Maximum PC

So in a month or two I'll be sitting down to do my bi yearly pc rebuild.
This time round I'm going to be delving again into dual video cards, last time i did that was back with a pair of voodoos early 2000.

My question is this, looking at mobo's i see a variety of runs @ 16x single, dual @ 8x etc.

What are you loosing by having this occur?
Should I focus on a mobo that runs both slots at full 16? Is this a trivial matter for them to halve when a second card is added? What about a third card dropping further down to 4x mode?

I imagine this is trivial in that I still have the bandwidth of 16x, just its split between the cards.

I plan to farm bitcoins when not gaming(mmo/fps). Would this affect that performance at all?

sli and crossfire will perform about the same in 16x or 8x, there are motherboards that have more 16x but you have to pay for them, as they are for the most part, expensive. but nvidia and amd both have dual GPUs on a single card. nvidia has several of them, the 590 is dual 580s, there is the 560ti 2 win card which is dual 560s, and a 460 2 win which is dual 460s.

amd has the 6990, 6870x2 are the newer ones. then the slot speed for sli xfire will not matter, since only a single slot is going to be used. and you could add a third if you wanted to. although it is not much of an improvement over two cards for gaming anyway.

it would have no effect on bitcoin farming, it uses the card not the bus. although I have heard it costs more for electricity to run the card or cards flat out than you get back in bitcoins.

also the SLI/xFire bridges really negate the bandwidth needed to cross talk between the cards. You only loose a 1-3% performance drop from going from 16x to 8x if you are running two 580's at 2560x1600 resolution with everything maxed. A third card only really adds about a 14% increase in overall performance and a 4th adds 8-10%. Really makes no sense in running triple SLI this is why I put my third 570 in a different computer. heh.

A few years ago, Tom's Hardware published an article experimenting how performance is related to PCI-Express lanes available. It was old (using PCI-Express 1.0), but basically they found out that graphics cards don't start having a big hit until about 4x. It wasn't that bad, but it was noticeable. 2x and 1x was when performance caved in.

Long story short, it won't do anything appreciable if you go from x8 to x16. Unless you're trying to run a 3DMark record or something.

And I think the bridge is just there so the PCI-Express bus doesn't have to handle frame buffer transfers and card coordination.

What all of this really boils down to is that modern cards aren't capable of saturating a 16 lane bus @ 2.x speeds. They are capable of clogging an x8 bus, but consider that when a card is communicating in this mode, the workload is usually being halved between two cards so any performance difference is negotiable.

This may all change soon. AMD's new 7000 series Radeons will be based on a new architecture, fabricated on a 28nm process, and will be the company's first "from the ground up" GPU re-design in years. The possible inclusion of XDR2 for VRAM purposes, coupled with GPU arrays in excess of 2000 stream processors, could theoretically offer enough throughput that the plumbing on an x16 2.x bus would be clogged up with FPS goodness. Then again, the 3.0 spec is right around the corner for mass adoption, so those like yourself who are looking to adopt the newest tech in a couple of months won't have anything to worry about.

The performance level of the gpu's is not directly proportional to the level of bandwidth needed to communicate between the gpu's and the cpu(s), system ram and hdd's. So even if you have two cards that are "sharing the load", you are actually increasing the amount of traffic on the pcie lanes in folds, but it really depends on a number of variables such as the software and what information is being shared between these resources. This mostly effects how well a system is scalable when adding a 2nd, 3rd or even 4th video card.

With SLI and xFire, the cross talk between gpu's are handled by the bridges obviously since pcie lanes can't handle the satuaration from high end cards. Although you can do SLI of low end cards without a bridge since you arne't pumping nearly the same amount of information.

Nobody suggested otherwise.



I think what you are trying to say is that SLI and CrossfireX are solutions that allow for workloads to me more evenly be distributed across processing platforms, but it doesn't necessarily increase the amount of traffic. Having a multi-GPU solution in place does not automatically increase the systems ability to keep said GPU's supplied with data.



J, what does this mean?



When you run a multi-GPU solution the card mounted in the primary slot is the master card, and any secondary cards are considered slaves (much like data drives during the golden age of computing). When data is sent to the graphics processing array, the load hits the primary card first. Part of the package is offloaded to supplementary cards (how much depends on the solution being run, and how drivers are designed to dedicate the load). Data bridges are common on high end solutions to supplement available bandwidth on systems that run a 2x8 PCIe config. Today's standards probably don't dictate the need for a bridge solution when a 2x16 lane bus is available in concert with a chipset that could actually handle it (that is important too), but thankfully hardware and driver solutions are not designed with top of the line hardware in mind.

16 lanes is considerable bandwidth. For the foreseeable future I would imagine that most of the energy dedicated to improvement of the bus will be centered on speed and not width.

But honestly, it all depends on the software and what information is being processed between the cpu and gpu. Some applications don't need much info to be feed from the cpu, which we see more and more with some specialized gpgpu apps.