Going from a GTX460 to GTX 600 Series.


log in

Advanced search

Message boards : Number crunching : Going from a GTX460 to GTX 600 Series.

1 · 2 · Next
Author Message
Profile D.A. Pinniger
Avatar
Send message
Joined: 16 May 02
Posts: 49
Credit: 9,931,729
RAC: 0
United States
Message 1310955 - Posted: 28 Nov 2012, 8:08:41 UTC

I'm currently running a Nvidia GTX-460 1 gig memory which has a memory bandwidth of 256 bits. On the less expensive GTX-600 series the bit width
drops to 192 bits. What do I sacrifice by going to the smaller bit width?

____________

juan BFBProject donor
Volunteer tester
Avatar
Send message
Joined: 16 Mar 07
Posts: 5414
Credit: 306,696,202
RAC: 331,194
Brazil
Message 1310960 - Posted: 28 Nov 2012, 8:34:48 UTC - in response to Message 1310955.

I'm currently running a Nvidia GTX-460 1 gig memory which has a memory bandwidth of 256 bits. On the less expensive GTX-600 series the bit width
drops to 192 bits. What do I sacrifice by going to the smaller bit width?

Only your big electric bill...
____________

N9JFE David SProject donor
Volunteer tester
Avatar
Send message
Joined: 4 Oct 99
Posts: 12001
Credit: 14,662,340
RAC: 12,019
United States
Message 1311039 - Posted: 28 Nov 2012, 14:38:50 UTC

This got me thinking... If I were planning to replace my GPU, I think it would be prudent to first set NNT and finish off the GPU work on board. Is there a way to set NNT for the GPU only and continue to receive work for CPU? It seems to me that setting Boinc not to use the GPU would immediately stop all GPU work and what I have wouldn't get done.

____________
David
Sitting on my butt while others boldly go,
Waiting for a message from a small furry creature from Alpha Centauri.


Richard HaselgroveProject donor
Volunteer tester
Send message
Joined: 4 Jul 99
Posts: 8634
Credit: 51,631,207
RAC: 49,026
United Kingdom
Message 1311047 - Posted: 28 Nov 2012, 15:07:54 UTC - in response to Message 1311039.

This got me thinking... If I were planning to replace my GPU, I think it would be prudent to first set NNT and finish off the GPU work on board. Is there a way to set NNT for the GPU only and continue to receive work for CPU? It seems to me that setting Boinc not to use the GPU would immediately stop all GPU work and what I have wouldn't get done.

Set the "Don't use [nVidia] GPU" preference in your account. Although a literal reading might suggest otherwise, in practice it does what you're asking - prevents new CUDA work being allocated (after one last scheduler update), but doesn't prevent you finishing off what you already have on board.

Profile D.A. Pinniger
Avatar
Send message
Joined: 16 May 02
Posts: 49
Credit: 9,931,729
RAC: 0
United States
Message 1311235 - Posted: 29 Nov 2012, 2:43:58 UTC

Gonk! The last 2 post missed it completely, It wasn't about stopping gpu or no new task. The question was what do I sacrifice from going from a gpu memory bandwidth of 256 bytes down to 192 bytes. Anyone else?
____________

Horacio
Send message
Joined: 14 Jan 00
Posts: 536
Credit: 75,130,246
RAC: 38,303
Argentina
Message 1311240 - Posted: 29 Nov 2012, 3:14:06 UTC - in response to Message 1311235.
Last modified: 29 Nov 2012, 3:15:03 UTC

Gonk! The last 2 post missed it completely, It wasn't about stopping gpu or no new task. The question was what do I sacrifice from going from a gpu memory bandwidth of 256 bytes down to 192 bytes. Anyone else?

The bandwith of the GPU memory is just one of several factors that makes the final performance of the GPU. Just comparing the bandwith there is not much to say... indeed, if everything else were the same then ovbiously you will lost some degree of performance... but if that lost is noticeable is beyond my guessing skills.

Ussually 600 series GPUs use memory that is much more faster than the memory used on several 400 and 500 series GPU, so it might happen that what you loose on bits its compensated by the raw speed...

