Sign in with
Sign up | Sign in
Your question

Performance bottleneck

Last response: in Graphics & Displays
Share
September 3, 2012 10:26:01 AM

Hello,

I just recently got myself the Palit GTX 680 JetStream 2GB, having read reviews that it had good performance thanks it its factory OC and cooler.

I benchmarked my system, using Unigine Heaven 3.0 to see how much difference the card would make.

With tesselation on extreme, AA at 8x and AF at 16x I got a score of 820 and average FPS of 30 at 1080p.

However I have seen other results online where people seem to be reaching 45 FPS and above with this card at 1080p, and since Heaven is supposed to test GPU performance and reduce its dependence on other components I find it hard to imagine it being held back by anything else.

My question is: Is there any other component or part of my system that could be holding back my GPU's performance?
My setup includes an i7 3770k, 16GB of 1600Hz dual channel RAM, and a Gigabyte Z77X-UD5H motherboard.

Another thing that occurred to me is that it might be the FXAA (it is supposed to be pretty taxing), or the 2GB buffer being too small?

Thanks for any help or suggestions!

More about : performance bottleneck

a c 107 U Graphics card
September 3, 2012 10:26:02 AM

over clock your ram
a c 87 U Graphics card
September 3, 2012 10:53:01 AM

fXAA is not very taxing. The whole point of FXAA is that it's not taxing. I don't recommend using it because it is not real AA, but is a blur effect. It'll smooth things out a little, but it's like looking at the screen through a slightly foggier window. Smooth, but more out of focus. Some people like it, but some such as myself consider it a waste of performance.

Chances are that the 2GB VRAM capacity is probably not your issue either, at least not at 1080p. What frequency does your CPU run at and are you sure that those performance numbers from others aren't overclocked performance numbers?
Related resources
September 3, 2012 10:54:33 AM

I've read in many place that overclocking RAM makes next to no noticeable difference. Also, since I'm testing the GPU, surely the GPU VRAM is more important than system RAM?
a c 87 U Graphics card
September 3, 2012 10:55:42 AM

The VRAM bandwidth with Kepler GPU-based graphics cards makes a huge difference. They are very memory band-width bottle-necked. Compare the GTX 680 and the 670, same memory bandwidth, about the same performance, but the 680's GPU is about 20% faster than the 670's GPU.

Overclocking the system RAM probably won't make a significant difference.
September 3, 2012 11:03:28 AM

Thanks for the input blazorthon.

Take a look here, I don't think they mentioned anything about overclocking their card. I'll keep looking around, but my scores do seem quite low.

BTW my CPU is at stock clock speeds (3.5 GHz).
a c 87 U Graphics card
September 3, 2012 11:07:49 AM

You might want to overclock your CPU too. It will probably make a difference.
September 3, 2012 11:58:18 AM

Ok, I'll try that. Thanks for the help! I most post back results when after I've OC'd and benchmarked my system.
a c 87 U Graphics card
September 3, 2012 12:27:01 PM

Good luck.
September 3, 2012 12:40:54 PM

Do you have a normal HDD? If you do, then an SSD would be a biggish boost in performance
a c 87 U Graphics card
September 3, 2012 12:45:54 PM

aggawk said:
Do you have a normal HDD? If you do, then an SSD would be a biggish boost in performance


An SSD would help loading times and such, but it shouldn't change the FPS by much, if at all. It could definitely make the experience more enjoyable, but improving raw performance as measured in gaming FPS is unlikely.
a c 107 U Graphics card
September 3, 2012 12:58:07 PM



Overclocking the system RAM probably won't make a significant difference.[/quotemsg said:



are you sure about that ? if you only get an extra 5 fps by over clocking ram it can make a world of difference in some games.

over clocking the processor was my second bit of advice but decided to delete that part. I would lower the multiplier and push the ram more.
a c 87 U Graphics card
September 3, 2012 2:10:03 PM

swifty_morgan said:


are you sure about that ? if you only get an extra 5 fps by over clocking ram it can make a world of difference in some games.

over clocking the processor was my second bit of advice but decided to delete that part. I would lower the multiplier and push the ram more.


