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

[djm34]

I just tried on one pool Blakecoin EU1 pool, I get only boo's.
the hashrate would be max 130Mhash, I just checked with the big amd electron's sucker (which doesn't seem to use that much on it),
 ouch 3.25GHash/s

this is the code from the blakecoin CPU miner... so the line on the bottom applies to Blake256.
I fixed this in cudaminer now.

Code:

if (opt_algo == ALGO_SCRYPT)
diff_to_target(work->target, sctx->job.diff / 65536.0);
else
diff_to_target(work->target, sctx->job.diff);

About the hashing speed: This is a straight port of the CPU code running mostly unoptimized in CUDA.  Bad speed is kind of
expected, although 50 MHash/s is a start...

If hashes are still getting rejected now, there must be another reason (maybe endianness problems....)

Christian

ok, I have recompiled, this time, I don't get that many share. The only one I got (in 3 runs) was rejected.

I recompile again, just to make sure everything was ok at that level, but it really take a long take to get a share...

[cbuchner1]

interesting: a CPU miner with all the algorithms the author could find: https://github.com/KaSt/AllMiner

This could serve as a reference for future cudaminer improvements, maybe.

Christian

[cbuchner1]

ok, I have recompiled, this time, I don't get that many share. The only one I got (in 3 runs) was rejected.

I recompile again, just to make sure everything was ok at that level, but it really take a long take to get a share...

ok, I guess I will have to start some pool mining myself and do some more validation.  but maybe not tonight

it's bed time.

Christian

[blade87]

Will the CUDA 6 release have any impact on mining?

so far it's a negative impact. So CUDA 6 RC is a no-go. There are serious
performance regressions for some kernels, and no additional gain for code
compiled specifically for the 750 Ti (Maxwell).

Christian

Isn't CUDA 6 about unified memory, ie host and device can share literally the same variable in the code? If they do, then there would be constant memcpy going on, which would slow down anything pre-Maxwell as Maxwell is the first to have hardware support for that.

Though I could be way off. Been a while since I've used CUDA.

[patoberli]

My Gainward 750 Ti arrived today. My mining rig has now GPUs from 3 different vendors and it is working under Win 8 x64
I have the onboard Intel 4600 (with connected display), a AMD Radeon 6950 1 GB for mining (395 kh/s stock clock, scrypt) and the new NVIDIA 750 Ti (290 kh/s stock clock, scrypt).
Driver installation was straight forward. Only downside to not having a display connected to either card, I can't use CCC or Nvidias control panel. If I would want to overclock I have to first connect a display.

[djm34]

My Gainward 750 Ti arrived today. My mining rig has now GPUs from 3 different vendors and it is working under Win 8 x64
I have the onboard Intel 4600 (with connected display), a AMD Radeon 6950 1 GB for mining (395 kh/s stock clock, scrypt) and the new NVIDIA 750 Ti (290 kh/s stock clock, scrypt).
Driver installation was straight forward. Only downside to not having a display connected to either card, I can't use CCC or Nvidias control panel. If I would want to overclock I have to first connect a display.

You can do it with either MSI AFterburner, EVGA Precision and Nvidia Inspector (but yes the nvidia control panel does not open).
I am thinking also about adding a 750ti (my psu should still have 50watt to spare even with the two big cards)

[djm34]

I discover in my bios, that it was possible the change the latency of the pci-e.
It is set by default at 32 something, but it can be increased.
Would there be an interest in increasing (or reducing) the latency when using risers or several pci cards ?

[Morgahl]

My Gainward 750 Ti arrived today. My mining rig has now GPUs from 3 different vendors and it is working under Win 8 x64
I have the onboard Intel 4600 (with connected display), a AMD Radeon 6950 1 GB for mining (395 kh/s stock clock, scrypt) and the new NVIDIA 750 Ti (290 kh/s stock clock, scrypt).
Driver installation was straight forward. Only downside to not having a display connected to either card, I can't use CCC or Nvidias control panel. If I would want to overclock I have to first connect a display.

You can do it with either MSI AFterburner, EVGA Precision and Nvidia Inspector (but yes the nvidia control panel does not open).
I am thinking also about adding a 750ti (my psu should still have 50watt to spare even with the two big cards)

Have you tried setting up a dummy plug for this? You just need a DVI-VGA adapter and 3x50-150 ohm resistors to create one. Instructions are literally all over the interwebz.