Now, if you are thinking in something like a 640 or lower GPU, it will not be worth (compared to the 460) unless you were looking for a reduction of the power bill and you dont care about the reduction in performance...
____________

tbretProject donor
Volunteer tester
Avatar
Send message
Joined: 28 May 99
Posts: 2861
Credit: 215,797,360
RAC: 190,831
United States
Message 1311245 - Posted: 29 Nov 2012, 3:35:03 UTC - in response to Message 1311235.
Last modified: 29 Nov 2012, 4:33:06 UTC

Gonk! The last 2 post missed it completely, It wasn't about stopping gpu or no new task. The question was what do I sacrifice from going from a gpu memory bandwidth of 256 bytes down to 192 bytes. Anyone else?


I've started to comment three times and gave-up. There are too many things going-on here to make a nice, clean response, so I'm going to try to answer the question you asked and one or two you didn't.

You do lose something when you reduce the memory bandwidth on a like processor. Take a look at the difference in performance between a 3/4 GB 460 and a 1GB 460. The reduced "amount" of RAM meant to make things work-out evenly a reduced bandwidth to the RAM was used and you ended-up with a somewhat slower GTX 460.

But you are talking about moving from one architecture to another so comparing the memory bandwidth of a 460 to a 660 isn't going to translate.

Here's why I stopped trying to answer before now: I don't know why it doesn't translate. I've experienced the difference, I've owned the difference, I've seen the difference and... well, it's all about the way the card was designed to begin-with and often there isn't any difference, or bizarrely, the lower bandwidth card is faster. I don't know why.

As you've no-doubt noticed, the 670 has the same "bandwidth" as our 460s. If you read the reviews you find that an overclocked 660Ti is almost the equal of, or sometimes superior-to, a reference 670. The cards use the bandwidth their own way, so even though, on paper, the 660Ti is *clearly* inferior in every respect to the 670, all you have to do is goose the throttle a little and lo-and-behold, the difference disappears.

But you said, "less expensive" and I don't know where you draw that line; with the 660Ti, or the 660, or the 640 which might have a 192-bit memory bandwidth depending on "which" 640 we're talking about.

My EVGA superclocked 660Ti is faster than a 3/4GB GTX 460. I don't know that it is faster than a 256-bit bus, 1GB, GTX 460.

If you're looking at replacing your 460s with 660Tis, I'm not sure you'll lose anything. If you replace them with the superclocked version, I'm not sure you'll gain much in the way of efficiency.

But I haven't run Fred's (or anyone else's) tests on the two cards under identical conditions to be able to say for sure.

Having said all of that and made everyone's eyes glass-over and riling the heck out of fans of one GPU or another, Juan's observation is right: the 600-series cards are cooler and tend to burn a little less electricity than the 400-series cards with more-or-less the same "crunching power." Even the heat difference between a 460 (which wasn't a hot card) and a 660Ti probably isn't as dramatic in practice as it is in theory.

I know, I know, that only clears things up from foggy to hazy, but the question you asked may not lead to the information you really wanted to know.

EDIT: All of my observations are based on number-crunching and have nothing to do with gaming ability.
____________

Profile D.A. Pinniger
Avatar
Send message
Joined: 16 May 02
Posts: 49
Credit: 9,931,729
RAC: 0
United States
Message 1311252 - Posted: 29 Nov 2012, 3:48:22 UTC - in response to Message 1311245.
Last modified: 29 Nov 2012, 3:54:44 UTC

Thank you both for good information. Makes sense, and agree it's hard to compare apples to oranges(460 vs 6xx) : )
____________

N9JFE David SProject donor
Volunteer tester
Avatar
Send message
Joined: 4 Oct 99
Posts: 12001
Credit: 14,662,340
RAC: 12,019
United States
Message 1311346 - Posted: 29 Nov 2012, 15:36:31 UTC

Pardon a couple of dumb questions, but what's the difference between a GT and a GTX whatever, and what does Ti denote?

____________
David
Sitting on my butt while others boldly go,
Waiting for a message from a small furry creature from Alpha Centauri.


ClaggyProject donor
Volunteer tester
Send message
Joined: 5 Jul 99
Posts: 4141
Credit: 33,637,129
RAC: 27,943
United Kingdom
Message 1311456 - Posted: 5 Dec 2012, 20:20:11 UTC - in response to Message 1311346.

