CrossFire internalizes, goes native
Nearly a year since the debut of its CrossFire multi-GPU scheme, the red team has finally integrated CrossFire's image transfer and compositing capabilities directly into a graphics processor. Before now, the high-performance implementations of CrossFire required the use of a specialized CrossFire Edition video card that came with an FPGA chip onboard to handle image compositing. Getting data from the other Radeon card to the CrossFire Edition required the use of an external dongle cable that hung out of the back of the PC like a hemorrhoid. Thanks to Preparation RV570, that's no longer necessary.
Any Radeon X1950 Pro can talk to a peer via the two "golden fingers" connectors on the top of the card.
Look familiar? These are similar to Nvidia's SLI connections, but they're neither physically nor electrically compatible with an SLI bridge. (The golden fingers thingy is wider, for one.)
Here's a look at SLI and CrossFire bridges side by side. Right now, SLI bridges come with SLI-ready motherboards, but ATI has a different plan. They will provide one CrossFire bridge with each Radeon X1950 Pro, so users with older motherboards won't be stuck without one.
Oddly enough, native CrossFire requires two connections between cards in order to work properly. Each of these connections is a 12-bit link, and native CrossFire will scale up to 2650x1600 at 60Hz via a dual-link arrangement. (I'm fairly certain SLI scales to that same resolution using just a single bridge connector.) ATI says CrossFire could work with just one of these two connectors between cards, but that its graphics drivers currently enforce a dual-link config. In fact, when first setting up our Radeon X1950 Pro CrossFire test config, I attached just one connector, and the system refused to go into CrossFire mode.
The presence of two CrossFire bridges naturally raises some suspicions in the curious mind. This is an all-new thing for ATI, and surely they've put some thought into it. Why complicate a dual-card setup unnecessarily by requiring the installation of two physical bridges? Could the excess capacity be there for future three- and four-card configurations? The two, staggered connectors per card could work well in a sort of daisy-chain of graphics cards, if need be.
Can't wait to see what they do with this one. Potentially, any model of Radeon with native CrossFire could be put into a serially connected team of three or four cards by interleaving the connectors used. With DirectX 9's three buffer limit (which I explain in my really cool Quad SLI article,) a triple-card CrossFire rig might be the sweet spot for performance. Mobos with three PCIe x16 slots are already beginning to show up in various quarters.