Enter XDMA, exit connectors, all needed data is now pulled over the PCIe bus via DMA. A requesting card can reach out to the required target and essentially grab, or at least ask nicely, for what it needs. This also allows for an unlimited number of cards in Crossfire but we forgot to ask how many. In any case eight cards should be doable if something else doesn’t bottleneck things first.
What is the XDMA engine? It is really just a display controller connected to the display engine so it can be timing aware but unlike the older way it isn’t timing dependent. This is a subtle but key difference. The card with the monitor connected is now able to prefetch the data it needs long before it needs it. The old Crossfire basically requested when it was needed, not before, and if anything burped things got ugly. The new XDMA version has much more flexible algorithms to both predict and fetch what it needs long before it needs the data. Storage is also larger and more flexible, again based on algorithmic control rather than fixed.
In short as soon as a render surface is enabled on the displaying card the data can be requested from the source. The storage space is also guaranteed to be adequate as well and nothing is just in time delivery so minor bumps in the night don’t turn in to missed frames. This should not only allow more scaling for card count, display counts, and screen sizes, but also higher reliability too. It will be interesting to see if this opens up multi-card, multi-monitor Crossfire solutions, with a bit of clever software hacks, multiple cards each with connected monitors is at least theoretically possible. Hmmm, Pi.
