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

[afixane]

For anyone that have Optimus NVIDIA card and using Linux, executing ./cudaminer will return

Code:

Floating point exception (core dumped)

Here's my solution :

1. Install Bumblebee.
2. run cudaminer with optirun (yes, you can use primusrun, but it had no effect)

Code:

optirun cudaminer blablabla

[1714183754]

1714183754

[1714183754]

1714183754

[1714183754]

1714183754

[1714183754]

1714183754

[1714183754]

1714183754

[cbuchner1]

is there any way to determine where I can (safely) set my texture-cache variable?

try -l 48x5 -C 2 then

The worst that could happen is a temporary driver crash (which it should recover from in 99% of all cases)

I had an idea how to further cut shared memory use. So Kepler based cards would see full occupancy on their SMX'es, maybe gaining 10% performance on these devices.

Christian

[Aggrophobia]

new nvidia beta driver increases my khashes/s   140x2



cudaMiner 22-04 third

[cbuchner1]

new nvidia beta driver increases my khashes/s   140x2



cudaMiner 22-04 third

the cudaminer directory says 2013-04-17 though.

[zeroedout]

Thank you *very* much for this. I only have an Nvidia card and before your project mining anything was just stupid. Cgminer would give me ~28 Mh/s fot BTC but LTC or FC is was only ~2.0 KH/s :< Cudaminer gives me ~32 Kh/s!!!!! This is on a 9800GT.

tl;dr You are frekin awesome and have my gratitude! I have never been able to mine a full unit of any crypto currency but IMHO my best shot is Feather Coin. As soon as I'm able to mine some, I will be more than happy to donate a portion to you. Thanks again!!

[Cheshyr]

new nvidia beta driver increases my khashes/s   140x2



cudaMiner 22-04 third

the cudaminer directory says 2013-04-17 though.

nVidia released a new driver today.  He was talking about that, not the new version of cudaMiner

I'd like to echo the other thoughts here... this tool has made mining on nVidia cards enjoyable again.  Not epicuberdoomminer, but effective when it wasn't previously.  Thanks.

[Lacan82]

new nvidia beta driver increases my khashes/s   140x2



cudaMiner 22-04 third

the cudaminer directory says 2013-04-17 though.

nVidia released a new driver today.  He was talking about that, not the new version of cudaMiner

I'd like to echo the other thoughts here... this tool has made mining on nVidia cards enjoyable again.  Not epicuberdoomminer, but effective when it wasn't previously.  Thanks.

But the user was saying he was using cudaMINER 22-04 with new drivers, but the screenshots show it is running 2013-04-17. Unless he extracted to the same directory everytime.

Check the File Path in the menu bar

[cbuchner1]

So, lol. I touch scrypt_core_kernelA() a bit, rearrange shared memory and yay! The hash rate drops from 220kHash/s to 32kHash/s on my fastest card, a GTX 560Ti 448core edition.

Well, lol. Results are still correct though

Christian

[Lacan82]

So, lol. I touch scrypt_core_kernelA() a bit, rearrange shared memory and yay! The hash rate drops from 220kHash/s to 32kHash/s on my fastest card, a GTX 560Ti 448core edition.

Well, lol. Results are still correct though

Christian

Sounds like you're on course to program for Apple!

[coalescent]

My 680 was mining at 205 kH/s yesterday afternoon on the older, WHQL drivers. Ill be interested to see if the betas help performance at all.

@cbuchner1: if you need anything, don't hesitate to PM me (w/ email since I dont have PM privs yet).

[cbuchner1]

So I eliminate shared memory alltogether from salsa kernel A, like this (pure elegance). Note the switch to uint4.

