Massive FPS gains in rf2 using PCI-e 3.0 x16 with higher end cards!

User ------------------ Spinelli - Ari
GPU Frequency (MHz) - 1163 - 1357
CPU Frequency (GHz) --- 4.5 - 5.1 (Haswell 4.8 is about equivalent to IB @ 5.1 and)
CPU cores ------------------ 6 - 4
PCI-E 3.0 lanes ---- 16x/16x - 8x/8x

Of course you will have a better Firestrike score than me. Firestrke, along with most games, is much more GPU dependant than CPU. Also, like most games, PCI-E 3.0 @ 8x is hardly a deficit to 3.0 @ 16x. So you definitely hold the advantage with your higher GPU clocks compared to my higher amount of CPU cores and PCI-E 3.0 lanes

Ari, we know that in Firestrike, and 99% of games that PCI-E 3.0 @ 16x/16x is only a 1-4% difference in frames-per-second compared to 3.0 @ 8x/8x. So don't worry, Ari, your 3.0 @ 8x/8x setup is absolutely fine for most games, HOWEVER, what we have discovered is that rFactor 2 is a MAJOR exception to this rule.

Ari, I'm still confused about what you're not agreeing with. Can you try to clarify it one more time, please?

 

Haswell supports up to three PCIe Slots coming from the 16 PCIe lanes of the processor itself, and can run in 16x, 8x/8x or 8x/4x/4x modes. Most Motherboards employs bridges that can simply route the lanes from one slot to another depending on if there is a card on it or not. Fixed 8x/8x should be good and viable because it is good enough for majority of users and removes some added complexity and cost due to the bridge chips.

Motherboards that advertise to have 16x/16x, uses a special switch between slot and processor that provides 16 lanes for each and this switch is limited by the fact than Haswell still has only 16 lanes.The only thing than the switch does, is reroute them on-the-fly based on GPU needs.This means that if you needed at some moment 12 lanes worth of bandwidth at one slot and the other card is Idle. (Switch actively gives one of the cards more bandwidth if it is using more.) With switch you would not have a bottleneck that you would on 8x/8x because it rearranges that on-the-fly and this most likely wont happen, you have the 16 lanes limit from Haswell, so if you needed more, you would still have a bottleneck and the switch will be good for nothing because it adds latency and overhead and limited performance benefit, because as most Video Cards can't saturate PCIe 3.0 8x Bandwidth.

 

Yes Ari, you are correct, PLX chips (switches) add latency and should be avoided if possible.

You are also correct with regards to Haswell - as well as Ivy Bridge and Sandy Bridge - only being able to provide 16 PCI-E lanes (3.0 in Haswell and Ivy Bridge, and 2.0 with Sandy Bridge).

The only way to get both GPUs running at true 16x PCI-E simultaneously is to use a Haswell-E (with tht exception of the 5820k which has been restricted to only 20 or 24 PCI-E lanes), Ivy Bridge-E, or Sandy Bridge-E processor.

I (and I'm pretty sure DrR as well) agree with everything you're saying. You're correct in all that.

What is it that you don't agree on? Or are we all just misunderstanding eachother while agreeing on the same thing?

 

That there is Massive FPS gains using PCI-e 3.0 x16 in rF2 with higher end cards when 2-way SLI is used, because if you look how things works with SLI, I just can not figure it out how there can bee huge fps gain.
Single card gives a big fps hit when PCI-e 3.0 x16 but SLI, I just don`t get it.

 

Spinelli, can it bee that in rFactor2 when benchmarking 2-way SLI that GPU load is so low to other card that with switch rF2 somehow takes benefit of PCI-e 3.0 x16?

 

Ari. I do not use PCI-E lane switching (PLX chip). I have an Intel Core i7-4930K CPU which is part of Intel's Ivy Bridge-E line of high-end/enthusiast desktop processors (LGA 2011 socket, and only compatible with X79-chipset motherboards).

These CPUs and motherboards, along with a few other types, natively support 40 PCI-E lanes, not 16. There is no switching involved to get 40 PCI-E lanes. If you have two GPUs in SLI, one will receive 16 dedicated PCI-E lanes, and the other will receive 16 dedicated PCI-E lanes. There is no switching involved. It is full, independent, seperate, dedicated 16 PCI-E lanes for each GPU - no switching involved.

Again, both cards are always fully benefitting from true, full PCI-E 3.0 x16 because each card always runs at full, independent PCI-E 3.0 x16 with no switching because my CPU/MB has a total of 40 native PCI-E 3.0 lanes, not a total of just 16 like the mainstream line of CPUs do.

I have both GPUs at almost 100% load (both always over 90%), and I have almost perfect SLI scaling. This is why I test rF2 and SLI only in 3D (Nvidia 3D Vision 2), because rF2 is completely broken with 2D-SLI, however, rF2 and SLI is the complete opposite in 3D; in 3D mode, SLI scaling in rFactor 2 is just about 100% perfect

 

So just performed the test at different pci-e modes and here are the results:

PCI-e 3.0 - 152 fps
PCI-e 2.1 - 130 fps
PCI-e 1.1 - 94 fps

It's not perfect linear scaling between 1.1 and 2.1 and i am curious why that is so. But here's the more interesting part.

I remember when i had GTX 460's in SLI that rf2 performance scaled very well (if i recall correctly 80% or greater) and i was not using stereoscopic 3d. I know that you (Spinelli) have said that only in stereoscopic mode in rf2 does SLI work correctly but this was not my experience from memory.

At the time i was using an i7-920 which i now realise had 32 PCI-e 2.0 lanes and i had the Asus P6T Deluxe which had full simultaneous x16/x16 SLI support. So each my cards were running at full pci-e 2.0 x16 mode.

