Performance of cuda-ecm on newer hardware?

fivemack · 2014-04-11, 18:14

On each of my GTX580 cards, I see 512 stage-1 curves run to B1=1e8 in six hours. Each card burns about 200 watts while doing this; my outbuilding is getting moderately warm and my electricity bill quite substantial.

Has anyone tried cuda-ecm on newer cards - 750Ti would be very interesting, whilst a Kepler-series (6xx or 7xx) would have the advantage that they'll be readily available on ebay once the 8xx start to appear ?

wombatman · 2014-04-11, 18:16

The 750 ti should be interesting as it is supposed to have the same computation power as older cards, but it has a significantly smaller watt pull (like 60 to 70W, I think?).

firejuggler · 2014-04-11, 18:38

60 watt; I just got one, will receive it next week.

Batalov · 2014-04-11, 19:07

Quote:

Originally Posted by fivemack

On each of my GTX580 cards, I see 512 stage-1 curves run to B1=1e8 in six hours. Each card burns about 200 watts while doing this; my outbuilding is getting moderately warm and my electricity bill quite substantial.

Has anyone tried cuda-ecm on newer cards - 750Ti would be very interesting, whilst a Kepler-series (6xx or 7xx) would have the advantage that they'll be readily available on ebay once the 8xx start to appear ?

I can try it for you on a 780.
Via nvidia-smi, it self-reports the "wattage" (true or not, that's another question).
What test case number (or at least size) would you propose?
I guess I'll take some newer Cunningham extension composites.

fivemack · 2014-04-11, 19:38

I'm currently running a couple of 17x-digit composites from aliquot sequences; why not try 5748.1526

Code:

1472719685748502449386007028264915543018048506185130332225484383890618193011424409284808724273216014754870734101746897136834531846846664397798081882555660328886965331001835843

Output from nvidia-smi on my machine (running one of these on each GTX580) is

Code:

Fri Apr 11 20:39:37 2014       
+------------------------------------------------------+                       
| NVIDIA-SMI 331.20     Driver Version: 331.20         |                       
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  GeForce GTX 580     Off  | 0000:02:00.0     N/A |                  N/A |
| 78%   83C  N/A     N/A /  N/A |    237MiB /  1535MiB |     N/A      Default |
+-------------------------------+----------------------+----------------------+
|   1  GeForce GTX 580     Off  | 0000:03:00.0     N/A |                  N/A |
| 65%   75C  N/A     N/A /  N/A |     62MiB /  1535MiB |     N/A      Default |
+-------------------------------+----------------------+----------------------+
                                                                               
+-----------------------------------------------------------------------------+
| Compute processes:                                               GPU Memory |
|  GPU       PID  Process name                                     Usage      |
|=============================================================================|
|    0            Not Supported                                               |
|    1            Not Supported                                               |
+-----------------------------------------------------------------------------+

So it is getting decidedly warm

wombatman · 2014-04-11, 21:31

Is there something akin to EVGA Precision X or MSI Afterburner for Linux? I use that to help keep the temperature down by adjusting the fan curve so it kicks up to higher speeds at lower temperatures. That seems to help a bit in keeping the temperature down. Still, yours seem a bit high.

Batalov · 2014-04-11, 22:42

No, there isn't. NVIDIA specifically disabled it in their linux drivers. (It used to be called "Coolbits" option.)
EDIT: ...or maybe they did restore it? See here.

Here's what temperatures look like on a quad-780 server:

Code:

+------------------------------------------------------+
| NVIDIA-SMI 5.319.37   Driver Version: 319.37         |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  GeForce GTX 780     Off  | 0000:02:00.0     N/A |                  N/A |
| 60%   70C  N/A     N/A /  N/A |      132MB /  3071MB |     N/A      Default |
+-------------------------------+----------------------+----------------------+
|   1  GeForce GTX 780     Off  | 0000:03:00.0     N/A |                  N/A |
| 71%   80C  N/A     N/A /  N/A |       72MB /  3071MB |     N/A      Default |
+-------------------------------+----------------------+----------------------+
|   2  GeForce GTX 780     Off  | 0000:83:00.0     N/A |                  N/A |
| 71%   80C  N/A     N/A /  N/A |      132MB /  3071MB |     N/A      Default |
+-------------------------------+----------------------+----------------------+
|   3  GeForce GTX 780     Off  | 0000:84:00.0     N/A |                  N/A |
| 71%   80C  N/A     N/A /  N/A |       72MB /  3071MB |     N/A      Default |
+-------------------------------+----------------------+----------------------+

(the card in slot 0 probably gets a better airflow)

I am running 4x768 curves. Let's see what the cat might drag in...

wombatman · 2014-04-11, 22:45

Man, that's disappointing...

