P3D V3.2 Performance SLI vs non-SLI and other stuff...

[Guest]

 

 

I have discovered that an Affinity Mask of 21760 (01,01,01,01,00,00,00,00) applied to the 5960X with Hyperthreading ON is providing the best performance in Prepar3D v3.2.

 

Hi Mike, I tried HT ON and your affinity mask 21760 and my results come out worse that with HT OFF no Affinity Mask.  See items 20-23 (22 twice is just a typo) below in updated results image:

 

 

Compare 20-23 with item 14.  In my system, 8C-8T (HT OFF no Affinity Mask) still provides the best results especially in the <0.1% providing a more consist time frame.

 

FRAPS time frame data results for tests 20-23 is HERE.

 

Cheers, Rob.

[Cruachan 705]

Hi Rob,

 

Interesting. I wonder why our experiences are differing? Certainly, I haven't done any frame time analysis - perhaps I should. My conclusion is based on delta line graph comparisons similar to those posted in recent threads and also what I am observing onscreen in a variety of situations and a/c. The sim is performing very smoothly while under quite heavy loads. Where are you running your tests? Probably shouldn't matter - just curious. I am also running Unlimited with VSync off and no Triple buffering for testing and general sim use.

 

Cheers,

Mike

[SteveW 1,792]

Trust your feelings Mike.

I would always suggest going with HT enabled, four LPs unmasked in the sim cfg, and leaving a core free from addon activity if you got one. Put the spare core in front of the sim cores, or after the sim cores if you want to share more with the system, maybe on a networked setup. Addon exe's by definition are likely running when the sim runs and do not want to be spilled onto the sim cores, or spilled onto the cores the sim spawned sub-processes may find themselves on.


I would like to reiterate that observing frame rates is a waste of time. Frame rates are disconnected from the sim.

Consider this:

1. We can alter the real performance of the sim by corralling it onto just one core with an AM, and see the fps is more or less maintained, and performance is not quite as bad as we imagine.

2. We can also move the autogen slider a notch to the left and gain 10fps.

We can be fooled into thinking we have better performance by observing the sim. But if we measure the sim properly, in the early stages we see poorer performance from the better setups. Concurrent background processes are taking less time to complete in the background and finishing in a different order - that can look lumpy.

Often a drop of a few fps, so that the main process can assimilate the data presented from the background tasks more timely, provides better overall performance. A good example is to experiment with FFTF. While most noise is made about reducing FFTF down to 0.1 to increase fps, and it may well do, but the result in quality is reduced. Why not try between 0.34 and 0.42, hands more time to the background process. Tune it with that and wait for two minutes of sim before measurements are taken.

The AM=116 configuration 01,11,01,00 (actually decimal 29 bit pattern 011101) is worth careful study. When used correctly even on a six core (where we could easily use four cores) with HT On AM=1856=01,11,01,00,00,00, beats 6 cores HT Off No AM. The decimal 29 bit pattern is a powerful number with FSX and P3D.

[JNS 106]

Post deleted

[SteveW 1,792]

What we do to set up a system is first we back off everything GPU related, AA, Vsync, maybe even reduce resolution output, and so on. Then we can concentrate on setting up the CPU side for max performance. After we're sure that's going on full steam, we can bring in VSync, and AA settings. AA is particularly good at limiting the sim and giving us improper information regarding the throughput capacity of the CPU setup.

[Guest]

Where are you running your tests?

 

Hi Mike,

 

The flight scenario files are linked in my original 1st post for this thread and all my graphics settings.  The table outlines variances from those graphics settings.  The situation I'm testing against is a setup I will typically fly in my daily flights under P3D V3.2.

 

I ran some more tests today going back to older nVidia drivers (based on Pete D. reporting he was getting much less stutters with 361.43 drivers) and of course back to HT off and no affinity mask.  I added two columns "GPU Driver" and "Count of Frames >40ms" ... frames >40ms are going to be noticeable "stutters".

 

Updated results:

 

The fraps raw data for tests 24-28 are HERE. 

 

Items of interest to me:

Test 23, 25 - HT ON vs. HT OFF (and/or possible GPU driver difference) - 8.1% better FPS and fewer long frames (>40ms) with HT OFF and no AM

Test 9, 26 - Driver 361.43 vs. 364.51 - 361.43 has 28.5% fewer stutters and 19% higher FPS in the 0.1% time

Test 27,28 - SLI provides 5.9% improvement even with NO AA

Test 20,21,22 - HT ON with an Affinity Mask produce 300-400% more stutters (frames >40ms)

 

Why our results are different I don't know, perhaps it's how we go about testing or something else different in our hardware and/or installed add-ons ... I'm working with a test scenario that is closer to what I'll actually be flying in sim with many add-ons.  I specifically included a default weather theme because it will introduce clouds but not be "random" like it would with ASN (clouds impact performance across AA settings) ... clouds will also produce cloud shadows out to 80,000.

 

Cheers, Rob.

