GTX 460 better times than GTX 660 TI

poppageek
poppageek
Joined: 13 Aug 10
Posts: 259
Credit: 2473733872
RAC: 0
Topic 197733

My GTX 460 1gig is doing Arecibo WUs in 5400 sec. The GTX 660 TI takes 6100. The 660 TI is now doing only Perseus and the times there are 19,000. Both GPUs are doing 2 tasks at a time. The 460 CPU is a dual core with nothing else running. The 660 has a 6 core running WCG on 5 cores, 1 for E@H.

GTX 460
Ubuntu 12.04 LTS
kernel 3.2.0-68
Boinc 7.0.65
Nvidia driver 331.38
CPU AMD 4450e 2.3gz

GTX 660 TI
Mint 17
kernel 3.13.0-24
Boinc 7.2.42
Nvidia driver 331.38
CPU AMD 6 core 3 gz

Any ideas or suggestions as to why the 460 is doing so much better? Both are dedicated crunchers.

Thanks

Jim1348
Jim1348
Joined: 19 Jan 06
Posts: 463
Credit: 257957147
RAC: 0

GTX 460 better times than GTX 660 TI

I suppose it is because the shader clock in the GTX 460 (Fermi) is frequency-doubled, whereas for the GTX 660 Ti (Kepler) it is not. Since the CUDA version is only 3.2 here at E@H, all the Kepler advantages do not come into play. It reminds me that I have a GTX 560 on the shelf that I have not looked at for a long time.

Phil
Phil
Joined: 8 Jun 14
Posts: 583
Credit: 228580955
RAC: 8065

Try running the 660 machine

Try running the 660 machine with no cpu tasks. At a minimum, each gpu task should have a full cpu core/thread reserved for it.

As a test, start with no cpu cores running any tasks. Then if you want to do cpu tasks, start running cpu tasks and increasing one core at a time until you see it start affecting your gpu times.

If you are running 2 gpu tasks, 2 cpu cores/threads, one per task, should always be reserved to support the gpu tasks.

I've stated all this without knowing the other specs on your computers as other factors can affect performance. Such as the chipset i.e. Z97, X79, etc. Memory quantity and speed. PCI bus spec, i.e. PCI2 or PCI3.

Hope this makes sense.

I personally don't run any cpu tasks on machines running gpu work. I treat the cpu as only existing to support the gpu.

Phil

I thought I was wrong once, but I was mistaken.

poppageek
poppageek
Joined: 13 Aug 10
Posts: 259
Credit: 2473733872
RAC: 0

Thank you jim1348 and AB9IZ -

Thank you jim1348 and AB9IZ - Phil for the replies.

I freed up another CPU core so each E@H task has 1 core and top showed the tasks going from 20-25% CPU to 30-35% CPU.

I will report back after a few WUs are completed but this looks fixed.

Thanks!

tbret
tbret
Joined: 12 Mar 05
Posts: 2115
Credit: 4863891961
RAC: 168937

RE: I will report back

Quote:

I will report back after a few WUs are completed but this looks fixed.

There are several things that might be contributing to my results, but my 460 also finishes Einstein CUDA tasks faster than my 660Tis.

Shafa
Shafa
Joined: 31 May 05
Posts: 53
Credit: 627005014
RAC: 0

I am quite surprised what you

I am quite surprised what you are saying about the 660ti, because 460s are a little old cards.

Thank you for posting that, I was about to buy a couple of 660ti or 670ti.

I have several GTX460s, different versions.
Great advantage of these cards is an OC potential.
All the cards are running on more than 800MHz (800-830), 24/7, pretty stable, without voltage adjustment.

ExtraTerrestrial Apes
ExtraTerrestria...
Joined: 10 Nov 04
Posts: 770
Credit: 578606871
RAC: 198543

Shafa, don't buy any Kepler

Shafa, don't buy any Kepler cards now that Maxwell is here. It's so much more power efficient, it will pay off over time. If you want to run nVidias at Einstein currently the GTX750Ti seems to be the most power efficient. If you prefer larger cards GTX970 should be for you, although it's not really stretching its legs compared to the 750 (see the Maxwell 2 thread).

MrS

Scanning for our furry friends since Jan 2002

Shafa
Shafa
Joined: 31 May 05
Posts: 53
Credit: 627005014
RAC: 0

Yep, that is clear. But cost

Yep, that is clear.
But cost of energy is not my concern, fortunately ;-)

Looks like 2nd hand GTX570 and GTX580 around 90EUR/each is the key :D

Petec888
Petec888
Joined: 3 Oct 06
Posts: 33
Credit: 1978975321
RAC: 0

I have a Gigabyte 660Ti and

I have a Gigabyte 660Ti and an eVGA 570 HD running in two separate machines. The 660Ti is running on a 2.66 GHz machine, the 570 on a 3.0 GHz machine. Both machines are quad core Xeons running Mac OS X and whatever the OEM driver is, under CUDA 6.0.54.

The 660Ti machine finishes a BRP4G unit in 9747 seconds, the 570HD machine does one in 7824 seconds, both of which are running two BRP4G's per video card.

Somehow I get the feeling nVidia chose to let raw CUDA compute take a back seat to graphics rendering between generations.

Petec888
Petec888
Joined: 3 Oct 06
Posts: 33
Credit: 1978975321
RAC: 0

Sorry, somehow double posted.

Sorry, somehow double posted.

archae86
archae86
Joined: 6 Dec 05
Posts: 3157
Credit: 7226004929
RAC: 1063135

My currently top-scoring two

My currently top-scoring two hosts have GTX 660s installed somewhat over a year ago as a swap for GTX 460s. I was looking to get power reduction and performance increase. I got some of each, but less than I hoped, and I think the power reduction was by a bigger factor than the performance increase.

One of these day, I intend to offer the 460s on eBay. And if I swap one or both 660s for Maxwell 2, those will go on offer as well.

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.