Not knowing much about SLI, can I ask why Nforce would cause an issue with bandwidth/bottleneck.
I've been thinking about the possibility of SLI for physX and 3D Vision, this has kind of made me think of what I'm missing, or need to consider.
-
Competitor rules
Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.
660 in SLI
- Thread starter snips86x
- Start date
Not knowing much about SLI, can I ask why Nforce would cause an issue with bandwidth/bottleneck.
I've been thinking about the possibility of SLI for physX and 3D Vision, this has kind of made me think of what I'm missing, or need to consider.
Because the CPUs which pair with those boards are so old they're likely to bottleneck today's graphics cards.
Soldato
- Joined
- 30 May 2012
- Posts
- 2,526
- Location
- Bristol
Plus the 660/660 Ti aren't really recommended for SLI setups due to the limitations of a 192 bit memory bus.
LOL says who ?
I used to run a pair of 5770s in Crossfire and they had a 128 bit memory bus and when Crossfire was working they trumped a 5870 which had double the memory bandwidth.
Please explain how extra memory bandwidth helps with alternate frame rendering that SLI uses? because the last time I did my research on how SLI and Crossfire worked the technique had pretty much nothing to do with the memory bandwidth and more to do with the fact that the cards simply split the work load and draw alternate frames.
Crikey I would love to know where you get your information from.
Does it look like the memory bandwidth is hampering the cards here?
http://www.guru3d.com/articles_pages/msi_geforce_gtx_660_twinfrozr_iii_review,24.html
Does it look like the memory bandwidth is hampering the cards here?
http://www.guru3d.com/articles_pages/msi_geforce_gtx_660_twinfrozr_iii_review,24.html
Soldato
- Joined
- 30 May 2012
- Posts
- 2,526
- Location
- Bristol
It's just what I've heard from most people.
Yes well it's wrong.
When ATI made the 5770 they thought they were putting a choke hold on it by cutting the memory bus in half. Reality was that it simply didn't matter.
See also - Kepler. It's a mid ranged chip that probably cost a whole chunk less than Fermi to produce (and Tahiti which was expensive) yet in the real world the cuts don't show at all as games are far less fussy about Direct compute and all of the other bells and whistles..
Same went for GTX 460 in SLI. They cost £100 less than a 480 and beat it.
When ATI made the 5770 they thought they were putting a choke hold on it by cutting the memory bus in half. Reality was that it simply didn't matter.
See also - Kepler. It's a mid ranged chip that probably cost a whole chunk less than Fermi to produce (and Tahiti which was expensive) yet in the real world the cuts don't show at all as games are far less fussy about Direct compute and all of the other bells and whistles..
Same went for GTX 460 in SLI. They cost £100 less than a 480 and beat it.
Soldato
- Joined
- 30 May 2012
- Posts
- 2,526
- Location
- Bristol
Then again, the 460 1Gb's had 256 bit buses, but point taken. I only mentioned it because big forum players like Gregster think so:
http://forums.overclockers.co.uk/showpost.php?p=23498785&postcount=4
Yes the memory bus was left intact on the 460 but it was heavily chopped in other areas. Mind you, what offset the heavy CUDA cut was the massive clock speeds. The less you put on a die the cooler it runs and thus the clocks can be very very high.
That's the trick with Kepler. Nvidia originally designed it to run at about half of the clocks it does.. Hence the 65% bios or whatever it was called. They just realised that due to it being mid range (compared to the technology of Tahiti) that they could ramp the clocks up to where they are now and the card would fly.
Fermi was disappointing because it was too hot and used a ton of power. All o f that heat and power were caused by things that don't equate into gaming performance. The same will go for Maxwell, which has not been designed with gaming in mind. It will no doubt have astronomical Direct Compute and so on, but none of that helps with gaming.
AMD found that out the hard way because Tahiti on paper should have demolished the mid range Kepler.
