llrCUDA

[Ralf Recker]

Originally Posted by em99010pepe

Ralf Recker,

First of all thank you. Second, can you post the specs of your machine (memory, harddrives, DVD-R, etc)? I want to make some calculations about energy efficiency so I need to know how many and type of components you have on it to make an energy consumption estimate.

A cheap uATX board (Series 3 Chipset), 4 GB DDR2-800 RAM plus a Core 2 Quad 9550 (E0 stepping) @ 3.4 GHz (400 MHz * 8.5) and a MSI Cyclone NGTX 460 (1 GB version, ugly cooler but effective ). The GPU ran at the standard (factory overclocked) settings @ 725 MHz Core / 1450 MHz Shaders / 900 MHz RAM (GDDR 5) @ 1.012V core (currently no working Linux overclocking but the card crunched without any problems @ 850 MHz / 1700 MHz / 1000 MHz under Windows). The temps during the test were around 45-48 degrees Celsius with the case closed.

The Q9550 runs with 1.1625V @ 2.83 GHz stock clock and with 1.20V @ 3.4 GHz. It's possible to run 3.4 GHz at lower voltages but I wanted to eliminate any risk of returning wrong results while crunching on PRP or LLR net.

[Ralf Recker]

Originally Posted by msft

GTX460:
5*2^1282755+1 is prime! Time : 4491.564 sec.
5*2^1320487+1 is prime! Time : 4447.951 sec.

I have two questions: Which OS and drivers did you use? Was the CPU otherwise idle or under load? The runtime differences (compared to my GTX 460) are significant. If your CPUs were under load it could be caused by GPU starvation...

[msft]

Dear Mr. Jean Penné
Happy New Year to you! I wish this year will be the happiest and best for you.
Yours sincerely,
Shoichiro Yamada

[msft]

Hi ,Ralf Recker

Originally Posted by Ralf Recker

I have two questions: Which OS and drivers did you use? Was the CPU otherwise idle or under load? The runtime differences (compared to my GTX 460) are significant. If your CPUs were under load it could be caused by GPU starvation...

OS:Linux ubuntu 2.6.28-19-server
driver:devdriver_3.2_linux_64_260.19.26.run
exec mprime same time.

[Ralf Recker]

Originally Posted by msft

Hi ,Ralf Recker

OS:Linux ubuntu 2.6.28-19-server
driver:devdriver_3.2_linux_64_260.19.26.run
exec mprime same time.

I've installed the 260.19.26 dev drivers and the 260.19.29 drivers (and the CUDA SDK 3.2) after my initial test (256.53 drivers and the CUDA SDK 3.1) and experienced a significant slowdown of the (BOINC/PrimeGrid) CWPSieve (CUDA) workunits from 2.5 minutes/WU to around 3.75 minutes/WU (CPU idle) or even 4.5 minutes (CPU under load - Sieving for NFS@Home). Since (older?) versions of Ken-g6's PPSieve (CUDA) and TPSieve (CUDA) are also working much better with the 256 drivers I've downgraded the drivers to 256.53 again (and the SDK to 3.1).

[Karl M Johnson]

Originally Posted by Ralf Recker

I've installed the 260.19.26 dev drivers and the 260.19.29 drivers (and the CUDA SDK 3.2) after my initial test (256.53 drivers and the CUDA SDK 3.1) and experienced a significant slowdown of the (BOINC/PrimeGrid) CWPSieve (CUDA) workunits from 2.5 minutes/WU to around 3.75 minutes/WU (CPU idle) or even 4.5 minutes (CPU under load - Sieving for NFS@Home). Since (older?) versions of Ken-g6's PPSieve (CUDA) and TPSieve (CUDA) are also working much better with the 256 drivers I've downgraded the drivers to 256.53 again (and the SDK to 3.1).

