So, I keep stating that there is a bandwidth issue with regards to SLI and current generation cards. What the issue is, is that a lot of tech requires a lot more bandwidth between cards than is currently available to work properly. Cards transfer data between the PCI/e bus and the bridge (if one is present). The standard SLI bridge is clocked at 400MHz and grants approximately 1GB/s of bandwidth between GPUs + bandwidth from the PCI/e connection. This is not enough for the amount of data in each frame (frame buffer data), or relevant memory between cards, to be transferred between frames for new tech, or technologies such as Temporal anti-aliasing in general (though we've seen TXAA and in some cases, albeit at a large performance hit, TAA function in SLI). It has gotten to the point where forcing SLI in certain recent unsupported games/game engines or at certain resolutions (usually, but not always, above 1080p) literally requires PCI/e 3.0 x16/x16 for the cards in SLI to even grant positive scaling. Yes, that means that negative scaling (less FPS than single GPU) happens otherwise. This means your CPU must cost a minimum of $500 USD for the intel mid-tier enthusiast CPU of any generation (4930K, 5930K, etc) and AMD CPUs which only use PCI/e 2.0 lanes cannot even be considered.
AMD offered a solution; one that I consider to be the best possible option. They decided to make use of the PCI/e bus bandwidth to talk to each card by configuring the memory in a certain way, which they call XDMA (Crossfire Direct Memory Access). This introduces a latency issue, but it's one that even AMD has found ways to minimize/eliminate, and I am certain nVidia could do the same if they cared (they apparently do not). In a PCI/e 3.0 x8/x8 situation (what a user is most likely to be using with a mainstream intel CPU as is the most popular PC configuration for gamers) a user would have 7.88GB/s bandwidth in this way, which is far more than the 900MB/s or so that the crossfire bridge granted previously, and far more than the 1GB/s that the SLI bridge grants. It requires the memory on the card to be designed in a certain way, however, so it cannot be retroactively added as a "technology" to any GPU. Also, as an added benefit, XDMA also helps to fix framepacing in multi-GPU, making it more smooth to use.
nVidia also made a solution to improve interconnectivity between cards... NVLink. However because NVLink replaces PCI/e, it is only sold on enterprise boards intended for use with Tesla cards. Supercomputers, to simplify. You are simply not getting a pair of gaming cards into one. Since it is a replacement for PCI/e, and apparently requires proprietary card connections, and it inevitably costs more money to have it, it would never work for mass-produced consumer boards. Whether this requires XDMA-style memory configurations or not, I do not know. But either way, it can't be sold to people like me.
So, recognizing the need for more bandwidth, but not being able to sell us simple XDMA-style cards for more $$, they decided to simply improve the bridge. There was already an overclocked LED bridge, that runs at 650MHz up from the standard 400MHz, granting 62.5% more bandwidth for 1.625GB/s. nVidia simply doubled up the connectors to use both SLI fingers and presto! 3.25GB/s bandwidth between cards! This is a huge step up from 1GB/s, however nowhere near the 7.88GB/s to 15.76GB/s that is present in XDMA-style x8/x8 or x16/x16 configurations respectively. Also, this effectively killed three-way and four-way SLI. I thoroughly dislike this approach that is being made, and it further reinforces my belief that SLI is basically being discarded by devs and nVidia alike. This is nothing but a band-aid on the bandwidth issue, and developers have zero reason to be coding AFR-unfriendly rendering techniques either. I hope things make a turn for the better in the future, but as of right now, things are exceedingly bleak. It also does not help that they have no competition and only need do the bare minimum that is necessary to get the population to love them, and most larger voices in tech and the majority of PC gamers don't care about SLI.