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

[SteveW 1,792]

 

More logical technical reasoning and referring to my point about "a problem halved"; worth mentioning is that disabling HT on the eight core CPUs still works very well compared to a four core which shows the task switching improvement better.

[westman 257]

I run AM 254 now

Bclk 133 give me 3200mhz on mems cl13 and uncore 4500.

 

Core 00 multi 35, core 01 multi 37 , and all other cores multi 35 my 5960x core2 (01) is the strongest core, can run multi 37 on core 2, multi 36 on core 1,3,4,5,6,7 the bad one is core 8 multi 35 with same vcore. With HT off you run atleast 0,2v higher vcore than with on, with less heat

Running maintread at 4.9ghz and the rest at 4.6ghz give me good performance

 

This is very fast and smoth,

[SteveW 1,792]

Running HT off to gain maximum GHz is a well known technique, and has been mentioned many times before. It is not some kind of reason that HT Off is better than HT On.

 

I run AM 254 now

254=11111110, the sim starts up on seven cores. We can show that the sim hardly improves going from three to four cores. So in effect you need only run the sim on four cores, which leaves the other cores free to run other processes, addons and sub-processes spawned by the sim. Instead with your setup, those extra three cores are used by the sim to no effect, but they do present the sim with wider affinity (7 rather than 4 cores). This has it's uses, but unfortunately when the sim gets going, those sim jobs utilise those cores fully, pushing aside other processes, like sub-processes of the sim. That results in a no-win situation. I would say you could do better but not by leaps and bounds.

[westman 257]

yes it start with 7 cores now, my intention was to load main tread to the strongest core first, and this happens to be core 2. If core 4 was my strongest core yes a started with 4cores.

Shall test to unasign cores until it be worse starting with the bad one nr 8.

[SteveW 1,792]

OK, I think it's worth a look. Maybe try 126 or 30, keeping the core 01 for the first core.

[Claude Troncy 4]

Bonjour,

 

Running HT off to gain maximum GHz is a well known technique, and has been mentioned many times before. It is not some kind of reason that HT Off is better than HT On.

 

My config :4790K AM = 85...
 

I decided just to see the impact of the AM to use only one core for P3D... AM = 1

Yes first core at 100% CPU,about a loss in FPS of 10% and a little more stutters... I expected a disaster... But No.
All those things seems to me very strange.

Cheers

Claude

[Guest]

However, I imagine my ASUS Strix GTX OC 980Ti's are closely matching the performance of your two Titan X's in SLI.

 

980Ti is essentially equal to a Titan X only the 980Ti has 6GB less VRAM ... they should produce very similar results.  In my table above I listed the OC setting for each Titan X (usually +200Mhz GPU and + 283Mhz VRAM ... I'll go higher with SGSS AA as GPU OC seems to be more beneficial with SGSS AA).  But normally I stay with 8X MSAA and don't use NI.

 

This flight was with HT OFF no AM and older 361.43 drivers using TrackIR ... add-on list used is extensive as are the graphics settings HERE.

 

 

I need to adjust my TrackIR profile as it's using a "Center" detent, but long frames are minimal considering the graphics settings and add-ons used.  So HT OFF no AM, TrackIR active, recording 4K content, and running P3D pretty smooth ... I'm just presenting my data/video on how it's working for me.  I have experimented with HT ON and many different AM settings, when it comes to my day to day flights, I'm finding HT OFF and no AM working the best for me, no need to "free" up any CPU, just terminate those "fake" CPUs wasting resources and L3 cache space

 

From what I can see when monitoring cores, Win10 with no AM will distribute the threads/loads better than restricting to an AM.

 

AA is entirely GPU based/accelerated it will make the GPUs work harder, but it will not impact the CPU's performance.  AA can be made to run "software only", but that would be nuts to operate AA that way as it would have serious impact on performance.  Final render is a combination of CPU/GPU, return to my long standing statement of getting the balance right in any system (from Low end to High end).

 

But to be clear, I'm definitely NOT telling or recommending what folks should do with settings, merely sharing my data/results with the given parameters ... they work for me, but they may not work for others.  My original intent was to show that SLI on my setup does indeed appear to have some benefit, even more significant with SGSS AA.

 

Cheers, Rob.

[GSalden 4,672]

I use AM 340 with my 6-core 5820k and W10..,

It runs very well.

Only when ASN is updating the weather ( ASN is on a client ) I see some jittering .

 

Over the next following days I am going to do some tests with no AM.

Just to test if I can improve it...

[Cruachan 705]

Hi Rob,

 

While checking out the 'extreme' settings used for the A2A T6-Texan flight I note you were running with VSync ON, Triple Buffering and Target Frame Rate (TFR)= Unlimited. VSync was OFF during your test flights.

 

Are you still syncing the video output to your 30Hz display?

 

Perhaps this goes some way towards explaining our different experiences. Like you, I too preferred Hyperthreading OFF with no Affinity Mask. Performance was good, but not quite up to what we are seeing in your videos. Moving away from dense urban areas improved things considerably, but this was because there was a commensurate rise in frame rates which naturally smoothed things out. Anything above 30fps was being perceived as smooth with barely perceptible stutters. However, while my TFR is also Unlimited, I am not using VSync and Triple buffering and my display refreshes at 120Hz.