Overclocking the RAM is unlikely to get much of a performance increase at all. It usually doesn't help gaming much, especially on the AMD platforms and Intel's platforms since LGA 1156. Going past 1600MHz dual-channel doesn't help LGA 1155 CPUs much. They aren't very RAM bandwidth/latency-starved because they have very good memory controllers and cache.
a b U Graphics card
September 3, 2012 2:19:03 PM

What do you have running in the background? antivirus? other programs? Disable everything and then run your benchmarks
September 3, 2012 2:40:01 PM

I have literally nothing running in the background at all
a b U Graphics card
September 3, 2012 2:52:14 PM

ctrl+alt+delete->start task manager->processes and check how many processes are running and how much ram is being used.
a c 87 U Graphics card
September 3, 2012 2:54:22 PM

Here's another suggestion: Boot into safe mode, run a program called Driver Fusion (it's a newer version of Driver Sweeper) and use it to delete all graphics drivers from Nvidia, Intel, and Ati/AMD that you have. Re-boot into a normal boot, install the latest graphics driver for your card (it should be available at nvidia.com), re-boot again, and give it a try.

Also, I recommend doing what egilbe suggested. If you can give us a screen shot or two (enough to see all running processes), we might be able to see if anything doesn't belong.
a c 130 U Graphics card
September 3, 2012 2:57:20 PM

Ram over clocking is not pointless in fact on the Z77 board it makes much more difference than it used to, however tightening the times will give you more of a boost. I would look at getting tighter timings over faster speeds.
Just because its testing the GPU don't think that the CPU makes no difference, over clock it if you can.
Also I have seen reviews where lower clocked card have actually outperformed overclocked cards, not by much you understand but its still strange. Speculating I would guess it could be down to how the memory is configured on different cards as they seem to be sensitive to memory.

Mactronix :) 
a c 276 U Graphics card
September 3, 2012 3:01:10 PM

swifty_morgan said:


Overclocking the system RAM probably wont make a significant difference.[/msgquoted said:



are you sure about that ? if you only get an extra 5 fps by over clocking ram it can make a world of difference in some games.

over clocking the processor was my second bit of advice but decided to delete that part. I would lower the multiplier and push the ram more.


@ swifty_morgan
On Intel systems, fast ram makes little difference in real FPS. Think 1%. Synthetic benchmarks look marvelous though.
Read this:
http://www.anandtech.com/show/4503/sandy-bridge-memory-...

@ fat-chunk
The reason one buys a "K" is to do some overclocking.
For a conservative OC, just gradually raise the multiplier from the default 35 up to 43 or so. Leave everything else on auto.
That should reduce or eliminate any cpu issue.
a c 87 U Graphics card
September 3, 2012 3:02:51 PM

mactronix said:
Ram over clocking is not pointless in fact on the Z77 board it makes much more difference than it used to, however tightening the times will give you more of a boost. I would look at getting tighter timings over faster speeds.
Just because its testing the GPU don't think that the CPU makes no difference, over clock it if you can.
Also I have seen reviews where lower clocked card have actually outperformed overclocked cards, not by much you understand but its still strange. Speculating I would guess it could be down to how the memory is configured on different cards as they seem to be sensitive to memory.

Mactronix :) 


Graphics cards can be sensitive to their VRAM bandwidth. CPUs, for gaming performance, are not sensitive to the system RAM bandwidth nor latency. Outside of gaming performance, it can matter a lot in some applications (some scale roughly linearly with linearly increases memory bandwidth, especially AVX accelerated accelerations), but games tend to not care about memory bandwidth with the LGA 1155 CPUs once you go to 1600MHz or better. 1333 and 1066 can be bad, 800 can be pretty bad, but going over DDR3-1600 doesn't make much difference in gaming.