Code:

template <int WARPS_PER_BLOCK> __global__ void
scrypt_core_kernelA(uint4 *g_idata)
{
    int warpIdx        = threadIdx.x / warpSize;
    int warpThread     = threadIdx.x % warpSize;

    uint4 *V = (uint4*)(c_V[blockIdx.x * WARPS_PER_BLOCK + warpIdx] + SCRATCH*warpThread);
    g_idata += 8 * (blockIdx.x * WU_PER_BLOCK + warpIdx * WU_PER_WARP + warpThread);

    uint4 B[4], C[4]; // registers to store an entire work unit

#define idxloop __pragma(unroll 4) for (int idx=0; idx < 4; idx++)

    idxloop { *V++ = B[idx] = *g_idata++; }
    idxloop { *V++ = C[idx] = *g_idata++; }

    for (int i = 1; i < 1024; i++) {

        xor_salsa8_uint4(B, C); xor_salsa8_uint4(C, B);

        idxloop { *V++ = B[idx]; }
        idxloop { *V++ = C[idx]; }
    }
}

And I get 108 kHash/sec on the 560 Ti 448 core.

I'm like "WTF dude?!"  I expected 250 from this. PTX is more compact than ever, global memory loads and stores are all .v4.u32 (vectorized). So what's going wrong here?

[Cheshyr]

But the user was saying he was using cudaMINER 22-04 with new drivers, but the screenshots show it is running 2013-04-17. Unless he extracted to the same directory everytime.

Check the File Path in the menu bar

I saw the file path bar.  Maybe he did extract it that way.  We have no way of knowing.

So I eliminate shared memory alltogether from salsa kernel A, like this (pure elegance). Note the switch to uint4.

Code:

template <int WARPS_PER_BLOCK> __global__ void
scrypt_core_kernelA(uint4 *g_idata)
{
    int warpIdx        = threadIdx.x / warpSize;
    int warpThread     = threadIdx.x % warpSize;

    uint4 *V = (uint4*)(c_V[blockIdx.x * WARPS_PER_BLOCK + warpIdx] + SCRATCH*warpThread);
    g_idata += 8 * (blockIdx.x * WU_PER_BLOCK + warpIdx * WU_PER_WARP + warpThread);

    uint4 B[4], C[4]; // registers to store an entire work unit

#define idxloop __pragma(unroll 4) for (int idx=0; idx < 4; idx++)

    idxloop { *V++ = B[idx] = *g_idata++; }
    idxloop { *V++ = C[idx] = *g_idata++; }

    for (int i = 1; i < 1024; i++) {

        xor_salsa8_uint4(B, C); xor_salsa8_uint4(C, B);

        idxloop { *V++ = B[idx]; }
        idxloop { *V++ = C[idx]; }
    }
}

And I get 108 kHash/sec on the 560 Ti 448 core.

I'm like "WTF dude?!"  I expected 250 from this.

Is this like the old sorting memory allocation problem?  Next Fit vs First Fit or whatever?  Sorting is expensive, so forcing organization is actually costing performance?

[cbuchner1]

My bet is that this new code now violates the memory coalescing rules for CUDA devices, resulting in low throughput-

Back to the drawing board... Let's see if I can keep the elegance while maintaining memory coalescing.

[falconae]

Thank you *very* much for this. I only have an Nvidia card and before your project mining anything was just stupid. Cgminer would give me ~28 Mh/s fot BTC but LTC or FC is was only ~2.0 KH/s :< Cudaminer gives me ~32 Kh/s!!!!! This is on a 9800GT.

tl;dr You are frekin awesome and have my gratitude! I have never been able to mine a full unit of any crypto currency but IMHO my best shot is Feather Coin. As soon as I'm able to mine some, I will be more than happy to donate a portion to you. Thanks again!!

Can you post your settings? I'm only pulling 15Kh that was up from 2 but still, seeing you 32 makes me curious

[cbuchner1]

Can you post your settings? I'm only pulling 15Kh that was up from 2 but still, seeing you 32 makes me curious

on these old cards it makes a difference whether you run Windows XP, Linux --- or Windows Vista/7/8. It's not a settings issue - I believe the WDDM driver model has issues with these old cards.