Are you sure that SLI only works in stereoscopic 3d?

 

Yes DrR, tried it a million times with a million different rF2 builds, a million different Nvidia drivers and a million different SLI bits.

From a few days ago...

Triple screens (5760x1080), multiview, level 5 AA, and different graphics settings (mostly max but not all), different track, and different car than the official benchmark (just doing a quick test to check the SLI bits suggested above....

Stock rF2 SLI profile:
Avg: 51 - Min: 45 - Max: 56

"SLI compatibility bits" "0x00500005 (Windows System Assessment Tool)":
Avg: 58 - Min: 51 - Max: 63

Single GPU (In NVidia Inspector "Number of GPUs to use on SLI rendering mode" - "SLI_GPU_COUNT_ONE"):
Avg: 90 - Min: 72 - Max: 110

in 2D mode, SLI is still absolutely broken in rF2 and rF1 based sims like GSC, FT, etc. - at least while using triple screens. Thank god for 3D (NVidia 3D Vision) where the SLI scaling is just about perfect (90-100%).

And for even more tests - rF2 mega benchmark 1xGPU, SLI, single screen, triple screen, 2D, 3D --> http://isiforums.net/f/showthread.php?p=299541

 

Would you mind checking 1920x1080 only? That was what i was using with my 2x 460's. I really don't think i'm remembering it wrong.

 

Look at the very bottom link of my post above, I edited it with a mega benchmark link with all sorts of tests

 

Oh nice one! So like you said, only in 3D for any res does it work for you.

It's alll very inconsistent and peculiar. I'm pretty sure i was running the graphics at the time on max as well which doesn't fit with your data. But my memory is irrelevant as good evidence, i'd really need to test again to know for sure.

 

Yes you are right but it don`t change the fact that Haswell supports 16 PCIe lanes of the processor itself, by going SLI you don't double the number of lanes, you are just splitting them between two cards, LGA 2011 socket or switch.
What I am trying to say is that truthful 2-way SLI PCI-E 3.0 x16 does not exist, nor will ever in current mainstream platforms.

 

I can also confirm that in 3D mode, SLI scaling in rFactor 2 is just about 100% perfect.

 

Please, can you confirm me with maximum settings of everything ( shadows, textures, reflections, soft particles, multiview activated, fxaa activated, x16 anisotropic...) except aliasing, how many fps ( minimum, maximum ) do you have with new Honda NSX mod at Autodromo di Mores, with 9 AI drivers at 5760x1080 please? And with 2x aliasing?

I´m interested in 970 GTX, 980 GTX and 780 Ti. If you can check also with SLI and 3D, would be nice

It could be very helpfull for me to compare with my system, because I want to know if I need to wait until 980 Ti or a bit later.


More than 50fps minimum could be good news


Thanks for your help.

 

Ari, what your saying is very confusing. You seem to agree with Spinelli only to then contradict your agreement with the following sentence that "2-way SLI PCI-E 3.0 x16 does not exist". Spinelli has an Ivy Bridge-E cpu which supports 40 lanes of PCI-e 3.0. Does that not mean he can have those 40 lanes split into 3 active pci-e slots of x16/x16/x8 simultaneously? And if you're saying no, then why not?

Also saying "nor will (it) ever in current mainstream platforms" seems sort of nonsensical. If the current mainsteam platform does not have it then it will never have it because in the future, those previously "current mainsteam platforms" will no longer be the current mainsteam platform.

Sure Spinelli has a high end (non mainstream) CPU but i fail to see the relevancy of that to your argument that simultaneous x16/x16 does not exist? Also, no one here has said that Haswell supports more than 16 PCIe lanes (meaning you can only have 1gpu in x16 or 2 gpu's in x8/x8...unless you have the PLX chip you speak of allowing a switching of x16 to each card when needed...but how efficient is this vs true x16/x16?).

 

Yes, Ari, we completely agree with you. The "mainstream" processors like Sandy Bridge, Ivy Bridge, Haswell, etc. only have 16 PCI-E lanes, so if you use SLI with them, then each card only get's 8 lanes maximum (or even less if you have other PCI-E devices, for example, a soundcard). We completely agree with you, and you are 100% correct.

The "enthusiast" line of processors, though, like Sandy Bridge-E, Ivy Bridge-E, and Haswell-E have a total of 40 TRUE PCI-E lanes on the chip itself. So you can have SLI and both cards will be running at 100% TRUE 16x/16x, no switching needed for that. In fact I also have PCI-E soundcard in another PCI-E slot and both my GPUs still get 100% true, dedicated 16x to one GPU and another 16x to the other GPU. My motherboard doesn't even have the capability to do PCI-E lane switching (PLX chip) even if i wanted to.

 

How efficient is PLX chip PCI-e lane switching vs true x16/x16 anyway?

edit: I just read that it's only beneficial for 3 or 4 way SLI due to the added latency and that it can be detrimental to 2 way sli when enabled or giving no difference in x8/x8 or x16/x16 PLX mode. Not sure what game and system they tested this on though.

 

You CPU has 6 cores and 12 Thread and a thread of execution, is a software term for the basic ordered sequence of instructions that can be passed through or processed by a single CPU core and that fact can not bee changed by any MBO chip etc and it don`t mater how many PCI-e lines mbo has your CPU has 6 cores and 12 Thread and by going SLI you don't double the number of CPU cores or Thread , you are just splitting them between two cards.There is no 100% TRUE 16x/16x because of this simple fact.

 

What do cpu cores have to do with PCI-e lanes?

 

Why do you think i7-4770K can only run PCI-e 3.0 x8x8 and i7 4930K PCI-e 3.0 x16x16 ?