 

Hyperthreading ON coupled with an Affinity Mask of 21760 (01,01,01,01,00,00,00,00) is producing measurable improvements in performance in a wider range of situations than before and this is without any anchoring of frame rates nor syncing with a monitor refreshing at 30Hz.

 

Cheers,

Mike

[Guest]

 

 

While checking out the 'extreme' settings used for the A2A T6-Texan flight I note you were running with VSync ON, Triple Buffering and Target Frame Rate (TFR)= Unlimited. VSync was OFF during your test flights.

 

When testing the max potential of CPU/GPU, I don't want any limiters, hence Vsync = OFF and Unlimited.

 

For my normal flights I run Vsyn = On (unlimited), hardware level limiter based on monitor refresh rate ... hardware limiters are more accurate and take no CPU resources.  My goal for normal flights at 30Hz refresh is 30 fps min ... I'll live with an occasional stutter but NOT frequent stutters ... pending location and add-ons used I'll adjust graphics settings such that I can sustain at least 30 fps.

 

Cheers, Rob.

[SteveW 1,792]

AA is entirely GPU based/accelerated it will make the GPUs work harder, but it will not impact the CPU's performance.

What is being said is that GPU setups can very obviously limit the sim performance and you don't want any of that to interfere with the CPU testing. It's best to back off the GPU to find differences in CPU throughput - if that's what you're testing.

 

 

I decided just to see the impact of the AM to use only one core for P3D... AM = 1

Yes first core at 100% CPU,about a loss in FPS of 10% and a little more stutters... I expected a disaster... But No.

All those things seems to me very strange.

In reality three cores are only just beaten by four cores. Five cores won't work well for a good reason (you need to know how P3D and FSX work).

 

Six works better but that's because it's only widened the affinity to six cores, leaving more holes for the sim sub-processes to run in. As the sim throughput ramps up, those extra two cores throughput are saturated by those two unused sim jobs 5 and 6, this reduces the bandwidth to those sub-processes. So in effect with low loads it may work well but eventually it'll get clogged up

 

HT off is only any good for saving a little heat with extreme overclocks. HT On doubles the available core count and confuses many multi threaded-apps so they need care taken with AMs.

[SteveW 1,792]

 

For those struggling with the concept of backing off the GPU function to test the CPU function, here's one analogy:

 

 

Let's say overall performance can be likened to the amount of light coming through a window with two roller blinds.

One blind represents the GPU, the other blind represents the CPU.

 

If we back off all the functions of the GPU, this fully opens the "GPU blind".

If we back off the sim complexity, this opens the "CPU blind".

 

We set up the sim complexity so we have desirable results, let's say that pulls the CPU blind down a third of the way.

We set up AA and VSync and we have desirable image quality, let's say that pulls the GPU blind down a third of the way.

 

Let's increase our AA quality and in so doing we pull the GPU blind down half way.

 

* We only have half the window showing wherever we set the CPU blind *

 

 

Maybe it's not a totally precise analogy but hopefully it goes some way to describe this dilemma, of testing the sim setup properly.

[SteveW 1,792]

 

I use AM 340 with my 6-core 5820k and W10..,

It runs very well.

Only when ASN is updating the weather ( ASN is on a client ) I see some jittering .

 

Over the next following days I am going to do some tests with no AM.

Just to test if I can improve it...

I remember suggesting 340 over 1360 as the HT enabled AM in an attempt to see if this helps the networked ASN client. That may be giving too much away to the system. I would also try 1360, and restrict all addon exe's to core zero (with AM=3 = 0,1). Leaves core 1 free rather than core 5.

[GSalden 4,672]

That is correct Steve.

 

I choose 340 over 1360 as I have 3 Prosim modules running on my server over the netwerk to my client pc and ASN from client to server.

 

That is taking some bandwith ...

 

The system is running great with 340 however.

Would you still suggest 1360 ?

[Claude Troncy 4]

you need to know how P3D and FSX work

Hi Steve,

For sure I need to know and I would like to !

What I understand is there is about 40 threads to run, and they must share 4 6 8 cores.

With such a number of thread it means for me that  FSX/P3D was designed with multithread in mind.

 

What I know about intel implementation of multithread, (Hyperthreading) is that it creates 2 logical processors on one core, and they must share the cache and the bus.

The cache default with HT on can affect the performance, but I do not know how much data is saved when FSX threads must be switched ! It can explain the debat AM or not AM that is the question !

 

To understand how FSX/P3D works, I think I need to know what are all those 40 threads for, and how they synchronize.... If you have information I 'll take it with pleasure.

 

Example of things difficult for me to understand.

Take off from Orly during the roll 50FPS.... a litle flying over the Eiffel Tower, and come back to the same position. Now the FPS is 40... Nothing has changed except the fuel... same weather. So it cost 10FPS. Why ? 

Cheers

Claude