Now, going over anything on a graphics card, especially very memory-bandwidth limited graphics cards such as the Kepler-based cards, can help a lot. Memory bandwidth improvements and latency improvements can be very helpful when you're using an IGP such as AMD's APU IGPs, but CPU performance is not very sensitive in gaming to it.

Overclocking the CPU will probably help significantly, but the system memory is not a big deal. Gaming is kinda notorious for not being sensitive to system memory performance when IGPs aren't in use.
a c 130 U Graphics card
September 3, 2012 4:35:49 PM

Yes gaming does not show anywhere near the performance difference as synthetics.
Ivy bridge and Z77 chip sets show an increase in the benefits of faster memory up to and including 2133. 1866 is probably where you want to stop sensibly but if you want to push for bleeding edge and money is not an option 2-3 more FPS are yours by using 2133.
With Sandy bridge 1600 was the sensible stopping point so there is an increase in the performance gains.
As someone else said 5-6 Fps can make all the difference in some games, making unplayable playable.

Also I don't remember reading anything about kepler being very bandwidth limited, please share the knowledge with some links.
from what I have seen its pretty good.

Mactronix :) 
a c 87 U Graphics card
September 3, 2012 4:58:55 PM

Look at the GTX 680. Now look at the nearly identically performing GTX 670. Same memory bandwidth, but the 680's GPU is about 20-25% faster than the 670's GPU. That's a memory bandwidth bottle-neck. Now look at the GTX 660 TI versus the GTX 670. They have the same GPU, but the 660 TI has 25% fewer ROPS and 25% lower memory bandwidth. It does not perform on-par with the 670. That's a memory bandwidth bottle-neck.

The GTX 660 Ti sucks at overclocking performance despite being able to push the GPU to very high frequencies. That's a memory bandwidth bottle-neck. The GTX 600 cards (especially the 660 TI) suck at AA (FXAA is not AA, just so you know, it's a blur effect and that's very different) in comparison to the AMD cards that have much more memory bandwidth despite having faster GPUs (when tessellation and such are not used so we can isolate the performance difference of differing levels of AA) as proven by their higher maximum FPS. That's a memory bandwidth bottle-neck.

Any up to date review will show that the Kepler cards generally have inferior AA efficiency. AA is a memory bandwidth and memory capacity eating feature.

No offense, but you need to learn more about the hardware and the benchmarks to actually know much about the cards and such. Benchmarks are often deceiving if you don't know exactly what to get from them. Nvidia likes to make up for their lack of memory bandwidth with their GTX 600 cards that have Kepler GPUs with hype. For example, several GTX 660 TI reviews used cherry-picked 660 Tis that were binned specifically to perform above what people would actually get from the 660 Tis in retail. One such review showed a GTX 660 TI with its memory overclocked all the way to 1.9GHz (7.6GHz effective). That's not even close to being realistic for an air-cooled GTX 660 TI. Maybe a water cooled model with the memory ICs also cooled by the water block could go that high, but even that is not a guarantee that you'll hit so much as 1.8GHz (7.2GHz effective).

Like I said, benchmarks are often made to deceive people who aren't already very knowledgeable about those cards. Another thing that many of the reviews don't tell you is that Kepler has inferior tessellation efficiency to GCN, inferior minimum frame rates in most games, and other issues that vary from minor to somewhat significant if you compare them to AMD and even previous gen Nvidia cards.

Kepler is basically a chopped-down Fermi with a different core organization and a die shrink. Nvidia took a good architecture and basically ruined many of its strengths (good memory bandwidth for the GPU performance, good tessellation performance, good dual-precision compute performance, good AA performance as a result of the better memory bandwidth, and more) and traded them for profit margins. Small, very unbalanced GPUs, as few memory chips as reasonably possible, and small, cramped PCBs.

