Archived

This topic is now archived and is closed to further replies.

tfm

Past the CPU bottleneck ...

Recommended Posts

... is a GPU bottleneck.Obviously all your components need to be pretty good to run FSX well, but the received wisdom is that "FSX is basically CPU-bound". My recent experience suggests that the technology may be reaching a point where this is no longer true.I recently bought a E8500 PC which I've overclocked to about 4GHz. This is significantly faster than my 3GHz Xeon system. The GPU in both PCs is an EVGA 8800 Ultra "factory super-overclocked". I've noticed a couple of things:First, frame rates have increased roughly proportionately to CPU speed in "easy" areas: ie, high altitude or nearer the ground in areas of low scenery density and complexity. This is consistent with the received wisdom about FSX being CPU-bound.Secondly, with GPU settings pretty much as Nick_N recommends, frame rates do NOT increase proportionately to CPU speed in "demanding" areas: ie, near the ground in areas of high scenery density and complexity. In such areas, framerates are sometimes only marginally faster than with my old 3GHz CPUs. This is inconsistent with FSX being CPU-bound.Thirdly, reducing the settings on the GPU and allowing FSX to take back control of AA and AF made no real difference with the 3GHz CPUs. But with the 4GHz CPU, taking these measures improves framerates considerably - although still not in direct proportion to the greater speed of the CPU. This is consistent with a bottleneck in the GPU, not the CPU.Accordingly, my first provisional conclusion is that for "demanding" areas - which is where (in my view) we most need higher framerates - the real bottleneck in my new PC is in fact the GPU, not the CPU. This leads me to the further provisional conclusion that Nick_N's GPU settings work best on CPUs with speeds of upto about 3GHz. To increase framerates for faster CPUs in the "demanding" areas where they're needed most, you need to start off-loading work from the GPU. In other words, once your CPU is faster than about 3GHz, it stops being the bottleneck in "demanding" areas.Tim

Share this post


Link to post
Share on other sites
Help AVSIM continue to serve you!
Please donate today!

The bottleneck is not the CPU? but the GPU? If that's the case, then many of us would be out buying the latest and greatest video cards. Your methodolgy is akin to finding a balance for the best performance in FSX and not in locating a bottleneck within FSX.So, to test your theory, go out and buy a brand new 9800GTX 0r GTX280, install it and tell me what great increase in performance you have witnessed within FSX. Now, increase your clock speed on your CPU another 0.2 to 0.5Ghz and see what kind of performace increase you get within FSX. I know where the gains are going to be made......and it's definitely not going to be found within the GPU. The reason "demanding areas" do not see a proportionate increase in framerate is because they are still DEMANDING on the CPU. Drop your complexity or autogen sliders (decrease the load on the CPU) and watch your framerates increase.

Share this post


Link to post
Share on other sites

>>Secondly, with GPU settings pretty much as Nick_N recommends,>frame rates do NOT increase proportionately to CPU speed in>"demanding" areas: ie, near the ground in areas of high>scenery density and complexity. In such areas, framerates are>sometimes only marginally faster than with my old 3GHz CPUs.>This is inconsistent with FSX being CPU-bound.>I don't see how you can logically draw that conclusion from what you observed...because:FS (any version) is primarily a software rendering program.You've drawn your conclusion based on a false premise, that is that the gpu is doing something. It's not doing that much in comparison to the cpu.Those scenery objects require CPU POWER to be drawn, not gpu power. FSX scenery is primarily rendered in software, not in the gfx hardware. This is a fact about the software. RhettFS box: E8500 (@ 3.80 ghz), AC Freezer 7 Pro, ASUS P5E3 Premium, BFG 8800GTX 756 (nVidia 169 WHQL), 4gb DDR3 1600 Patriot Cas7 7-7-7-20 (2T), PC Power 750, WD 150gb 10000rpm Raptor, Seagate 500gb, Silverstone TJ09 case, Vista Ultimate 64ASX Client: AMD 3700+ (@ 2.6 ghz), 7800GT

Share this post


Link to post
Share on other sites

>You've drawn your conclusion based on a false premise, that is>that the gpu is doing something. It's not doing that much in>comparison to the cpu.>Rhett,Would you consider that the following test would prove/indicate anything about CPU/GPU bottleneck:1. Measure fps with the system as it is.2. Measure fps with the GPU underclocked 25% (using RivaTuner) and CPU as normal.3. Measure fps with the CPU underclocked 25% and the GPU as normal.Could this indicate what component is most critical (bottleneck)?Ulf BCore2Duo X6800 3.3GHz4GB RAM Corsair XMS2-8500C5BFG 8800GTX, Creative SB X-FiFSX Acc/SP2, Vista 32

Share this post


Link to post
Share on other sites

I entirely agree that my observations are unexpected; they are certainly not what I had predicted - but I respectfully suggest that my inferences are supportable, and I am not sure that I follow how your reasoning contradicts them.Furthermore, my provisional conclusions are actually consistent with something I found last year when, in my 3GHz system, I replaced an 8800GTS with the 8800 Ultra. Received wisdom said it should make no difference. But it did: quite a big one, actually. I posted about it at the time on this forum and I think attracted the same sort of scepticism back then.Just to be clear: I am not for a moment discouraging the use of fast CPUs: overall, framerates are significantly better. It's just that the CPU seems only to be able to "break through" the most demanding areas if the GPU is taken out of the equation.Tim