[Cial]

Been away from the thread for a while. What's the average hashrate for scrypt on a 750ti atm? I'm getting ~300 on mine.

[rahrahrah]

Been away from the thread for a while. What's the average hashrate for scrypt on a 750ti atm? I'm getting ~300 on mine.

Same.

[Morgahl]

Been away from the thread for a while. What's the average hashrate for scrypt on a 750ti atm? I'm getting ~300 on mine.

Same.

300 is pretty good, Some are seeing as high as 340 with heavy OC and as low as 280.

[Omega0255]

I discover in my bios, that it was possible the change the latency of the pci-e.
It is set by default at 32 something, but it can be increased.
Would there be an interest in increasing (or reducing) the latency when using risers or several pci cards ?

probably not unless youre running 1x

[lumberinvestments]

You can get a 4x multiplier card here:

http://www.aliexpress.com/item/PCI-e-PCIe-Express1X-to-4port-1X-multiplier-switch-riser-card-adapter-cable-for-diy-bitcoin/1444567810.html

The only manufacturer that does that. Probably the guy from GPU coin is having some kind of deal with this manufacturer to get 8x multiplier.
I think that 4x is enough, it will save you from investing in a lot of computers.

As of now it is the single manufacturer in China that offers such things.

Good luck in your mining adventures !

[ivanlabrie]

You can get a 4x multiplier card here:

http://www.aliexpress.com/item/PCI-e-PCIe-Express1X-to-4port-1X-multiplier-switch-riser-card-adapter-cable-for-diy-bitcoin/1444567810.html

The only manufacturer that does that. Probably the guy from GPU coin is having some kind of deal with this manufacturer to get 8x multiplier.
I think that 4x is enough, it will save you from investing in a lot of computers.

As of now it is the single manufacturer in China that offers such things.

Good luck in your mining adventures !

Yeah, really expensive...if that guy can get em for half the price, it might be worth it.

Nvidia driver supports up to 16 gpus per rig AFAIK.

[djm34]

but how much of them can support the cpu itself ?

[bigjme]

but how much of them can support the cpu itself ?

well using -H 2 the cpu load should be tiny, at 10% per card you could manage 9

[djm34]

but how much of them can support the cpu itself ?

well using -H 2 the cpu load should be tiny, at 10% per card you could manage 9

I was not really thinking in terms of cpu load but rather of an intrinsic limitation (like the number of sata 3 port)

[ivanlabrie]

but how much of them can support the cpu itself ?

well using -H 2 the cpu load should be tiny, at 10% per card you could manage 9

I was not really thinking in terms of cpu load but rather of an intrinsic limitation (like the number of sata 3 port)

I'm not sure, depends on the driver and board bios as well as the electrical connecions on each slot.

[coercion]

I can't seem to compile the latest commit for blake and I'm not familiar with how to handle nvcc errors like the following:

Code:

nvcc -g -O2 -Xptxas "-abi=no -v" -arch=compute_10 --maxrregcount=64 --ptxas-options=-v  -o blake.o -c blake.cu

### Assertion failure at line 1917 of ../../be/cg/cgemit.cxx:
### Compiler Error in file /tmp/tmpxft_000010b1_00000000-9_blake.cpp3.i during Assembly phase:
### operand 0 is not a register
nvopencc INTERNAL ERROR: /usr/lib/nvidia-cuda-toolkit/lib//be returned non-zero status 1

[cbuchner1]

I can't seem to compile the latest commit for blake and I'm not familiar with how to handle nvcc errors like the following:

Code:

nvcc -g -O2 -Xptxas "-abi=no -v" -arch=compute_10 --maxrregcount=64 --ptxas-options=-v  -o blake.o -c blake.cu

### Assertion failure at line 1917 of ../../be/cg/cgemit.cxx:
### Compiler Error in file /tmp/tmpxft_000010b1_00000000-9_blake.cpp3.i during Assembly phase:
### operand 0 is not a register
nvopencc INTERNAL ERROR: /usr/lib/nvidia-cuda-toolkit/lib//be returned non-zero status 1

you might want to get the latest tagged release instead.

by the way, compilation on windows seemed to work fine for blake.cu

EDIT: I can repro the problem now under Linux. I think I am going to replace the entire blake implementation anyway, as yesterday's approach with some simple and clean blake256 implementation didn't produce compatible hashes. So it's likely the problem will go away as a result.

Christian