NVIDIA NVLink
With Turing, NVIDIA retires the old SLI connector, which used a now-outdated connection protocol that was for point-to-point connections only, and also one that only transferred rendered frame data to the master card in the SLI configuration.
We have seen NVIDIA's NVLink connector before on the Titan V, as well as on Tesla and Quadro cards. It is built around a communication protocol that is similar to PCI-Express, with support for mesh networking for easier communication among multiple devices. Turing TU102 (RTX 2080 Ti) comes with two x8 NVLink 2.0 links, whereas TU104 (RTX 2080) has a single x8 link to help further differentiate the product stack. Along that same thought process, TU106 (RTX 2070) does not support SLI and thus, there are no NVLink connectors present.
Compared to previous-generation SLI, NVLink-based SLI offers a significant bandwidth advantage, which brings a peak bandwidth of 25 Gb/s per x8 link, per direction. As such, the RTX 2080 Ti has a combined bi-directional peak bandwidth of 100 GB/s, and the RTX 2080 has a similar peak bandwidth of 50 Gb/s. This increased bandwidth offers more display support than ever before, enabling SLI with 8K Surround on the RTX 2080 Ti, whereas the RTX 2080 supports single display 8K@60Hz or 4K@144Hz surround over SLI. Note that the actual driver enabling said support is slated to drop post-launch, however.
Another important change with NVLink SLI is that each GPU can now access the other's memory in a cache-coherent way, which lets them combine framebuffer sizes, something that was not possible with SLI before. The underlying reason is that the old SLI link was used to only transfer the final rendered frames to the master GPU, which would then combine them with its own frame data and then output the combined image on the monitor. In framebuffer-combined mode, each GPU will automatically route memory requests to the correct card no matter which GPU is asking for which chunk of memory.
Multi-GPU shared memory comes off as a tremendous promise, but I doubt it will have any significant impact for games. Market research has shown that multi-GPU configurations are extremely rare and thus, developers will be hard pressed to invest substantial time and money to support this tiny subset of the market. For the professional space, this could be a big thing, though, as scenes of much higher complexity can be rendered from VRAM (up to 96 GB with 2x Quadro RTX 8000).
The introduction of NVLink-based SLI to the consumer space means that existing SLI bridges will not work on TU102, TU104, and TU106-based GeForce RTX cards. You will thus have to buy new NVLink bridges from the NVIDIA store at $79 each. These bridges are available to support motherboards with 3-slot or 4-slot spacing between cards at this time, and presumably, we will see board partners come up with their own take on the NVLink bridge sooner than later.