Pardon a couple of dumb questions, but what's the difference between a GT and a GTX whatever, and what does Ti denote?

a GTX is faster than a GT, and Ti means it's got a Turbo and injection too!! ;-)

Claggy
____________


Profile ivan
Volunteer tester
Avatar
Send message
Joined: 5 Mar 01
Posts: 625
Credit: 143,605,391
RAC: 150,187
United Kingdom
Message 1311507 - Posted: 5 Dec 2012, 22:04:30 UTC - in response to Message 1311456.

Pardon a couple of dumb questions, but what's the difference between a GT and a GTX whatever, and what does Ti denote?

a GTX is faster than a GT, and Ti means it's got a Turbo and injection too!! ;-)


Oh, I thought it meant titanium, and thus stronger and lighter!
(I once used a titanium tweeter (audio speaker) pressing as a window for an antimatter-chemistry experiment, to let low-energy positive muons into a chamber of hydrogen gas at red-hot temperatures and several atmospheres of pressure...)
____________

Profile dancer42
Volunteer tester
Send message
Joined: 2 Jun 02
Posts: 436
Credit: 1,159,660
RAC: 130
United States
Message 1311523 - Posted: 5 Dec 2012, 22:45:22 UTC - in response to Message 1311507.

sounds like a lot of hot air to me.

LOL
____________

Profile Fred J. Verster
Volunteer tester
Avatar
Send message
Joined: 21 Apr 04
Posts: 3252
Credit: 31,897,232
RAC: 3,739
Netherlands
Message 1311528 - Posted: 5 Dec 2012, 22:50:06 UTC - in response to Message 1311507.
Last modified: 5 Dec 2012, 22:56:21 UTC

---Snipped----


Oh, I thought it meant titanium, and thus stronger and lighter!

----snipped---

Titanium is indeed used, but I don't know for what reason and/or replacing
an other (conducting) material?

I stiil use an GTX470 and 480, cause 470 has twice the CUDA cores as the 460,
in stead off using 2 460s, I rather use 1 470.

But I'm thinking of adding 2 GPUs the 600 series is the better choice,
also a 470 maybe hard to get new ofcoarse.

Although the 600 series has a different architecture, also 64K regsize,
400 series (FERMI) has 32K regsize and most support PCIe 3.0 x16.
And are more efficient compaired too 470 or 480, which (can) run hot, 98C
or the fan can make a lot of RPM c.q. noise.
B.t.w. The 400 or FERMI series already were/are using GDDR5 memory,
quad pumped.

The CUDA app. provides no info about estimated GFLOPS, on 600 series.
Don't know if BOINC 7.0.28 displays an estimate
Atleast the 600 GPU has to have the same or more CUDA cores as
the 470 or 480 NVidia GPUs.
____________

jravin
Send message
Joined: 25 Mar 02
Posts: 941
Credit: 102,627,602
RAC: 90,459
United States
Message 1311953 - Posted: 6 Dec 2012, 22:23:36 UTC - in response to Message 1311942.

I use Vaseline (petroleum jelly) for all my CPUs - works great, costs zippo and does what thermal paste really is supposed to do: fill the air spaces between the processor and the heatsink. And it doesn't turn solid and crack, or evaporate or run out. And is easily cleaned up with a tissue. (I had 2 dual quad core Opteron systems on SETI 24/7 for a couple of years using the jelly and never had a problem with CPU temps).

I was inspired by an old post at Dansdata (Aussie) from 2002 (!) that you can find by Googling "thermal goop".
____________

Profile dancer42
Volunteer tester
Send message
Joined: 2 Jun 02
Posts: 436
Credit: 1,159,660
RAC: 130
United States
Message 1312011 - Posted: 7 Dec 2012, 1:32:59 UTC

THE LARGE AIR COOLER will fail softer than the water cooler.

the water cooler will do a better job.

the big air cooler will continue to work even if the fan die's just not as well.

If you use the water cooler be vary careful to install the hose's correctly

any mistake here will cost you dearly later.

also you may want to set the bios cpu cutoff temp low say 55c better to have an

occasional shut down then buy new cpu.

____________

1 · 2 · Next

Message boards : Number crunching : Going from a GTX460 to GTX 600 Series.

Copyright © 2014 University of California