[citation][nom]blazorthon[/nom]CF doesn't inherently scale better than SLI. How well it scales depends on the GPUs. For example, VLIW5 GPUs tend to scale slightly worse in CF than Fermi GPUs scale in SLI and the VLIW5 GPUs tend to be much more stutter-prone. VLIW4 GPUs and GCN GPUs tend to scale much better than Fermi and have less stutter. CF does tend to scale down to the weakest link except in asynchronous CF modes such as one or two discrete cards in CF with an AMD IGP, so yes, CF can be considered a little inferior to Lucid's multi-GPU technology, but CF tends to scale a little better and if situations where the weakest link isn't much weaker than the strongest; CF can beat Lucid somewhat in many situations. SLI on Kepler is also very well-scaling, although it doesn't seem as friendly with three GPU setups as CF is with GCN. About where I've read all of this? Well, it pops up in many different places. When I want to learn something, I'll often spend quite a bit of time looking into it. Crossfire doesn't need to be software, it is also hardware, but software CF can do so much more than hardware CF. Hardware CF, like hardware SLI, must be supported by a motherboard and the cards for it to work, whereas software CF can work on pretty much any motherboard and a wider variety of cards and in a wider variety of ways.[/citation]
Thank you for all those clarifications and specific case details. So what you're saying is that while SLI only has a hardware less flexible hardware implementation, CF has both a hardware and a more flexible software implementation? If this is the case, does the hardware implementation of CF have any advantages over the software version? Is there a way to choose between hardware and software implementations with CF if both are possible? When you say "software," you mean it's managed by the drivers? Is this the reason why mobos support CF but not SLI? I'm thinking that last one might be more of chipset thing. Thanks!
[citation][nom]blazorthon[/nom]If you have boards with great enough PCIe bandwidth, the bridges are still not always necessary. There are many factors for this. For example, the bridge only needs to be able to transmit data equal to the total amount of displays to and from the card that it is on between each card in the system. For example, a CF system with three 7970s and one 1080p display attached to each card would make a 5760x1080 Eyefinity configuration and each card's bridges need to be able to send in at least enough bandwidth for the one 1080p display that they have for frames rendered by the other two cards and enough bandwidth to send out data to the other two displays for frames rendered by that card. So, each CF dongle needs to have at least enough bandwidth for one 1080p display in this example because you would ideally use two dongles for three cards and each dongle only needs to be able to transmit a single 1080p frame at a time. Assuming V-Sync of some sort is enabled and frames are limited to 60FPS and that we're using 32 bit color (24 bit color + 8 alpha), that's just under 500MB/s. So, theoretically, if you are using a board that has at least 32 PCIe 3.0 lanes and each card is getting at least 8 PCIe 3.0 lanes, you could probably not use the bridges and still be just fine. If you had only two 7970s in CF and still used this resolution and board, then the 16 PCIe 3.0 lanes per card would almost be enough, especially considering that with the bridge, even PCIe 3.0 x8 is enough. Even with a 120Hz 2560x1600 display (needing a maximum of about 2GB/s in these circumstances) should be far within practical limits. Theoretically, Nvidia could do something similar with SLI, but I don't think that current high end Nvidia cards can do this and assuming that they can't, I don't think that Nvidia cares to add such functionality. However, I'll not pretend that ALI never had advantages. It didn't always scale better, but it did often have less micro-stutter than a similarly CF setup. However, with VLIW4 and GCN, this advantage waned quickly. You can do even 7970 CF without a bridge. Yes, the 7770 does not need one for all situations. You need to have sufficient PCIe bandwidth and not have a ridiculously high MP and Hz display configuration. For example, you're better off using something like 1080p with very high AA instead of similarly performing 2560x1600 with little to no AA. Basically, you need each card to have enough PCIe bandwidth for all of the displays in addition to its other PCIe bandwidth needs. A motherboard with PLX would be preferable, but not always necessary.For example, consider a Z77 motherboard without PLX and thus having 16 PCIe 3.0 lanes for graphics. The board can run two PCIe x16 cards in x8/x8. The 7750 and 7770 can probably get away with not needing more than 8 lanes of PCIe bandwidth when they don't use a bridge (granted the 7750 obviously doesn't have a bridge as being an option). You probably couldn't do this properly with much higher end cards without more PCIe bandwidth, but I have no doubt that the 7770 is capable of it. I'll see if I can find some benchmarks of it.[/citation]
I thought that when you want multi-monitor setups, or at least Eyefinity setups with CF, you connect all monitors to one card while the other(s) remain free. Or is it more efficient to distribute the monitors so that whole frames from the other cards don't suffer from a bottleneck getting to the card with all the monitors and possibly alleviate the need of CF bridges? Or is that just the way you make an Eyefinity setup with CF? Wait, does one card even handle a whole Eyefinity frame, 5760x1080 for example, on its own or does a frame get divided depending on the monitors attached to each card? I'm guessing the former since the latter sounds more idealistic only when there are equally powerful cards with equal amounts of monitors, 1 or 2 each.
Is that the only purpose of CF bridges nowadays, to transfer sync frames in a way? How much bandwidth does a CF bridge offer?
Also, there's pretty much no harm in using a CF bridge even if your mobo/CPU already has enough PCI-E lanes right? I'm not sure if this is the case with every card that can use a CF bridge, but they come with CF bridges right? I'm guessing not, but do mention if so or not. I was thinking it was a general AMD thing to make cards come with CF bridges, but maybe the two Radeon cards I've had so far from XFX and Gigabyte just happened to come with them.
[citation][nom]blazorthon[/nom]PCIe bandwidth has nothing to do with the SATA ports except for SATA ports on PCIe adapters and even then, it doesn't affect the other PCIe lanes that are intended for graphics cards.[/citation]
Are PCI-E lanes reserved in a way that only graphics cards have access to some/most of them? Oh wait! Or do you mean that when you already have (a) graphics card(s) inserted, the chipset or something assigns lanes like x16, x16/x8, x8/x8/x8, etc for example, for exclusive use of (a) graphics card(s)?
SO MANY QUESTIONS!!! Sorry about that. A lot of thoughts popped in my head.
BTW, on a personal side note, I just remembered an old topic we talked about blaz concerning PCI-E 2.0 and 3.0 cross-compatibility, where you also mentioned a PLX chip as a factor. Would you mind elaborating on PLX chips, PCI-E ones in particular in case there are others? I found a company on Wikipedia named PLX that might have something to do with it as it was described. Anyway, back to the side note. You, blaz, along with tourist had a pretty heated debate around that time and I'm just happy to see that you guys are getting along.
Facebook
Twitter
Instagram