Restarting GPU TF

[MatWur-S530113]

hiho,
I try to restart my TFing on my Geforce 1070 Ti, but I get an error when starting it:
running a simple Selftest
ERROR: CudaGetLastError () returned 8: Invalid Device function


I tried it with mfactc-0.20 and mfactc-0.21, always the same error (and the Wagstaff-version and the LessClasses version...)



2 years ago I used another GPU (Geforce 970) and now I have 2 monitors, one of them is running with the internal GPU of the Processor (oldish Intel 4670K). Maybe this forces the error.
Any help about this? Tia

[kriesel]

Review the setup steps at https://www.mersenneforum.org/showpo...18&postcount=1.

[MatWur-S530113]

Thanks, but no positive result here.
I tried different versions of mfactc (0.19, 0.20, 0.21), I installed Cuda from the Nvidia-page, I tried different Cuda versions, but always the same. With version 11.2 of Cuda I got a different error:


---------------------copy of selftest file



mfaktc v0.21 (64bit built)

Compiletime options
THREADS_PER_BLOCK 256
SIEVE_SIZE_LIMIT 32kiB
SIEVE_SIZE 193154bits
SIEVE_SPLIT 250
MORE_CLASSES enabled

Runtime options
SievePrimes 25000
SievePrimesAdjust 1
SievePrimesMin 5000
SievePrimesMax 100000
NumStreams 3
CPUStreams 3
GridSize 3
GPU Sieving enabled
GPUSievePrimes 82486
GPUSieveSize 128Mi bits
GPUSieveProcessSize 8Ki bits
Checkpoints enabled
CheckpointDelay 30s
WorkFileAddDelay 300s
Stages enabled
StopAfterFactor class
PrintMode compact
V5UserID ****
ComputerID OZI-GPU1070Ti
AllowSleep no
TimeStampInResults no

CUDA version info
binary compiled for CUDA 11.20
CUDA runtime version 11.20
CUDA driver version 12.0

CUDA device info
name NVIDIA GeForce GTX 1070 Ti
compute capability 6.1
max threads per block 1024
max shared memory per MP 98304 byte
number of multiprocessors 19
clock rate (CUDA cores) 1683MHz
memory clock rate: 4004MHz
memory bus width: 256 bit

Automatic parameters
threads per grid 622592
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 | class Pct | time ETA | GHz-d/day Sieve Wait
Nov 03 17:15 | 3387 0.1% | 0.001 n.a. | n.a. 82485 n.a.%ERROR: cudaGetLastError() returned 209: no kernel image is available for execution on the device


-------------------------------end copy



At least the 1070 Ti was recognized correctly.
All other versions had the error 8 returned as said in the OP

[kriesel]

Originally Posted by MatWur-S530113

Thanks, but no positive result here.
...

CUDA version info
binary compiled for CUDA 11.20
CUDA runtime version 11.20
CUDA driver version 12.0

CUDA device info
name NVIDIA GeForce GTX 1070 Ti
compute capability 6.1
max threads per block 1024
max shared memory per MP 98304 byte
number of multiprocessors 19
clock rate (CUDA cores) 1683MHz
memory clock rate: 4004MHz
memory bus width: 256 bit

Automatic parameters
threads per grid 622592
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 | class Pct | time ETA | GHz-d/day Sieve Wait
Nov 03 17:15 | 3387 0.1% | 0.001 n.a. | n.a. 82485 n.a.%ERROR: cudaGetLastError() returned 209: no kernel image is available for execution on the device


-------------------------------end copy



At least the 1070 Ti was recognized correctly.
All other versions had the error 8 returned as said in the OP

GTX10xx will happily run CUDA 8 executables, with compatible dlls and drivers, and do not support CUDA 10.x or CUDA 11.x compiled without downlevel support. Try this exe, assuming you're on Windows. (You can conduct such experiments in separate folders so as not to destroy the setups you tried previously.)
See also re CUDA & driver compatibility. That can be a significant consideration in selecting a suitable version of mfaktc or whatever for a specific GPU model.
While it's possible to compile CUDA apps to support multiple Compute capability levels & so support older GPU models, and level-independent PTX, there's no guarantee that the CUDA 9.x, 10.x or 11.x Mfaktc.exe builds available from download sites were built to either include PTX or CC 6.1 support.

[MatWur-S530113]

ah, I got it to work with mfaktc-0.21.win.cuda80.zip (the last version I tried out...) , no more errors, both selftests passed.
Thanks


Edit: what are the differences betwenn mfaktc-win-64.exe, mfaktc-less-cuda80-64.exe and mfaktc-more-cuda80-64.exe ? Only the number of outputs on the screen?


Ah, ok thanks again, I read the answer.

[kriesel]

You could get more performance by tuning with a -2047 flavor (2047M max GPUSieveSize instead of 128Mbit). See https://www.mersenneforum.org/showpo...99&postcount=8

[MatWur-S530113]

tested all 3 versions (for short time):


mfaktc-win-64.exe -> ~808GHzdays/day
mfaktc-less-cuda80-64.exe -> ~817 GHzdays/day
mfaktc-more-cuda80-64.exe -> ~827 GHzdays/day


Using the 'more'-version, thanks all, done

[kriesel]

Why not the recommended https://download.mersenne.ca/mfaktc/....21.win_cuda80-2047.zip? That's where the full performance is. (-more-classes for normal work, -less-classes for very low bit levels that would have very short class times in -more-classes, hence many console outputs and checkpoint saves per second, or on very high exponents. And if experiencing short class times, use SSD for faster save of checkpoints and logging of console output by redirection or tee.)
Proper tuning takes time, after settling on one executable; stop program, change ini file settings, save, restart program, run at least 30 classes, tabulate performance of last 20, repeat, one ini parameter at a time.
There's also the option of subsequently tuning power level for efficiency (throughput rate per watt used).
And of running multiple instances that may squeeze out a little more performance, after other tuning is finished.