Share this post


Link to post
Share on other sites

The frame counter (i.e., FPS) cannot be considered a reliable metric. A strong Vcard will simply repeat the same frame over and over until the next new frame arrives from the CPU. The frame counter will ramp higher, but the visuals "march to the beat" of the subjective viewer. As this occurs, the entirely subjective analysis of the observer becomes the metric. So, 1) An analysis that does Not consider this dynamic cannot provide a meaningful result. 2) An analysis that Does take this dynamic into consideration cannot provide a meaningful presentation. Good luck there!

Share this post


Link to post
Share on other sites

That's an interesting possibility Sam, but is it right? Surely it implies that - with a good Vcard - reported framerates will remain more-or-less constant whatever the situation, which evidently is not the case.Tim

Share this post


Link to post
Share on other sites

Of course not. That's the principle. When you get the "stutters," dats what's goin' on. The frame counter is meaningless. If it's visually smooth, fly-on. The Vcard Can influence frames . . . as long as the CPU can keep it fed.

Share this post


Link to post
Share on other sites

Sorry, you've lost me. But I agree with you about visual smoothness: now that I've got it, I'm just going to go back to enjoying the game - without the #### FPS counter on.Tim

Share this post


Link to post
Share on other sites

All depends on resolution as well. At high res. above 1680x1050, there will be a general increase in performance with a very good card...especially a 9800GX2.http://www.tomshardware.com/reviews/nvidia...80,1953-15.htmlAnd I am convinced that basically getting the best of both CPU and G Card is the way to go. At very high setting levels, both will be important. Arguing as to which is "most-dependent" is as futile as arguing that it was the knife that killed him, not the bloke holding it....I've yet to see 'convincing' evidence of either argument...inherent in FSX is massive variability in frame-rates.Typically, my frame-rates are jumping between 40 and 60 or 50 and 80...within seconds....When I perform benchmarks on Serious Sam II, they are 'replicable' - FPS scores are the same over and over, accurate to 1/10th of a frame/second - just about the antithesis of FSX.I even question the validity of Tom's benchmarks...showing 'statistically significant' differences at such low FPS (mid-20's) would be absolutely impossible...the 'error-bars' (s.e/mean) would be bigger than the Eiffel Tower.Better would be to devise a benchmark that is actually 'typical' of the settings and resolutions that a typical person actually uses......this would also produce larger FPS scores, the differences of which, would be more objectively, measurable.

Share this post


Link to post
Share on other sites

>... is a GPU bottleneck.I've looked at it this way all along. Here's my particular twist on reality: FSX is a CPU intensive program. There are situations in every flight where "the bottleneck" shifts around from CPU, to CPU-memory interface, to the GPU. This has been obvious to me all along, and came from watching, for example, how a weaker video card buckles in high density clouds. Just look at the range of frame rates between the various "high end" video cards in FSX. It's HUGE!:http://forums.avsim.net/user_files/192206.jpgIf he used the same or similar testbeds, then the common denominator is the CPU and memory, and the GPU obviously plays A HUGE ROLE in the final performance outcome. And what's worse, you can't evaluate this based on the "raw" horsepower of a given video card, which is born out in this comparison too. It's not only about how each component performs, but also how they all get along with each other. It's a matter of where the bottleneck exists during any given moment in any flight and in any given machine. Just depends what kind of task is at hand, moment by moment. So, what's the best arrangement? Parts that essentially can cope with each other well. A video card paired with memory and CPU that can "feed" it well, is all you can ask for. And "bottleneck": I think it's flawed to think of problems in performance being related (exclusively) to bottlenecks, if by that you are thinking bandwidth. Timing matters . . . in digital systems. Less so in analog systems.

Share this post


Link to post
Share on other sites

>It's a matter of where the bottleneck exists during any given moment in any flight and in any given machine. to quote someone recently... Ding Ding Ding DingStop the calls... we have a winnerthis is why that Raptor on PCIe RAID0 or SCSI drive is just as important as the memory subsystem and its latency/bandwidth which is just as important as the CPU speed (3.6 and above is best for FSX) which is just as important as a quad vs/dual in the system which is just as important as the video card core/drivers, etcits how the application addresses the system and visa versa at any given timePut them all together in the right way, and you have FSX goodness.. one difference in the chain link can mean different results at different points in time which makes it hard for those who are not tech savvy to observe and make the needed change.In normal game play without controls established the frame counter means nothing in FSX in reality.. you can not place a linear perf value on that number as it increases or decreases. There are too many operations going on in the background we do not see and/or can not measure to place ---->any<---- perf value on the frame counter number.