Batalov · 2014-04-12, 00:39

On a Tesla:

Code:

+------------------------------------------------------+
| NVIDIA-SMI 4.304.54   Driver Version: 304.54         |
|-------------------------------+----------------------+----------------------+
| GPU  Name                     | Bus-Id        Disp.  | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap| Memory-Usage         | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  Tesla M2075              | 0000:42:00.0     Off |                  Off |
| N/A   N/A    P0   148W / 225W |   1%   66MB / 6143MB |     99%      Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Compute processes:                                               GPU Memory |
|  GPU       PID  Process name                                     Usage      |
|=============================================================================|
|    0     22022  ecm-gpu20                                             53MB  |
+-----------------------------------------------------------------------------+

Xyzzy · 2014-04-12, 00:44

http://mersenneforum.org/showpost.ph...postcount=2301

http://mersenneforum.org/showpost.ph...postcount=2305

Batalov · 2014-04-25, 07:15

Quote:

Originally Posted by fivemack

I'm currently running a couple of 17x-digit composites from aliquot sequences; why not try 5748.1526

Code:

1472719685748502449386007028264915543018048506185130332225484383890618193011424409284808724273216014754870734101746897136834531846846664397798081882555660328886965331001835843

Here is the output on a GTX780

Code:

GMP-ECM 7.0-dev [configured with GMP 6.0.0, --enable-asm-redc, --enable-gpu, --enable-assert, --enable-openmp] [ECM]
Running on avva
Input  number is  1472719685748502449386007028264915543018048506185130332225484383890618193011424409284808724273216014754870734101746897136834531846846664397798081882555660328886965331001835843  (175 digits)
Using MODMULN [mulredc:0, sqrredc:1]
Computing batch product (of 144266969 bits) of primes below B1=100000000 took 7804ms
GPU: compiled for a NVIDIA GPU with compute capability 3.5.
GPU: device 3 is required.
GPU: will use device 3: GeForce GTX 780, compute capability 3.5, 12 MPs.
GPU: Selection and initialization of the device took 18ms
Using B1=100000000, B2=110131572,  sigma=3:1249387146-3:1249387913 (768 curves)
dF=512, k=4, d=4620, d2=13, i0=21633
Expected number of curves to find a factor of n digits:
35      40      45      50      55      60      65      70      75      80
326     1658    9378    58245   393298  2863792 2.2e+07 1.9e+08 1.7e+09 1.8e+10
Computing 768 Step 1 took 1469430ms of CPU time / 32719214ms of GPU time
Throughput: 0.023 curves by second (on average 42603.14ms by Step 1)
Using 21 small primes for NTT
Estimated memory usage: 2272K
Initializing tables of differences for F took 0ms
Computing roots of F took 9ms
Building F from its roots took 9ms
Computing 1/F took  10ms
...

and on CPU:

Code:

GMP-ECM 7.0-dev [configured with GMP 6.0.0, --enable-asm-redc, --enable-gpu, --enable-assert, --enable-openmp] [ECM]
Running on avva
Input  number is  1472719685748502449386007028264915543018048506185130332225484383890618193011424409284808724273216014754870734101746897136834531846846664397798081882555660328886965331001835843  (175 digits)
Computing batch product (of 144266969 bits) of primes below B1=100000000 took 7786ms
Using MODMULN [mulredc:0, sqrredc:1]
Using B1=100000000, B2=110000000, polynomial x^2, sigma=2:3383285554610732763
dF=512, k=4, d=4620, d2=13, i0=21633
Expected number of curves to find a factor of n digits:
35      40      45      50      55      60       65      70      75      80
284     1436    8069    49838   334914  2428189 1.9e+07 1.6e+08 1.3e+09 7.9e+09
Step 1 took 532986ms
Using 21 small primes for NTT
Estimated memory usage: 2272K
Initializing tables of differences for F took 1ms
Computing roots of F took 7ms
Building F from its roots took 8ms
...

2014-04-11, 18:14	#1
fivemack (loop (#_fork)) Feb 2006 Cambridge, England 2×7×461 Posts	Performance of cuda-ecm on newer hardware? On each of my GTX580 cards, I see 512 stage-1 curves run to B1=1e8 in six hours. Each card burns about 200 watts while doing this; my outbuilding is getting moderately warm and my electricity bill quite substantial. Has anyone tried cuda-ecm on newer cards - 750Ti would be very interesting, whilst a Kepler-series (6xx or 7xx) would have the advantage that they'll be readily available on ebay once the 8xx start to appear ?

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Newer milestone thread	Uncwilly	Data	3668	2023-04-20 12:24
Newer X64 build needed	Googulator	Msieve	75	2022-06-13 14:22
CUDALucas / cuFFT Performance on CUDA 7 / 7.5 / 8	Brain	GPU Computing	13	2016-02-19 15:53
Xbox for TF (aka not higher performance hardware)	spaz	Hardware	4	2009-12-14 17:11
Newer msieves are slow on Core i7	mklasson	Msieve	9	2009-02-18 12:58

2014-04-11, 18:16	#2
wombatman I moo ablest echo power! May 2013 1856₁₀ Posts	The 750 ti should be interesting as it is supposed to have the same computation power as older cards, but it has a significantly smaller watt pull (like 60 to 70W, I think?).

2014-04-11, 18:38	#3
firejuggler "Vincent" Apr 2010 Over the rainbow 2²·3⁶ Posts	60 watt; I just got one, will receive it next week.

2014-04-11, 21:31	#6
wombatman I moo ablest echo power! May 2013 2⁶×29 Posts	Is there something akin to EVGA Precision X or MSI Afterburner for Linux? I use that to help keep the temperature down by adjusting the fan curve so it kicks up to higher speeds at lower temperatures. That seems to help a bit in keeping the temperature down. Still, yours seem a bit high.

2014-04-11, 22:45	#8
wombatman I moo ablest echo power! May 2013 2⁶×29 Posts	Man, that's disappointing...

2014-04-12, 00:44	#10
Xyzzy Aug 2002 2²×2,161 Posts	http://mersenneforum.org/showpost.ph...postcount=2301 http://mersenneforum.org/showpost.ph...postcount=2305