I'm not saying that someone can't be happy with a GTX 600 card that has a Kepler GPU, just that they really aren't as good as they're said to be for a wide variety of reasons. AMD isn't perfect, but most of their issues were software/driver, not hardware, and AMD fixed most of those while making significantly higher-performing drivers than they had before. AMD basically fixed most of their issues whereas Nvidia, at best, won't even start fixing their until GTX 700, if even that soon.
a c 87 U Graphics card
September 3, 2012 6:03:12 PM

You don't need links for that info because you can confirm it yourself by reading ANY review with the mentioned cards. What I said is entirely relevant. That was fact, not mere opinion.

If you want proof of what I said for the cherry-picked card issue, look at the Tom's hardware review of the 660 TI. They didn't get as cherry-picked of a card as some of the other review sites and they weren't too happy about that.
a c 130 U Graphics card
September 3, 2012 6:03:23 PM

Yea so anyway....... faster memory does matter to Z77 and Ivy.


You just wasted a lot of time typing that. Why I don't know as none of its relevant.
I asked for links not your opinion.
Thanks for the attempt at a lesson but your way of the mark my friend.

Mactronix :hello: 
a c 276 U Graphics card
September 3, 2012 8:05:11 PM

Motherboard DDR3 ram and graphics card Vram are entirely different things.

For Intel motherboard ram, there is minimal difference in real app performance or fps when overclocked to run faster.

Graphics card vram speed definitely does impact the card performance, and perhaps the quantity too.
a c 87 U Graphics card
September 3, 2012 8:23:08 PM

VRAM capacity, except on Nvidia's cards where the capacity doesn't match with the bus width, only improves performance when there isn't enough of it. If you have 2GB of VRAM and you only use 800MB, then you won't have any performance difference if you go down to 1GB or up to 4GB with the (otherwise) same card.

Still like you said (and what I've been saying), Intel's CPUs really aren't memory-bandwidth bottle-necked for gaming performance with 1600MHz memory. 1866MHz might make a tiny, very tiny difference... However, it won't be perceptible in-game, just in some benchmarks. There might be a few games that do rely on memory bandwidth more than most do, but the vast majority do not care much at all about it with the LGA 1155 CPUs.
a c 130 U Graphics card
September 3, 2012 11:48:48 PM

No blazorthonl you are wrong and have yet to post anything other than your opinion to support your stance.
Over clocking % results in excess of percentage gain with Kepler while its less with AMD.
I should point out here just for clarity that I an very much a AND fan but wont lower myself to slag off a really good Archetecture like kepler just to try and prove a point.

blazorthon
Post something and I mean a link that shows how kepler is bandwidth limited and supports you stand point or I will assume as will everyone else that you are just an AMD fanboy that has taken offence to Nvidia beating out AMD at ever relevant sec recently.

No waffle. links or you fail.

Mactronix :) 
a c 87 U Graphics card
September 4, 2012 12:04:03 AM

http://www.tomshardware.com/reviews/radeon-hd-7970-ghz-...
http://www.tomshardware.com/reviews/geforce-gtx-660-ti-...

http://www.anandtech.com/show/6025/radeon-hd-7970-ghz-e...
http://www.anandtech.com/show/6159/the-geforce-gtx-660-...

http://www.techpowerup.com/reviews/MSI/GTX_660_Ti_Power...
http://www.techpowerup.com/reviews/Palit/GeForce_GTX_66...
http://www.techpowerup.com/reviews/Zotac/GeForce_GTX_66...
http://www.techpowerup.com/reviews/AMD/HD_7970_GHz_Edit...
http://www.techpowerup.com/168056/AMD-Introduces-the-Ra...

I could have kept going for dozens of more links, but this seemed like it was enough. They all show that Kepler GPUs have memory bandwidth bottle-necks. I shouldn't need to give you links after what I said, but here they are. Hell, just looking at the memory bandwidth specifications at Nvidia's site and for non-reference cards at newegg or any other such site should tell you that GTX 600 has less memory bandwidth than competing Radeon 7000 cards. That you don't seem to understand how that affects performance doesn't help.
!