Share this post


Link to post
Share on other sites

There is a priority system at work. When the weakest link starts a chain of events then that priority system changes to deliver the content, therefore the result is dynamic not only to the scene but tumbles like a domino effect until the system can catch up.. and that can not be gauged by frames. If we had access to a list of successful completed operations per second and the benchmark control list of what is possible and should be completed per second, then we have a way to cross reference that back to the frame counter. At that point you not only have a FPS value but a SCOPS value which then will tell you if something you did or upgraded made a real difference..Past the numerical values the last part of that is observation of screen result in clarity and smoothness, something no counter or register will ever tell you.Right now it requires very keen observation, and, knowledge of what should be on the screen vs. what is not on the screen to assess FSX.Most can

Share this post


Link to post
Share on other sites

It would be very interesting to know what FSX should look like on a properly configured system, Nick: is there any chance you could post an HD video somewhere to show us all, along with your specs, using some of the really demanding scenery (eg - my particular bug-bear - Aerosoft's Heathrow 2008X)?I agree that it comes down to balance. But for non-engineering amateurs like me, getting the balance right is a pretty hit-and-miss experience. The recent debate about DDR2 vs DDR3 RAM is an example: the argument in favour of DDR2 makes perfect sense - if the extra cost is a factor. If not, DDR3 (from your posts) sounds like a "no brainer".I also agree that it is terribly easy to become fixated with the framerate counter. During cruise, for example, the difference between 30fps and 60fps is of no consequence at all. But on the other hand, I'm not sure the FPS counter can be discounted completely. For example, when I notice a choppy approach into Heathrow, I look at the framerate counter - and, sure enough, the choppiness in the display is accompanied by the reported FPS dipping into the low teens instead of hovering in the high teens/20s/30s. In that situation, saying that the FPS has dropped from 28 to 14 has meaning. A rough tool, admittedly: it gives no insight into the snappiness of texture loading and no doubt numerous other things, but it is at least a useful shorthand for conveying the experience of (relative) choppiness in at least SOME situations.Tim

Share this post


Link to post
Share on other sites

what does that frame counter tell you at low frames you can not see on the screen without the frame counter running?The only advantage to that frame counter is it will tell you if the problem is low frames (max system saturation vs. system ability has been achieved) or if the problem is a visually interpretable interrupted frame timing (poor frame transitions) due to an undefined, at that point, influence.that

Share this post


Link to post
Share on other sites

>There are too many operations going on in the background we do>not see and/or can not measure to place ---->any<---- perf>value on the frame counter number.I disagree with that analysis. Frame counter number below a certain value is the mother of all show stoppers. Once you get to acceptable frame rates then I would agree other variables trump frame rate since frame rate is no longer a pressing issue. Some set that at acceptable level at 15, or 20, or 30 or more. When you get below 15 it gets pretty dicey as you know. You have to give SOME credence to how many times an image is painted, and how motion is reflected in that. In frame counter rates of 100, sudden moves display as smooth motion. Not so in low FPS situations. There is fluidity which is clearly enhanced, barring stuttering behaviors, with higher frame rates I believe. QX9650 w/ Retail HSF|ASUS P5E3 Premium WiFi|4GB Muskin Ascent 7-6-6-18 1T DDR3-1600|EVGA 8800GT|Seagate SATA 2 x 2|Seagate Cheetah 15K.x|XP Pro SP2|Vista 64--maybe never to be installed

Share this post


Link to post
Share on other sites

you must have missed the post above yours... there is no "perf" value in that counterthere are only 2 things that counter will tell you and they are based on if playable frames are there or notneither tells you your system is better or worse than another in a linear format. That counter can drop to 10 because of a bad scenery package installed and lets forget addons.. it can also drop to 10 because you dont have your sliders set right for the hardware or something in the OS or background is affecting it I am not speaking to performance benchmarked against yourself, I am speaking to having that number compared to another persons system tell you anything of valueComparing to your own result depends on the type of frames and visual clarity/smoothness you are seeing and how it may change with a system tweak, hardware or a setting move as I posted aboveEDIT:LOL thread format... I meant this posthttp://forums.avsim.net/dcboard.php?az=sho..._id=43258&page=

Share this post


Link to post
Share on other sites

Here's some ammunition to support the view that the GPU may be the limiting factor in performance- at least in FS9.My system is now in its 6th year- truly a Smithsonian relic- AMD XP2200 1.8GHz- 2GB RAM. But it very nicely runs triple synchronized views on 3 monitors via a GeForceFX 5200 AGP card and a same type PCI card. FS9 settings are not maxed out but are very comfortable near mid range. Autogen Off.Now some numbers- (same situation/settings for all)Single view on 1 monitor- 34-46 FPS -Average 40Triple views on 3 monitors- 18-22 FPS -Average 20In the second case, the pixel output of 3 monitors has tripled (300%) with only a 50% decrease in FPS. A nice tradeoff!!Clearly, multiple views has enabled the CPU to generate more output. Phrased another way, the CPU was handicapped by lack of GPU!In fact this holds true for only 2 monitors being driven by same GPU.Here my pixels double with only a 33% reduction in FPS.Conclusion? Expand your GPU capability and watch your CPU strut its stuff!! Alex Reid

Share this post


Link to post
Share on other sites