Hi,

I just re-read the CUDA documentation about streams... it is a bad idea to hope that the independent streams are executed in a specific order.

I'll try to fix this in 0.10.

Oliver

[quote=ckdo;220964]ERROR: cudaGetLastError() returned 8: invalid device function[/quote]
A simple "--gpu-architecture sm_10" flag while compiling fixes that.

[QUOTE=TheJudger;220987]Hi Ethan,



my first guess is that you modify those h_ktabs [B]before[/B] they are uploaded completely to the GPU.
[/QUOTE]

Ah, good call! In fact I was always writing the next h_ktab directly on top of the one most recently queued for transfer, so this would be a very likely occurrence. The failures are non-deterministic so this argues for the same problem.

Benchmark in GTX 460
 

Hi TheJudger,
[QUOTE=TheJudger;220929]Here is mfaktc 0.09! :smile:
[/QUOTE]

My GTX460 is cheaper version.
Inno3D N460-1DDN-G5GW
GDDR5 768MB (192bit-bus)
core clock:675MHz
mem clock:3600MHz (real clock:900MHz)

Thank you,

Hi msft,

thank you for the benchmark. :smile:

I can see that the GTX 460 ist at least as fast as my GTX 275 for the 75bit kernel. Seems to be still CPU limited (SievePrimes too high).
Can you rerun it with the mfaktc.ini in the current directory (it seems to be missing in your run, mfaktc.ini contains the runtime parameters)?

Thank you!

Oliver

[QUOTE=TheJudger;221228]Can you rerun it with the mfaktc.ini in the current directory (it seems to be missing in your run, mfaktc.ini contains the runtime parameters)?[/QUOTE]

rerun.:smile:

Thank you, msft!

so for the 75 bit kernel we have
~175% performance compared to my (mild factory overclocked) GTX 275 :smile:
~55% performance of an GTX 480

The 1GiB variant of the GTX 460 shouldn't change those timings, memory bandwidth isn't very important for mfaktc.
There are allready some GTX 460 with 1600MHz shader clock rate (1350MHz default) :max:

You made my decission easy... I'll upgrade to a GTX 460! :smile:
According to the reviews in the web the GTX 460 consumes less power than my GTX 275 and generate less noise. :smile:

Oliver

[B]msft[/B], I wonder why your GTX 460 has only 224 unified shaders.
Should be 336.
Could mfaktc be wrong?

nvidia Quatro doesn't work
 

Hi,

I compile mfakt 0.09 (with CUDA 3.1) and the program doesn't run.

Config:
Ubuntu 10.04
nvidia driver 256.35
CUDA 3.1
GPU: NVIDIA QUATTRO 140M

when I run mfakt -st here is the result:

[CODE]Compiletime Options
THREADS_PER_GRID 983040
THREADS_PER_BLOCK 256
SIEVE_SIZE_LIMIT 32kiB
SIEVE_SIZE 230945bits
VERBOSE_TIMING disabled
MORE_CLASSES disabled

Runtime Options
SievePrimes 25000
SievePrimesAdjust 1
NumStreams 3
WorkFile worktodo.txt
Checkpoints enabled

CUDA device info
name: Quadro NVS 140M
compute capabilities: 1.1
maximum threads per block: 512
number of multiprocessors: 2 (16 shader cores)
clock rate: 800MHz

cudaStreamCreate() failed[/CODE]

Any suggestion ?

PS:(I also tried the compilation option --gpu-architecture=sm_10)

Thanks

[QUOTE=Karl M Johnson;221242]
Could mfaktc be wrong?[/QUOTE]
Maybe.
[QUOTE]
if(deviceinfo.major == 1)i=8; /* device with compute capability 1.x have 8 shader cores per multiprocessor */
else if(deviceinfo.major == 2)i=32; /* assuming 32 shader cores per multiprocessor for compute capability 2.x */
printf(" number of multiprocessors: %d (%d shader cores)\n", deviceinfo.multiProcessorCount, deviceinfo.multiProcessorCount * i);
[/QUOTE]
Need "48"...

Hi Aillas,

[QUOTE=Aillas;221245]
[CODE]
Ubuntu 10.04
nvidia driver 256.35
CUDA 3.1
GPU: NVIDIA QUATTRO 140M
...
cudaStreamCreate() failed[/CODE]

Any suggestion ?
[/QUOTE]
no, not really. :sad:
Did you try the examples from the CUDA SDK?
---
hi msft, Karl

[QUOTE=msft;221246]Maybe.

Need "48"...[/QUOTE]
Yes, I need to adjust this. But it is only a cosmetic error. I calculate the number of shader cores only for display, nothing depends on this calculation.
I thought that some users might feel uncomfortable with the number of multiprocessors: "Hey, my GPU has e.g. 256 cores, why does mfaktc only show 32 (multiprocessors)?")
The calculation of shader cores was easy before Fermi: just multiply the number of multiprocessors by 8.
Now Nvidia has other configurations, too...
32 cores per multiprocessor (compute capabiltiy 2.0 / GTX 465/470/480)
48 cores per multiprocessor (compute capabiltiy 2.1 / GTX 460)
But Nvidia doesn't tell the number of cores per multiprocessor before they launch the products... But again, it is just a cosmetic issue!

Oliver