mfaktc: a CUDA program for Mersenne prefactoring

[storm5510]

The below is from a newer version than I have been using. I was trying to run the self-test using the -st parameter.

mfaktc v0.21 (64bit built)

CUDA version info
  binary compiled for CUDA  11.20
  CUDA runtime version      11.20
  CUDA driver version       11.70

CUDA device info
  name                      NVIDIA GeForce RTX 2080
  compute capability        7.5
  max threads per block     1024
  max shared memory per MP  65536 byte
  number of multiprocessors 46
  clock rate (CUDA cores)   1830MHz
  memory clock rate:        7000MHz
  memory bus width:         256 bit

Automatic parameters
  threads per grid          753664
  GPUSievePrimes (adjusted) 82486
  GPUsieve minimum exponent 1055144

########## testcase 1/2867 ##########
Starting trial factoring M50804297 from 2^67 to 2^68 (0.59 GHz-days)
Using GPU kernel "75bit_mul32_gs"
  Date   Time |   Cls    Pct |   Time     ETA | GHz-d/day |                 Bits 
Nov 05  13:17 |  3387   0.1% |  0.001    n.a. |      n.a. |    50804297    67-68  
ERROR: cudaGetLastError() returned 209: no kernel image is available for execution on the device

It requires cudart64_110.dll to run which I found. I have seen this before, but I do not remember the solution.

[AlvinBunk]

Originally Posted by kriesel

CUDA SDK 8 is required for GTX10xx. You need higher for RTX20xx, higher still for RTX30xx. (Assuming PTX is not there.) Read the last link I posted. SDK level != CC level.

So I guess I'm confused. In the previous post you mentioned that there is no need to build and there are downloads available at this link: https://download.mersenne.ca/mfaktc/mfaktc-0.21
But when I check there, how do I know which one is a Windows build with "Compute Compatibility" 8.6. I can see my RTCX 3060 is defintely CC 8.6.

Maybe there are CC 8.x builds somewhere else? Couldn't find them though...

[AlvinBunk]

Originally Posted by storm5510

The below is from a newer version than I have been using. I was trying to run the self-test using the -st parameter.

mfaktc v0.21 (64bit built)

CUDA version info
  binary compiled for CUDA  11.20
  CUDA runtime version      11.20
  CUDA driver version       11.70

CUDA device info
  name                      NVIDIA GeForce RTX 2080
  compute capability        7.5
  max threads per block     1024
  max shared memory per MP  65536 byte
  number of multiprocessors 46
  clock rate (CUDA cores)   1830MHz
  memory clock rate:        7000MHz
  memory bus width:         256 bit

Automatic parameters
  threads per grid          753664
  GPUSievePrimes (adjusted) 82486
  GPUsieve minimum exponent 1055144

########## testcase 1/2867 ##########
Starting trial factoring M50804297 from 2^67 to 2^68 (0.59 GHz-days)
Using GPU kernel "75bit_mul32_gs"
  Date   Time |   Cls    Pct |   Time     ETA | GHz-d/day |                 Bits 
Nov 05  13:17 |  3387   0.1% |  0.001    n.a. |      n.a. |    50804297    67-68  
ERROR: cudaGetLastError() returned 209: no kernel image is available for execution on the device

It requires cudart64_110.dll to run which I found. I have seen this before, but I do not remember the solution.

Have you tried using this version: mfaktc-0.21.win.cuda100
I have an RTX 2060 and the same compute compability version and it works fine. Not sure if you are using a Windows build though?

[storm5510]

Originally Posted by AlvinBunk

Have you tried using this version: mfaktc-0.21.win.cuda100
I have an RTX 2060 and the same compute compability version and it works fine. Not sure if you are using a Windows build though?

Yes. This is the one I have been using. A Windows build. I suspect I may have a PSU issue. I have to use Afterburner to throttle the GPU down. 70% power is high as I dare go. If I try going above, I often end up with a blank screen and the GPU fans running full throttle. It takes a reboot to correct the problem.

[AlvinBunk]

Originally Posted by storm5510

Yes. This is the one I have been using. A Windows build. I suspect I may have a PSU issue. I have to use Afterburner to throttle the GPU down. 70% power is high as I dare go. If I try going above, I often end up with a blank screen and the GPU fans running full throttle. It takes a reboot to correct the problem.

Are you sure?
It should be showing "binary compiled for CUDA 10.0" if you are using the one I mentioned. But your screenshot shows version 11.

[storm5510]

Originally Posted by AlvinBunk

Are you sure?
It should be showing "binary compiled for CUDA 10.0" if you are using the one I mentioned. But your screenshot shows version 11.

Yes, I am sure. The below is from the working build which I have been using for quite a long time.

CUDA version info
  binary compiled for CUDA  10.0
  CUDA runtime version      10.0
  CUDA driver version       11.70

I am trying to track down an issue where the screen goes blank, and the GPU fans go full throttle. The only way to prevent this is to use Afterburner to cut the power back to 70%, or lower.

[lycorn]

Originally Posted by storm5510

Yes, I am sure. The below is from the working build which I have been using for quite a long time.

CUDA version info
  binary compiled for CUDA  10.0
  CUDA runtime version      10.0
  CUDA driver version       11.70

I am trying to track down an issue where the screen goes blank, and the GPU fans go full throttle. The only way to prevent this is to use Afterburner to cut the power back to 70%, or lower.