It is good at Winzip benchmarks though
Because I game at 5760*1080, I can see how the memory bus of the 680 fails against the memory bus of the 7970. Even a 7950 will leave me standing at my resolution. This is the reason I wouldn't recommend a 660 or TI for SLI. As soon as AA is applied, the frames drop off rapidly (and this is at 1080P). To get the best performance, the 7970/50/680/70 bring all the detail to the table. Go above 1080P and the 7970/50 starts to pull ahead quickly compared to the 680/70 and the lower memory bus of the 660/TI will be hit even harder(nothing to do with the amount of VRAM).
It's not wrong.
You end up heavily limited by memory bandwidth even at 1920*1080. You have considerable GPU grunt with two 660Ti's/660's but not the memory bandwidth to complement it. So any memory intensive effect such as AA (which you'd expect to be able to run easily with two cards) can cripple performance in today's games (i.e. not FC2 as in your link
).The effect is even greater on super HD resolutions or Eyefinity/Surround:
http://forums.overclockers.co.uk/showthread.php?t=18455827
The FPS numbers on this are still at playable levels but it's more about quantifying the drop off when AA is applied.
For reference:
7870 = 256 bit
7950 = 384 bit
660Ti = 192 bit
670 = 256 bit
Last edited:
It's not wrong.
You end up heavily limited by memory bandwidth even at 1920*1080. You have considerable GPU grunt with two 660Ti's/660's but not the memory bandwidth to complement it. So any memory intensive effect such as AA (which you'd expect to be able to run easily with two cards) can cripple performance in today's games (i.e. not FC2 as in your link
The FPS numbers on this are still at playable levels but it's more about quantifying the drop off when AA is applied.
For reference:
7870 = 256 bit
7950 = 384 bit
660Ti = 192 bit
670 = 256 bit
Due to the way SLI and Crossfire work the memory bus won't stop it accelerating where it matters.
I'm not saying that the memory bus doesn't affect the performance, all I'm saying is that it doesn't hurt SLI and Crossfire due to the way they work.
SLI-ing a pair of cheap cards (especially now that it's supported very well and looked after by EVGA with their patches) always gains you more performance than buying a single high end card.
It has done for about 3 years now. AMD know this now as they stop you crossfiring more than two 7850s for example.
Nvidia also limit their low end cards to 2 way SLI. That was my only complaint about the 460 tbh.
Yes of course it doesn't hurt but I think the point is that you aren't getting the maximum out of the cards due to being limited by the memory bandwidth and therefore it isn't a very good option.
You're getting the maximum out of what they can offer as the cards are but it represents bad value for money especially when you've got 2 GPUs to be limited by the memory bandwidth (depending on the game).
Raw FPS numbers will be higher, I agree, but 660s/660Tis are pretty much on the edge of not being able to play some of today's games at acceptable frame rates already due to the memory bandwidth so adding a second GPU will only exacerbate this internal 'bottleneck' of sorts.
(I concede that FPS numbers will still increase by adding a second card of course
)
Last edited:
Would I see the issue if I was using 660 in SLI on 2x 24" 3D monitors like these:
http://www.overclockers.co.uk/showproduct.php?prodid=MO-051-BQ&groupid=17&catid=1851&subcat=
I'd hate to shell out on a 680 as the prices are still too high!
I'll most likely be using x16/x16 or x16/x8.
http://www.overclockers.co.uk/showproduct.php?prodid=MO-051-BQ&groupid=17&catid=1851&subcat=
I'd hate to shell out on a 680 as the prices are still too high!
I'll most likely be using x16/x16 or x16/x8.
Well you are getting the maximum out of them as that's what they are capable of per the spec they have been made at.
I just don't agree with what people are saying about the memory bandwidth hampering SLI or even Crossfire when the figures (benchmarks and games) totally disagree. There is an enormous performance increase over running one card.
And that's down to how Crossfire and SLI work.
You're slightly missing the point. It's not hampering CF/SLI at all and I've already agreed that they're running at the maximum as per spec but the spec itself is the problem and doesn't represent that good value for money IMO especially for an SLI option.
By pairing 2 mid range GPUs with a 192 bit bus you're just exacerbating the problem when the cards are running individually (i.e. they're bottlenecked by their memory bus). You're adding another load of GPU power but the bus width is obviously the same and becomes an even larger limiting factor.
Last edited: