CPU Performance Database

[Donziboy2]

Im looking forward to seeing what kind of performance increases come from 0.23.

And I have a hard time swallowing the physics limited to one CPU core issue. I have a G15 keyboard that lets me view the CPU cores and I have never seen a single core maxed out or even close to maxed out playing KSP stock. I did add ECLSS a few weeks back and it was more then happy to peg one of my cores at 100% for no reason, lets just say I didn't play with that mod for long.

[DMagic]

And I have a hard time swallowing the physics limited to one CPU core issue. I have a G15 keyboard that lets me view the CPU cores and I have never seen a single core maxed out or even close to maxed out playing KSP stock. I did add ECLSS a few weeks back and it was more then happy to peg one of my cores at 100% for no reason, lets just say I didn't play with that mod for long.

It's possible for one core to max out while the others remain idle or low, but that's not generally what happens. The processing load is spread out evenly across all cores, so most of the time when someone says that a thread is maxing out one core you will see all four cores running at around 25%. There are probably lots of reasons for this that I'm not aware of, but that is the way it works. Being single-threaded doesn't limit a process to running only on core, it just means that it can't be split up to run simultaneously on all cores.

[_Aramchek_]

larger numbers to be handled.

.

Which has the effect, in video game engines, of speeding things up.

More information executed per clock=faster.

[DMagic]

Thanks VirtualCLD, I added both sets of results. I see a little bit of improvement from the Q6600 to the 4670k. It's nice to see that your results match up well with Slugy and Nao's.

[TUFOM]

I disabled AMD TurboCore and performance improved grately:

CSV FILE

[DMagic]

That's interesting, these results look much more like what I would expect from your CPU, but they never reach the levels of the old results. I guess AMD's turbocore can do weird things to KSP performance.

Actually, looking at this carefully, it seems like your old tests might have been done with the physics delta time set above 0.03 s/frame. Comparing the timing for stages and the way the simulation time compares between the two tests makes this seem likely.

Edited December 13, 2013 by DMagic

[DMagic]

So, anyone else getting better results? I tried this on a Surface Pro 2 and only got a very slight increase in performance. I'll see what I get on my desktop later today.

[Donziboy2]

So, anyone else getting better results? I tried this on a Surface Pro 2 and only got a very slight increase in performance. I'll see what I get on my desktop later today.

The Pro 2 is an Intel 1.6Ghz Dual core with onboard HD4400 correct?

[_Aramchek_]

I haven't tested it, but looking at your graph and my own "feel" of the game, performance is considerably more consistent/smoother.

[DMagic]

My desktop looks a little better. I'm getting around 20-30% increases in FPS at a given part count, which is pretty good, maybe not the 50-100% improvement that some people might have liked, but that's still a big increase. It's about the equivalent of moving up to a CPU a few notches higher.

And more subjectively, it seemed to me like things were smoother in 0.23. I don't know for sure, but it seemed better in ways that the fps data don't necessarily show. I'll have to run a frametime test sometime to see if that comes out any different.

Edit:

I haven't tested it, but looking at your graph and my own "feel" of the game, performance is considerably more consistent/smoother.

Yeah, that's what I saw too. I need to run a frametime test, those results in older versions were all over the place. Maybe some of that has been smoothed out now.

I ran these tests more carefully and ran a new test on 0.22 (which I didn't do on the Surface).

The Pro 2 is an Intel 1.6Ghz Dual core with onboard HD4400 correct?

Yeah, it's an i5 4200U at 1.6GHz with turboboost up to 2.6GHz. It runs about the same as the i7-4650U in the Macbook Air I used to have (despite the much better GPU in that part).

Edited December 18, 2013 by DMagic

[Kadathan]

Hey, my first post maybe, at least I don't recall posting earlier...

Anyways, I figured I'd try the benchmark out and give you some more data on the new .23 update.

<a href=http://www.filedropper.com/ksp2013-12-1718-28-47-91fps><img src=http://www.filedropper.com/download_button.png width=127 height=145 border=0/></a><br /><div style=font-size:9px;font-family:Arial, Helvetica, sans-serif;width:127px;font-color:#44a854;> <a href=http://www.filedropper.com >file storage online</a></div>

I'm running a 4670k clocked to 4.5ghz, nothing else notable.

I was kind of slow staging the thing as I found the staging to be pretty crowded and had to think about whether I was about to drop an active stage or not, but at least I didn't blow the thing up

If it's not usable I can try running it again.

[DMagic]

Lots of updates for this thread here:

First of all, I updated the first page with a lot of new information. I altered the rocket for 0.23, so anyone who wants to run the tests please download the new version from SpacePort.

Using tweakable settings I reduced the amount of fuel in every stage to bring the total in-game running time down to just under 4 minutes from about 9 minutes in the old version. Each stage is shorter, the SRBs fire for about 20 seconds instead of 30 for instance, but the major time savings come at the end. That way it's easier for people to run the test all the way to the last stage and not have several long burns to wait through.

I also adjusted the staging so that the next set of rockets immediately fire, removing the need for so many double staging events. There are now only 14 stages, instead of 22.

This is a comparison chart of the old and new rockets, it's obvious how much shorter the new version's running time is. It also preserves that tell-tale stair-step pattern, which is helpful in determining if someone is CPU or GPU limited (GPU limited results tend to look like more of a steadily rising curve). Performance at each stage seems to be about the same, so results from the old rocket version should still be valid, but it would be better if everyone ran the new one. Everything seems stable with this version, let me know if something goes wrong.

I also was curious about why 0.23 seems to be so much smoother despite fairly small increases in framerates. I ran a frametime benchmark with FRAPS (using the old version of the rocket), you can see the results below (the frametime numbers have been inverted to give FPS). While overall performance is obviously higher in 0.23, there doesn't seem to be much of a difference in the frametime pattern; the results are still all over the place, with as many as four different frametime "channels". FRAPs isn't really the best way to accurately measure frametimes, but it still seems that this might not explain why gameplay seems smoother in 0.23.

[DMagic]

Thanks for the results Bean, I updated the first page.

Looking at those frametime results again I notice that there are many fewer of those very slow frames, the ones running at about 5 FPS, or 200ms per frame. In the 0.22 results those frames show up pretty consistently about every 5 seconds. In 0.23 it's more like once every 15 seconds, although, curiously, they do seem to increase when the overall framerate goes up.

That could explain a lot of the smoothness that I've noticed in the newest version.

[_Aramchek_]

https://dl.dropboxusercontent.com/u/80346/KSP%202014-01-13%2013-11-37-33%20fps.csv

https://dl.dropboxusercontent.com/u/80346/KSP%202014-01-13%2013-11-37-33%20frametimes.csv

Same specs as before.

[Bean]

Probably not applicable to my run (My PC is a beast), but could you clarify what the settings should be, possible with a screenshot, though it only matters if the PC is GPU limited, which isn't usually the case with KSP. KSP doesn't seem to have a default settings button. I did my run with Terrain Scatters and absolutely every settings maxed. I think the Pixel Light Count and Shadow Cascades start around 8 (not 64 like my run). Also do you want a particular resolution?

If none of the above is relevant to the CPU comparison then I apologize, though as you are going by FPS graphics settings may come in to play, so minimum settings on everything but physics might be a better standard for the CPU comparison.

[jfx]

If someone has a compareable 0.23 / 0.22 fraps frametime csv I can run a brief statistical analysis with frappo on them (no time to play/test myself at the moment). Bonus late 90s style graphs included.

One reason for the felt leap in smoothness (no idea, haven't even started 0.23 yet) may be this:

If you have alternating frametimes of (e.g.) 10ms and 60 ms your game will always feel like you play it on ~17 fps (60ms/frame) despite actually measuring an average ~28fps. The slow frames give you a visual impression of a low framerate*. So if your slow 60ms frames (the lower"channel" in his last weird frametime/fps hybrid-diagram) get boosted to a 40ms "channel" you gain a visual impression of 25fps - a boost of 50% despite having only gained ~40%fps. Not the best example, but I hope you get the idea.

tl;dr:

FPS (especially average FPS) suck for a objective comparison of fluidity when playing a game. Better measurements are: % of time spent in frames rendered slower than 16.66ms/33.33ms (60fps/30fps) or better yet a graph of the cummulative distribution of frame times.

*This effect is known as microstuttering and was (and partially still is) a major factor for the suckiness of Crossfire/SLI Setups (They tended to have wildly varying frametimes, giving a verry sluggish impression despite having good fps ... which is why its always important to check the frame time distribution)

Edited January 13, 2014 by jfx

[DMagic]

Probably not applicable to my run (My PC is a beast), but could you clarify what the settings should be, possible with a screenshot, though it only matters if the PC is GPU limited, which isn't usually the case with KSP. KSP doesn't seem to have a default settings button. I did my run with Terrain Scatters and absolutely every settings maxed. I think the Pixel Light Count and Shadow Cascades start around 8 (not 64 like my run). Also do you want a particular resolution?

I've spent some time considering this, and I've run a lot of tests, but I don't think it matters all that much. The biggest single graphics hog is the terrain settings on planets with an ocean. Because of the way water works it's basically like rendering two planets at the same time. You can play around with terrain settings, or tweak the ocean terrain details in the settings.cfg file, but I've found that just tilting the camera up to keep the ground out of view works just as well, which is why that is the only graphics related part of my instructions in the first post.

I also want to keep the instructions short and require as few steps as possible; never underestimate people's ability to ignore directions.

Most of the other graphics settings don't have a huge effect on performance, rendering and texture settings make a bit of a difference (I don't think there is any terrain scatter around the KSC, so that probably doesn't really matter), but it doesn't take much of a GPU to handle those at their max values. Resolution obviously can make a big difference, but I assume most people running at very high resolutions (above 1080) also have decent graphics cards.

I'm also assuming that most people who have very powerful CPUs (and are therefore more likely to run into GPU limited situations) also have powerful graphics cards. I do see some results from laptops that look a little GPU limited though. Their framerates increase in more of a steady curve, instead of the stair step pattern I usually see. But the majority of the results I've seen don't appear to be GPU limited.

If someone has a compareable 0.23 / 0.22 fraps frametime csv I can run a brief statistical analysis with frappo on them (no time to play/test myself at the moment). Bonus late 90s style graphs included.

Here are my two frametime .csv files.

https://www.dropbox.com/s/c7s1omly1ken31d/KSP%20Frametime%20v22.csv

https://www.dropbox.com/s/4xkng8krlckg6gv/KSP%20Frametime%20v23.csv

[AcidEric]

Right then I've just done these bench mark thingys, totally new to this so I've kinda gone over kill a bit there is probably wayyyy too much info here but wasn't sure exactly what you needed

basicly all the graphics settings are all the way up & Set the physics delta slider all the way to the right; a physics delta-time per frame of 0.03s.

Turned off V-sync.

Launched the rocket without any additions or modifications in a fresh sand box session of KSP (didn't launch or fly anything else before this rocket).

Flew straight up until out of fuel and stages.

did 2 runs of the following, 1st just recording FPS data, 2nd recording FPS, Frame Rate & max/min/average. (once I'd figured out how fraps works)

• 2x Looking slightly up, No horizon visible, View stayed fixed
  
• 2x Side view of rocket, Horizon visible, View stayed fixed
  
• 2x Moved view around as I felt like
  

PM'ed the resulting .csv files over to DMagic or look here

Then had a look at CPU-Z

whew, over my head most of it, but got it to save a txt report. Here are the highlights (again sorry if it is more than you need) Think my Graphics card is holding me back should really get round to upgrading it.

CPU-Z TXT Report

-------------------------------------------------------------------------

Processors Information

-------------------------------------------------------------------------

Processor 1 ID = 0

Number of cores 8 (max 8)

Number of threads 8 (max 8)

Name AMD FX-8350

Codename Vishera

Specification AMD FX-8350 Eight-Core Processor

Package Socket AM3+ (942)

CPUID F.2.0

Extended CPUID 15.2

Core Stepping OR-C0

Technology 32 nm

TDP Limit 125 Watts

Core Speed 4113.7 MHz

Multiplier x Bus Speed 20.5 x 200.7 MHz

Rated Bus speed 2608.7 MHz

Stock frequency 4000 MHz

Instructions sets MMX (+), SSE, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2, SSE4A, x86-64, AMD-V, AES, AVX, XOP, FMA3, FMA4

L1 Data cache 8 x 16 KBytes, 4-way set associative, 64-byte line size

L1 Instruction cache 4 x 64 KBytes, 2-way set associative, 64-byte line size

L2 cache 4 x 2048 KBytes, 16-way set associative, 64-byte line size

L3 cache 8 MBytes, 64-way set associative, 64-byte line size

FID/VID Control yes

Min FID 7.0x

Base TDP 6 Watts

Boosted P-States 2

Max non-turbo ratio 20.00x

Max turbo ratio 21.00x

TSC 4013.5 MHz

APERF 4119.6 MHz

MPERF 3997.0 MHz

Chipset

Northbridge AMD RD9x0 rev. 02

Southbridge AMD SB910/950 rev. 40

Graphic Interface PCI-Express

PCI-E Link Width x16

PCI-E Max Link Width x16

Memory Type DDR3

Memory Size 16 GBytes

Memory Speed 2207.4MHz

Mainboard Model CROSSHAIR V FORMULA-Z

Display Adapters

Name NVIDIA GeForce GT 240

Revision A2

Codename GT215

Technology 40 nm

Memory size 1024 MB

Windows Version Microsoft Windows 7 (6.1) 64-bit Service Pack 1 (Build 7601)

DirectX Version 11.0

Hope this is of some use

Thanks AE

Edited January 14, 2014 by AcidEric
added Video & csv's link added

[Duke23]

Just ran this on my rig. How'd I do? lol

Mobo: MSI 970A-G46 (may be a limiting factor*)

CPU: AMD FX-4350 running at stock 4.2 GHz

GPU: MSI Radeon R7 260x OC Edition (overclock disabled at the moment)

RAM: 8GB DDR3 GSkill RipJaws @ 1600MHz

*I've read that this Mobo can't fully handle the 125w power draw of this CPU and will sometimes throttle it to keep from overheating the VRM. It will be replaced when budget allows.

Also I derped and decoupled one or two booster stages late, if that makes a difference.

https://drive.google.com/file/d/0B4VOBam9oo-3dXJXUDh6Njl6UUE/edit?usp=sharing

Edit again to add: Holy crap, I just went to run this test again for my own satisfaction... And I forgot to throttle up the rocket before activating because most of my rockets at the moment don't use launch clamps or SRBs (and some will destroy the payload if I don't throttle up gradually). So when I realized that of course I tried to throttle up and save it. The crash was amazing.

Edited January 14, 2014 by Duke23
add info

[Padishar]

I saw this thread and thought I'd give it a go on the pitiful excuse for a laptop I'm using at the moment. It has:

CPU: Core Duo T2600 2.16GHz

RAM: 3GB

OS: Win7 Ultimate 32bit

Video: NVidia Quadro NVS 110M 128MB dedicated, 1279MB shared @ 1360x768

Still had most of the graphics settings (including ocean config tweak) set for max performance but did set the 0.03s physics time and made sure vsync was off.

Only did the one run with view straight up.

https://www.dropbox.com/s/xgvlw1f1bnjst36/KSP%202014-01-14%2018-40-28-83%20fps.csv

Not exactly stellar performance I think you'll agree... Oh well, it's good enough for now to launch stuff like this and max out the tech tree with a few drives around on Mun and Minmus...

Edit: Did another run for frametimes in case anyone is interested...

https://www.dropbox.com/s/6mrjewdwvxv6s13/KSP%202014-01-14%2019-20-27-51%20frametimes.csv

Pad.

Edited January 14, 2014 by Padishar

[jfx]

Frame time differences between consecutive frames: Notice the "estimated - feels like x FPS" calculated Leonidas3 (Named after the guy who had the idea for the scaling). According to that 0.23 brought a felt FPS boost of 17%.