I think you´re addressing two different issues in your last couple of posts. The latter appears to boil down to insufficient cooling of the GPU - have you tried cleaning the fans? Dust can accumulate to a point where their operation is no longer effective.
The former is about the CUDA Version, and there are conflicting data in your posts: post #3554 indicates Version 11.2 whereas your last post refers to 10.0, which should in fact be the one used with GTX 20xx.

[James Heinrich]

Thanks to kriesel for recommendations, the top of https://download.mersenne.ca/mfaktc/mfaktc-0.21 now includes suggested download links for common GPU families, since it is admittedly quite confusing.

[AlvinBunk]

Originally Posted by James Heinrich

Thanks to kriesel for recommendations, the top of https://download.mersenne.ca/mfaktc/mfaktc-0.21 now includes suggested download links for common GPU families, since it is admittedly quite confusing.

Awesome! Thank you very much for adding this. My RTX 3060 is running the self test correctly now.

[kriesel]

Originally Posted by storm5510

The below is from a newer version than I have been using. I was trying to run the self-test using the -st parameter.

mfaktc v0.21 (64bit built)

CUDA version info
  binary compiled for CUDA  11.20
  CUDA runtime version      11.20
  CUDA driver version       11.70

CUDA device info
  name                      NVIDIA GeForce RTX 2080
  compute capability        7.5
...

########## testcase 1/2867 ##########
Starting trial factoring M50804297 from 2^67 to 2^68 (0.59 GHz-days)
Using GPU kernel "75bit_mul32_gs"
  Date   Time |   Cls    Pct |   Time     ETA | GHz-d/day |                 Bits 
Nov 05  13:17 |  3387   0.1% |  0.001    n.a. |      n.a. |    50804297    67-68  
ERROR: cudaGetLastError() returned 209: no kernel image is available for execution on the device

It requires cudart64_110.dll to run which I found. I have seen this before, but I do not remember the solution.

That attempt demonstrated a few things; RTX2080 is compute capability 7.5, the CUDA 11.2 binary tried did not include kernels supporting CC 7.5, and what the error message will look like in such a case. Also that Storm5510 is like most of us capable of forgetting what works given enough time. (Why I write a lot down!)

Originally Posted by storm5510

Yes, I am sure. The below is from the working build which I have been using for quite a long time.

CUDA version info
  binary compiled for CUDA  10.0
  CUDA runtime version      10.0
  CUDA driver version       11.70

I am trying to track down an issue where the screen goes blank, and the GPU fans go full throttle. The only way to prevent this is to use Afterburner to cut the power back to 70%, or lower.

I've seen such indicate dirty GPU fan blades or failed GPU fan motors, a hot running case due to slow or failed case fans, elevated environment temperatures (air conditioning not keeping up with lots of gear on a hot day), fan bearing lubricant becoming thick and tarlike, and occasionally a GPU that runs hotter than typical even when new. I suppose it's possible a heat sink gasket or compound is being ineffective, or a heat pipe sprung a leak. Reduced power improves power efficiency much more steeply than it reduces throughput per day. It's useful to do that regardless.

Originally Posted by lycorn

about the CUDA Version, and there are conflicting data in your posts: post #3554 indicates Version 11.2 whereas your last post refers to 10.0, which should in fact be the one used with GTX 20xx.

Neatly providing examples of what does not work and what that looks like in the first, and what does work in the second. CUDA apps are tricky for compatibility. The latest download may be incompatible with older hardware.

Originally Posted by James Heinrich

Thanks to kriesel for recommendations, the top of https://download.mersenne.ca/mfaktc/mfaktc-0.21 now includes suggested download links for common GPU families, since it is admittedly quite confusing.

James is so good to us, and this is a further example. He thought of adding that little table, and PMed a draft. I played editor/fact checker and added a line. The link to a forum support thread is a nice touch he included later.

[storm5510]

Originally Posted by kriesel

...I've seen such indicate dirty GPU fan blades or failed GPU fan motors, a hot running case due to slow or failed case fans, elevated environment temperatures (air conditioning not keeping up with lots of gear on a hot day), fan bearing lubricant becoming thick and tarlike, and occasionally a GPU that runs hotter than typical even when new. I suppose it's possible a heat sink gasket or compound is being ineffective, or a heat pipe sprung a leak. Reduced power improves power efficiency much more steeply than it reduces throughput per day. It's useful to do that regardless.

I had the system mostly disassembled yesterday. The fan blades were not dirty, just a little discolored. I cleaned them anyway. There was very little dust in the heat-sink itself. All three fans were running after I put it back together and started it.

At the 70% setting, mfaktc generally runs between 2,900 and 3,200 GHz-d/day. The core temperature rarely reaches 60°C. I still feel the PSU is no longer capable of producing what it used to. It is four years old. There are times when I would like to take it out and throw it as far as I can. It is not modular. It has a bunch of cables I do not need. It is a real pain the get that mess tucked somewhere out of the way.

All of it needs to be in a much larger case. Getting the GPU back in its slot is a chore. If the inside were any narrower, it would not be possible. I have to plug the PSU cables before I make the attempt. They are not reachable after it is mounted.