When I have researched the effects of the PCIe bus, Reject Threshold and BP=0 offers improvements in FPS and stutters at times when changes in PCIe bandwidth is the only hardware change that have any significant impact. So if BP=0 or Reject Threshold is giving you improvements. Higher PCIe bandwidth will also give you improvements. The opposite is also true. Lower PCIe bandwidth will give you even worse performance.The GTX590 is equiped with a NF200 PCIe bridge chip. So the CPU is feeding 16 lanes to the NF200 that in turn is feeding 16 lanes to each of the two GPUs on the card. So one GTX590 is a lot better then GTX580 in SLI on the 2600K for FSX.

I believe the NF200 doesn't really increase the bandwidth. You get more lanes, but at half the speed. It's a multiplexer but it can't create BW out of nowehere Lars. The PCIe controller is in the CPU and it has the 16x 2.0 bandwidth max limitation.About your PCIe 3.0 tests, and dont get me wrong, you're the first one ever to expose the PCIe bottleneck here that I'm aware of so big kudos to you for that, I personally get very different results.I really don't know what to think, but I'm under the impression that PCIe 3.0 will make little to no difference with complex aircraft like PMDG's that run so heavy on the CPU. I mean, we keep getting almost 1:1 improvements with CPU speedups, while PCIe bandwidth has remained the same for years...I'll be waiting for results, and ready to eat crow if wrong ^_^

Edited February 17, 201214 yr by dazz

I believe the NF200 doesn't really increase the bandwidth. You get more lanes, but at half the speed. It's a multiplexer but it can't create BW out of nowehere Lars.

Does it not enable x16 comunication between the two GPUs while normal SLI on SNB would only allow x8 communication? But the NF400 will also add latency. 2x16 is better then x16+NF400 that is better than 2x8. I happily admit that I am no expert PCIe bridge chips. I do agree with you that CPU performance will continue to be the important thing with complex aircraft. And I don't think IVB will add much over SNB at a clock for clock comparison in FSX. I cant wait to se proper results in how much higher (if any) you can actually clock IVB using "non-exotic" cooling. Tri-gate transistors yes, but also the first production of them. Lower power consumption at the same clockspeed yes. But higher max overclock?I havent tested PCIe 3.0. Just PCIe 2.0 with and without BP=0 at several buswidths. That made me believe that PCIe 2.0 + BP=0 will perform very similar to PCIe 3.0 without BP=0.

I was reading in this form in a thread that mentioned that going from PCI-E 8x to PCI-E 16x resulted in an % performance increase with lots of autogen. I never got an % reduction in frame rate going down to PCI-E 8x from PCI-E 16x, I only lost 1 fps. Did the person mean the loading of autogen was faster but not the frame rate when the PCI-E bandwidth was increased? Maybe I didn't have the ideal scenario and settings to make it a difference, I wasn't moving the aircraft while testing and just sitting on the runway?

At times when you are fully limited by the CPU you wont see any benefit by a better GPU or better PCIe bandwidth. When limited by the PCIe bus you would get lower FPS and a possibly stutters. Sitting on the ground would really restrict your view and the amount of autogen that you would see. So no surprise that there is no difference there. Try flying at around 1000' over a dense forest with full autogen. Then try that with your second card in once you get that back.Thanks for reporting your experience with and without SLI on the GTX580. What resolution do you use? I believe a single GTX580 won't restrict you using a single display in FSX.

Does it not enable x16 comunication between the two GPUs while normal SLI on SNB would only allow x8 communication? But the NF400 will also add latency. 2x16 is better then x16+NF400 that is better than 2x8.

With a regular SLI in LGA1155/LGA1156 you get 2x8 2.0 (16 lanes split across two slots)With a GTX590 in a single slot you get 1x16 PCIe 2.0 lanes, so it's esentially the same thing.Even in Tri-SLI boards with the NF200 chip, the PCIe BW is limited to what the controller in the CPU is capable of , which is 16 lanes at 2.0 speeds. The NF200 works multiplexing the 3 GPU's into those 16 lanes because there's no 16x board capable of running 3 GPU's without them dropping to 8x4x4x for a total of 16 lanes. Then the NF200 allows for 32 PCIe 1.0 lanes to balance the load across all three GPU's.For dual GPU setups, be it 2 way SLI or dual GPU's like the GTX590/GTX290/HD6990, the NF200 just adds latency (not in a way that it will really effect performance anyway) and negates PCIe 3.0 operability, just like switches do.You don't need multiplexing when you can get 8x8x operation, it's just for 3x GPU configs that it's needed

Ok. Thanks Dario. Did a bit more reading up on SLI and muxes as well.The GPU to GPU communication is through the SLI bridge. And in the rare cases that you dont have an SLI bridge fitted the communication would go GPU1->CPU(PCIe controller)->GPU2. Not GPU1-> GPU2So the only thing the NF200 is doing is making it possible to put more than one PCIe user on one PCIe output by braking up the incoming PCIe 2.0 lanes in to 32 PCIe 1.0 lanes (half the bandwidth of PCIe 2.0) that can be split between 4 users.The s1156/s1155 CPUs have two PCIe outputs. So if you want more than one user off one output (dual GPU card) you need a PCIe bridge.So two separate cards in SLI should not add the latency of the NF200 and that should be slightly better. And still provide the same PCIe bandwidth to each GPU.Or 2x16lanes is better than 2x8lanes that is better than 16lanes+NF200

I don't have the technical knowledge of alot of the guys posting in this thread, but from my experience with my GTX 590, I find I that experience a 25-30% increase in performance on the FSMark11 benchmark when I disable one of my GPU's through the driver. I personally think that's a solid vote against running SLI.

Not to sidetrack the conversations but I'm also hoping Nvidia gets a clue and includes 3 monitor support in the upcoming Kepler/680.ATI has had 3 monitor support on their top and even mid level cards for more than 2.5 years now and it's getting kind of ridiculous that we still need to do SLI just for 3 monitor support on Nvidia.

I wish they would go this way too, I would love to get back those frames that I lose from running SLI.