Yes.
NV messed something up in 260.xx .
OpenCL and CUDA(at least the Driver API ones) apps, compiled with 260.xx, dont work on older drivers(like 258.96).
OpenCL suffers major slowdown(up to 25%) in performance, while CUDA up to 5%.
That's my observations on applications I use.
What's curious, using 260.xx and 3.1 toolkit/sdk doesnt solve anything, so it's definitely the drivers.
258.96 are the last proper drivers for windows. I use them
For Linux, yeah, 256.53 are last proper too.

[msft]

$ time ./llrCUDA -q5*2^1282755+1

real 19m42.907s
user 8m58.070s
sys 7m21.650s
$ cat lresults.txt
5*2^1282755+1 is prime! Time : 1182.847 sec.

[Ralf Recker]

Looks like it reached the speed of a single CPU core

ralf@quadriga ~/llrcuda.0.11 $ time ./llrCUDA -q5*2^1282755+1 -d
Starting Proth prime test of 5*2^1282755+1, FFTLEN = 131072 ; a = 3
5*2^1282755+1, bit: 40000 / 1282757 [3.11%]. Time per bit: 0.819 ms.

To quote myself:

Quick comparison: Time per bit on the CPU: ~0.812 ms.

(Core 2 Quad Q9550 @ 3.6 GHz - Single Core)

Update:

ralf@quadriga ~/llrcuda.0.11 $ time ./llrCUDA -q5*2^1282755+1 -d
Starting Proth prime test of 5*2^1282755+1, FFTLEN = 131072 ; a = 3
5*2^1282755+1 is prime! Time : 1056.916 sec.

real 17m37.006s
user 4m51.894s
sys 9m27.603s
ralf@quadriga ~/llrcuda.0.11 $ cat lresults.txt
Bit 32105 / 1282757
5*2^1282755+1 is prime! Time : 1056.916 sec.

[Ralf Recker]

Originally Posted by Ralf Recker

Looks like it reached the speed of a single CPU core

ralf@quadriga ~/llrcuda.0.11 $ time ./llrCUDA -q5*2^1282755+1 -d
Starting Proth prime test of 5*2^1282755+1, FFTLEN = 131072 ; a = 3
5*2^1282755+1, bit: 40000 / 1282757 [3.11%]. Time per bit: 0.819 ms.

To quote myself:

(Core 2 Quad Q9550 @ 3.6 GHz - Single Core)

Update:

ralf@quadriga ~/llrcuda.0.11 $ time ./llrCUDA -q5*2^1282755+1 -d
Starting Proth prime test of 5*2^1282755+1, FFTLEN = 131072 ; a = 3
5*2^1282755+1 is prime! Time : 1056.916 sec.

real 17m37.006s
user 4m51.894s
sys 9m27.603s
ralf@quadriga ~/llrcuda.0.11 $ cat lresults.txt
Bit 32105 / 1282757
5*2^1282755+1 is prime! Time : 1056.916 sec.

Update 2:

Code compiled for sm_20 shows no significant difference:

ralf@quadriga ~/llrcuda.0.11 $ time ./llrCUDA -q5*2^1282755+1 -d
Starting Proth prime test of 5*2^1282755+1, FFTLEN = 131072 ; a = 3
5*2^1282755+1 is prime! Time : 1053.793 sec.

while code compiled for sm_21 is noticeably slower: Time per bit: 0.865 ms

[em99010pepe]

Originally Posted by Ralf Recker

Looks like it reached the speed of a single CPU core

Next step is to be 4x faster than CPU.

EDIT: Where can I find a list of all Nvidia GTX TDP's?

[em99010pepe]

For

GPU: GTX470 TDP of 215 W
CPU: Q9550 TDP of 121 W ( 4 cores at 3.6 GHz)
CPU timing for 5*2^1282755+1 of 1056 secs

Then

the GPU client should be 7.1 times faster than the CPU version to achieve the same relation Watts/candidates of CPU LLR client in 24 hours. This is the optimal for this pair, Q9550 and GTX470.