It is calculated by diminishing the influence of the fast frames on overall FPS. Leonidas3: Mean value of the 90% slowest frames. Fast frames <20ms are scaled back logarithmical.

0.22

0.23

More stats

Edited January 17, 2014 by jfx
versions were switched

[Padishar]

Here is an i7-3770 (not K, machine needs VT-x for VMs) at 3.4GHz with 8GB RAM on Win8 64bit though it doesn't have a graphics card in at the moment so is using the onboard HD4000. I did two runs, first at fairly decent graphic settings and then with everything dragged right down (though I should have probably reduced the texture res even more). Both runs were at 1650x1080 fullscreen.

Reasonable quality fps

Reasonable quality frametimes

Low quality fps

Low quality frametimes

Not sure how useful these are as I had a lot of other stuff running (nothing really active but using lots of memory). I'll try and do a run after a clean boot at some point (and stick some sort of graphics card in it).

Pad.

[DMagic]

Thanks for all of the results, I've added them all to the first page.

AcidEric, were you recording video during all of your tests? Your results seem really slow and recording video with FRAPS can really kill the frame rate. The FX-8350 is pretty similar to the FX-4350, it just has 8 cores to the 4350's 4 cores (and more cache), so I would expect performance to be similar between the two. On the other hand the GT 240 is a pretty low end video card, so you could just be very GPU limited, especially with all of the settings maxed out.

Those T2600 numbers actually look pretty good. They are somewhere around 50-75% higher than what similar types of CPUs were getting in 0.22, even if you are still running at 7 or 8 FPS that's a pretty big improvement. I can see how the HD4000 would limit the i7 3770 in your case. That's a pretty big disparity in CPU and GPU power. But the low quality results seem reasonable, they are similar to the i5 3470, which is about what I would expect since HT isn't really doing anything for KSP.

I think the frametime result labels for 0.22 and 0.23 are switched jfx. But those look interesting. It looks like each frame is alternating between two speeds primarily; one frame is 40ms then the next is 60ms, then back to 40ms. Then there are those occasional really slow frames, around 150ms per frame that really make things stuttery. I guess any reduction in those is good.

[Duke23]

Interesting info here! It's nice to see how everything compares, I actually feel a little better after seeing all the figures side by side. I agree with you something is up with the test on the FX-8350. I could see how the graphics might bog it down a bit, my 260x isn't that much better in the grand scheme of things and I find it limiting my FPS in some situations, particularly looking at a large ship with Kerbin in the background in LKO. ETA: I would expect the 8350 with the .1ghz overclock to run pretty much neck and neck with my stock 4350 as the extra cores aren't really going to help in this game or most games for that matter.

I may actually run this test again soon because I've made a couple changes to the computer hardware that could make a slight difference. I reconfigured my RAM to run properly in dual channel mode (long story) and applied the Microsoft scheduling patch for AMD Bulldozer / Piledriver chips, also a slight GPU overclock but that will be irrelevant when my new card shows up. Before running dual channel the lack of the patch didn't make much difference but between the three configurations (dual channel w/o patch, single channel w/o patch, and dual channel with patch) there are noticeable differences in FPS stability, load times, and stuttering in my modded Skyrim. I just did it yesterday and haven't run KSP again yet to test.

Also something to note is that with some mods in KSP I noticed an unbearable FPS drop. My tanker station with a bunch of crap docked to it went from ~20fps down to ~10fps. I think Kethane was the culprit but I can't prove it because I removed another mod or two at the same time when it got bad and my framerates went back to normal.

Edited January 22, 2014 by Duke23
OCD or something.

[Demodus]

AcidEric, were you recording video during all of your tests? Your results seem really slow and recording video with FRAPS can really kill the frame rate.

The free Fraps program caps you to 30fps while recording. AcidEric you should run it again and just use the benchmark option, defualt is 'F11' key.

Dmagic, I've PM'ed you the benchmark run for my laptop rig. Thx, keep up the good work on this project.