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

[falconae]

Thank you for all your work on this,
I'm running an older card, non OC 9800GT it's getting 15.6KHash where before it was only seeing 3.5
wow!

I've just been testing it out solo (yes I know that speed is better suited for pool) and have been switching to each new version.  When you do that, does the work continue from the last point you were at? or does it start fresh each time and the previous work is lost?

when you get it perfected, what are the chances of moving it over to bitcoin mining?

[cbuchner1]

I've just been testing it out solo (yes I know that speed is better suited for pool) and have been switching to each new version.  When you do that, does the work continue from the last point you were at? or does it start fresh each time and the previous work is lost?

when you get it perfected, what are the chances of moving it over to bitcoin mining?

I am puzzled that your 9800 GT delivers "only" so little hashing power. It has 112 CUDA cores after all! My 9600M GT in the laptop does 9.6 kHash/s on Linux with just 32 CUDA cores.

It doesn't really matter from where your miner continues. Hashing is a like lottery game. Each hashing attempt has the same chance of being below the difficulty target. On average every 2.5 minutes we have a new block on the litecoin network, and hence all work begins from scratch. Even if a new transaction is added to the current block in your "litecoind" program, then hashing has to begin from scratch.

All the work you can attempt is to to try 2^32 different hashes - 4294967296 exactly (the "nonce" field has 32 bits). Also you can vary the timestamp field in the block ever so slightly, which gives you a few some more combinations to try. This is important for ASIC miners on Bitcoin. They are so fast that they have to vary more than just the nonce.

If you solo mine at this low hash rate, you might get a block (50 LTC + transaction fees) once per year or so. Bad idea, unless you get lucky and find a block in a few days already (hello, it's a lottery!)

Pools set much easier difficulties, so they can "count" the work sent in by clients. This is then called shares, and used to compute payouts. wemineltc.com have just started to roll out variable difficulty (VARDIFF) which adjusts to the individual client hashing powers. So the number of shares reported per hour stays constant, but for payout the shares are weighed according to their difficulty.

Christian

[falconae]

Thank you for the in depth answer 
Any thoughts to get it faster? the miner reports a compute capability of 1.1 (is that correct for this card?)
and I've tried without autotune and it gets between 7 and 9 KHash

thanks again

[danyy]

i think nvidia is still worthless, with a $90 hd 5570 u can get 80 khs, with a 6670 u get 200 khs.. and with that nvidia $300 dll u only get 120 lol...

with a gt440 i was getting 20 khs, how many khs do you think im gonna get now with this app?

[Schleicher]

I am puzzled that your 9800 GT delivers "only" so little hashing power. It has 112 CUDA cores after all! My 9600M GT in the laptop does 9.6 kHash/s on Linux with just 32 CUDA cores.

9800 GT has Compute Capability 1.0
9600M GT has Compute Capability 1.1
https://developer.nvidia.com/cuda-gpus
Maybe that's the reason

[cbuchner1]

i think nvidia is still worthless

I am only trying to make nVidia suck a bit less.

a GT 440 could make 30-40 kHash/s.

Let me tell you that $300 cards now make around 150-180 kHash with my tool.

Christian

[cbuchner1]

9800 GT has Compute Capability 1.0
9600M GT has Compute Capability 1.1

Capability 1.1 only added atomic operations in global memory - everything else stayed pretty much the same.

As far as I know the capability 1.0 only existed on the original G80 die, e.g. the 8800GTX, the 8800GT.

I think the 9800 series was based on the G92 die (an architectural re-spin of G80). so it is likely it has Compute 1.1.

And then also they added the GT200 die which added double precision support and was also sold as the first Tesla cards for professional use. That's what my GTX 260 cards were based off. But now they've reached end of life for me

Christian

[cbuchner1]

Thank you for the in depth answer  
Any thoughts to get it faster? the miner reports a compute capability of 1.1 (is that correct for this card?)
and I've tried without autotune and it gets between 7 and 9 KHash

you might want to try -l S14x3  or -l S28x3. It's an exact multiple of your multiprocessor count, and running 3 warps per multiprocessor.