[Aggrophobia]

But the user was saying he was using cudaMINER 22-04 with new drivers, but the screenshots show it is running 2013-04-17. Unless he extracted to the same directory everytime.

Check the File Path in the menu bar

This! Extracted in the same directory, thats why i've edited my post

[FalconFour]

FYI: oughtta post the "put your results here" GDocs link into the first post so it's easy to find... I'm off to go look for it, since I just got access to a GTX 660 to play around with on cudaMiner. I'm installing Win7 on that test machine

edit: yeah, hidden on page 14, and I'm not even about to try following the discussion around to that post where I found the "sorted" one...

[Lacan82]

FYI: oughtta post the "put your results here" GDocs link into the first post so it's easy to find... I'm off to go look for it, since I just got access to a GTX 660 to play around with on cudaMiner. I'm installing Win7 on that test machine

edit: yeah, hidden on page 14, and I'm not even about to try following the discussion around to that post where I found the "sorted" one...

I fixed the sorted sheet. sort2 

[FalconFour]

FYI: oughtta post the "put your results here" GDocs link into the first post so it's easy to find... I'm off to go look for it, since I just got access to a GTX 660 to play around with on cudaMiner. I'm installing Win7 on that test machine

edit: yeah, hidden on page 14, and I'm not even about to try following the discussion around to that post where I found the "sorted" one...

I fixed the sorted sheet. sort2 

Right, but the problem is trying to find the link to these docs in this complete clusterfuck of a 27-page thread. Forum layout isn't always conducive to information archiving and retrieval... posts between the first two and last two pages are pretty much lost in history. :/

[gork68]

I have MSI 660 ti pe running stock.

I am using the 4/22 build and the following flags < -i 0 -C 2 -l 42x7 >

           *** CudaMiner for nVidia GPUs by Christian Buchner ***
                     This is version 2013-04-22 (alpha)
        based on pooler-cpuminer 2.2.3 (c) 2010 Jeff Garzik, 2012 pooler
               Cuda additions Copyright 2013 Christian Buchner
           My donation address: LKS1WDKGED647msBQfLBHV3Ls8sveGncnm

[2013-04-24 21:56:45] 1 miner threads started, using 'scrypt' algorithm.
[2013-04-24 21:56:45] Long-polling activated for http://127.0.0.1:8332/lp
[2013-04-24 21:56:45] GPU #0: GeForce GTX 660 Ti with compute capability 3.0
[2013-04-24 21:56:45] GPU #0: interactive: 0, tex-cache: 2D, single-alloc: 1
[2013-04-24 21:56:45] GPU #0: using launch configuration  42x7
[2013-04-24 21:56:45] GPU #0: GeForce GTX 660 Ti, 9408 hashes, 67.01 khash/s
[2013-04-24 21:56:48] GPU #0: GeForce GTX 660 Ti, 573888 hashes, 176.02 khash/s
[2013-04-24 21:56:48] accepted: 1/1 (100.00%), 176.02 khash/s (yay!!!)
[2013-04-24 21:57:03] GPU #0: GeForce GTX 660 Ti, 2671872 hashes, 179.91 khash/s

[2013-04-24 21:57:04] accepted: 2/2 (100.00%), 179.91 khash/s (yay!!!)
[2013-04-24 21:57:11] GPU #0: GeForce GTX 660 Ti, 1430016 hashes, 179.74 khash/s

[2013-04-24 21:57:12] accepted: 3/3 (100.00%), 179.74 khash/s (yay!!!)
[2013-04-24 21:57:45] GPU #0: GeForce GTX 660 Ti, 6115200 hashes, 180.31 khash/s

[2013-04-24 21:58:03] GPU #0: GeForce GTX 660 Ti, 3226944 hashes, 180.03 khash/s

[2013-04-24 21:58:03] accepted: 4/4 (100.00%), 180.03 khash/s (yay!!!)


After it stabalizes I maintain +180 KH/s although the pool stats show higher for some reason...