[JNS 106]

 

 

Why our results are different I don't know, perhaps it's how we go about testing or something else different in our hardware and/or installed add-ons

 

Rob, are you both running 8-core processors?  I'm just wondering if somehow different models of processor(within the same class, i.e. 4-core, 6-core) handle the AM command and/or HT a little differently.  Of course, if you're both on 8-core (5960X, only model I know of) then that blows that theory.

 

Jeff

[Guest]

Rob, are you both running 8-core processors?

 

Yes, Mike and I are running the same CPU 5960X and same motherboard I think (Asus Rampage V Extreme with the same EFI/BIOS version) and I think at the same frequency 4.6Ghz.

 

I'm in the process of uploading a video of A2A's new T-6 Texan out of Orbx CNK4 with some lovely ASN weather and very high graphics settings (the T-6 shows off dynamic reflections very well on it's fuselage and wings), super smooth with HT OFF and no AM.

 

I did go back to nVidia 361.43 drivers as they produce far far fewer long frames (stutters) even in extreme settings/add-ons (Vsync 30Hz Unlimited).

 

Cheers, Rob.

[westman 257]

Iam new at P3D v3.2 , running a 6700k and 5960x system.

The 5960x works very good in P3D compared to FSX.

Test a lot with HT on and Off and AM21760.

 

The CPU runs very fine, and works best with HT Off as Robs system.

5960x with same MoBo , similar mems and single 980Ti

 

What a dont understand here, is the point that all guys that run HT off have not properly set up the system??

 

Keep up your good work Rob

[SteveW 1,792]

What a dont understand here, is the point that all guys that run HT off have not properly set up the system??

 

I'm simply unable to get ANY HT Off scenario working as well as with HT enabled. Seems logical to me that would be the case or Intel would be in trouble. Even though the sim main jobs don't utilise HT, they spawn upward of 50 threads that do.  Plus all the other system stuff runs better with HT enabled. I would say the HT Off guys are running something odd.

 

HT Off guys... what happens when you turn HT OFF? You halve the LP count, that is all. If this is "kind of" improving things for you, then you got something needs taming in HT ON, that's all it is.

 

And the upshot of that is all that's been done is halve a problem. A problem halved is not a problem solved.

[Cruachan 705]

Hi Rob,

 

8 Core - 16T / 21760

GPU Driver: 364.72

 

The main differences between our testing setups that are standing out are:

 

1. My 5960X is o/c to 4.4GHz (Base clock = 100Mhz, Multiplier x44)

2. Ram is not running as fast as yours.

3. Screen resolution = 2560x1440 (Refresh = 120Hz / GSync OFF)

4. AA is off for testing purposes as it can skew the results.

5. I have been employing the excellent IdealFlight 10 Pro (with Steve's ever-helpful guidance) as the front-end as it has the ability to collect data over accurately defined periods and data collection can be delayed to allow the sim time to settle down (I define 1 minute for the delay and 2 minutes for collection). It ensures consistency of testing environment (including weather and traffic) using the same flight scenario and it can be configured to run under its own affinity mask so will not interfere with the sim. That way I know that I'm comparing apples with apples with each test run.

 

Otherwise our test beds are very similar with the exception of the GPUs. However, I imagine my ASUS Strix GTX OC 980Ti's are closely matching the performance of your two Titan X's in SLI.

 

Normally I run very smooth with VSync OFF, Unlimited, 4x SSGAA in NI and MSAA = 4 Samples in P3D. Antialiasing is very good with minimal impact flying through clouds.

 

Cheers,

Mike

[SteveW 1,792]

 

Another thing. No attempt seems to have been made by any "HT Off Guys" to describe some logical, technical reasoning behind why they should be able to improve matters with HT Off. At least I've seen nothing.

 

Perhaps if the HT Off guys looked into that, they may come up with a way to run with HT enabled.

[westman 257]

I dont get dep into this, bad eng.

I run HT on apps that need it.

Overclocking for very very long time ,

Can give you some example on 3dmark from futuremark.

3D mark 06 benefit HT and 3D mark 05 not, and run HT on do nada.

 

Correct me 8core HT on is 16LP?

If you run 4LP on the sim what do the rest 12LP.s

With HT off 8LP with less heat , the Haswell Skylake IvyBridge SandyBridge that run 4core with HT off for years ar we all stupid?

 

With all respect , stupid or not a gone run HT off until a see any benifit with HT on.

[SteveW 1,792]

ar we all stupid?

No because there's the possibility of many problems with HT enabled, and that means guys like you will gravitate to HT Off, with otherwise no other way to cure the problems.

[SteveW 1,792]

 

If you run 4LP on the sim what do the rest 12LP.s

Well, as you may know, four of these are simply the HT partner LPs of the four cores allocated to the sim. What that means is we've allocated four whole real cores, half the cores of an 8 core, not a quarter. The underlying architecture is that each core can emulate a pair, it's still one core even if you only allocated one LP. If one can't get good results with the sim on four cores, that's an obvious problem showing up like a fireworks display.