I've had cudaminer seen crashing during regular operation.
Weirds me out.

I got a std::bad_alloc from uint32x4_t::operator new after leaving it for about 8-9 hours. The process had used 1.7GB so I assume there's a memory leak somewhere... I think one problem is that in scanhash_scrypt, the 2 Xx4 buffers are allowed to leak.

--Tom

[cbuchner1]

Why can't it be 64-bit? I can only chroot to gentoo-prefix (which does not support multilib) to compile cudaminer.

Mostly due to lack of testing opportunity (at the moment),  and different priorities of the developer.

You could try disabling the "tex" kernels (comment them out, don't call them) to see if it builds in 64bits.
The texture caching feature was a failure, anyway.

But there's no guarantee that it will compute correctly in 64 bits. At least the CPU will validate this now.

[cbuchner1]

I got a std::bad_alloc from uint32x4_t::operator new after leaving it for about 8-9 hours. The process had used 1.7GB so I assume there's a memory leak somewhere... I think one problem is that in scanhash_scrypt, the 2 Xx4 buffers are allowed to leak.

That explains my crash that I saw in the morning. I will be fixing that pronto.

UPDATE!  the leak fix version is posted

Some latest performance stats:
GTX 460  -l28x4   100kHash/s
GTX 560Ti -l32x4  144kHash/s

Now if I tear out the power hungry GTX 260 and put in a GT 640, I should be able to hit close to 300kHash and NOT blow out my PSU.

[Notanon]

Just switched to mining Mincoins on the vircurpool pool and the hash rate on the GT 220M is now averaging around 9.8kh/s. interesting that it's a bit higher compared to the Ozcoin Litecoin pool.

[cbuchner1]

cheers! are there guides for these things? or is it just trial and error? Sorry i don't want to annoy you too much but do you know what would be some settings to throw at a 580gtx to see any improvement or should i just let the auto tune take care of it and not worry?

Trust the autotune. Try tuning a couple of times, note down the configs and speeds that it found - and then pick the one that consistently puts out the highest kHash.

This is how I just found my 32x4 and 28x4 configurations for my GTX 560Ti and GTX 460 cards.

The thing is, with every cudaminer update, these ideal configurations can change. Even overclocking your card can move the ideal configuration to some other "sweet spot".

Christian

[nst6563]

I definately like the speed improvement on the last update.  I went from a total hashrate of 230kh/s (2 gtx560se's and a gt430) to 264kh/s.

Unfortunately the hashrate of my gtx260 went down by 10kh/s.  I plan on replacing it soon anyway.

[cbuchner1]

Unfortunately the hashrate of my gtx260 went down by 10kh/s.  I plan on replacing it soon anyway.

I will see what I can do about restoring the performance of the "S" kernels... Also I am trying to fix the driver crash on CTRL-C

Old vs. New.

GT 640 4GB 128bit GDDR3 memory
vs.
GTX 260 896MB 448-bit memory

Both do about same kHash/s, but difference in power draw is factor 3 or 4.
My PC is now doing 280 kHash/sec on 3 GPUs.

[cbuchner1]

Hey, the CTRL-C issue has been fixed. Grab the update from the first posting.

Don't directly close the DOS box with the X button though, that doesn't leave my program enough time to do cleanup.

The CPU utilization is not really predictable. Sometimes it's just 10%, other times  it goes up to 40%. Weird.

[TiWu]

GTX 570 Club3D
Win 8 x64 - 314.22 WHQL driver

• 2013-04-09: 120x2 150 kh/s
• 2013-04-10: 30x7  175 kh/s
• 2013-04-14: 30x7  210 kh/s (update 2, crashes on CTRL-C)
• 2013-04-14: 30x8  220 kh/s (update 3)

Autotune for latest (update 3) version, no debug:

Code:

[2013-04-14 18:57:01] GPU #0: GeForce GTX 570 with compute capability 2.0
[2013-04-14 18:57:01] GPU #0: interactive: 0, tex-cache: 0, single-alloc: 0
[2013-04-14 18:57:01] GPU #0: Performing auto-tuning (Patience...)
[2013-04-14 18:58:26] GPU #0:  222.38 khash/s with configuration  30x8
[2013-04-14 18:58:26] GPU #0: using launch configuration  30x8
[2013-04-14 18:58:26] GPU #0: GeForce GTX 570, 7680 hashes, 0.09 khash/s
[2013-04-14 18:58:27] GPU #0: GeForce GTX 570, 7680 hashes, 119.87 khash/s
[2013-04-14 18:58:37] GPU #0: GeForce GTX 570, 2242560 hashes, 221.51 khash/s
[2013-04-14 18:58:37] accepted: 1/1 (100.00%), 221.51 khash/s (yay!!!)

Shutdown issue is handled nicely too.

Code:

[2013-04-14 19:00:37] accepted: 13/14 (92.86%), 217.86 khash/s (yay!!!)
Ctrl-C
[2013-04-14 19:00:38] workio thread dead, waiting for workers...
[2013-04-14 19:00:38] worker threads all shut down
Terminate batch job (Y/N)?