Christian

[danyy]

you were right, my gt440 went from 17-20 khs in the past to 33 khs now.. more than 50%, good tool
the only thing is that i get a lot of stales/invalids, why

[cbuchner1]

If you see a lot of stales, try setting up a stratum proxy. e.g. wemineltc.com has a binary for download that runs with zero configuration.

Then you point cudaminer at -o http://localhost:8332 -O yourworkername.1:yourpassword

and you are mining with almost no stales.

[falconae]

you might want to try -l S14x3  or -l S28x3. It's an exact multiple of your multiprocessor count, and running 3 warps per multiprocessor.

Christian

no change with either, oh well, just a crappy card

[peacefulmind]

I am running the two titans in surround mode hooked up to a 5760x1200 display.

Could this configuration prevent cudaMiner from working?  I have tried multiple versions over the past few days and seem to get only a driver crash (314.22) or nothing.

I am not new to mining and am pointing it to localhost stratum proxy that is properly configured.

- Do I need to disable 5760x1200 for this to work?

[danyy]

i got stales only the first time i connected to the pool, now its fine with no stales and at good speed thanks

[cbuchner1]

I am running the two titans in surround mode hooked up to a 5760x1200 display.

try single memory allocation mode -m 1,1

Alternatively run just one card per cudaminer, i.e. -d 0  and -d 1 on separate cudaminer instances

[falconae]

Here's my results after connecting to WeMine
better than a poke in the eye with a sharp stick 
(and they are showing up properly on WeMine)
https://i.imgur.com/AH9Y0hA.jpg

[haderach]

I am running the two titans in surround mode hooked up to a 5760x1200 display.

Could this configuration prevent cudaMiner from working?  I have tried multiple versions over the past few days and seem to get only a driver crash (314.22) or nothing.

I am not new to mining and am pointing it to localhost stratum proxy that is properly configured.

- Do I need to disable 5760x1200 for this to work?

I only have the one titan, but I've been having similar issues. First run of cudaminer (4/14) and I get a stream of messages saying the CPU could not verify the block. Second run it seems to work, but because of the crash I can't get the cock speed to return to normal levels and it runs at 512Mhz (should be boosted to 1167MHz).

I tried going back to 4/13 but it seems to have issues as well. Prior to this I had only started and used 4/13 once for about a continuous day of hashing with no issues. Now it has been running for half a day and had 0 shares where as I had 9+ in the same period before. I also get the CPU not verifying the hash fairly often, about once a minute yet I can run the card with games or stressing benchmarks for hours and never see an issue.

Edit: The latest release with the Ctrl-C fix seems to have fixed the problem with the core being stuck at 512MHz. Now to see if I get any shares at a reasonable pace.

[E]

141kH/s on GTX-470; I left it (the 4-14 build) running overnight and came back to:

[2013-04-14 13:50:30] HTTP request failed: Failed to connect to 198.100.146.95: No buffer space

and cudaminer.exe had terminated.

Thanks for your work on this!

[Lacan82]

GTX 570 is now getting around 240.

650M is now getting around 40

[gchil0]

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

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

A 32-bit compile on a 64-bit host is possible but very tricky thanks mainly to libcurl.  Here's a link to a 32-bit Linux binary from the April 14 code.  Give it a try.  Hopefully you won't run into shared library issues...
https://www.dropbox.com/s/twjdug4l0z4rnz0/cudaminer_414.gz

[haderach]

Following the computation on total memory used, what is the reason that not even half of the memory available on the Titan is being used? Up until the last release today (4/14) autotune has always picked a configuration of 300-307x2 (now 263x2) which comes out to about 2.5GB of used memory at most. Since scrypt is a space/computation tradeoff algorithm isn't there potential for almost a doubling in hash performance if the full 6GB could be used?

Btw, the lastest work has been fantastic. My card hashes at a much more consistent rate and about 30khash/s more than previous builds. Fixing the Ctrl-C is just a nice usability improvement.