Suffice it to say that the adaptation of this concept to feature the new RV770 GPUs needed a few slight changes in the bus interface, including the internal interconnects to the latest PCIe 2.0 standard to provide enough data bandwidth for the combined 1600 Stream processors. In a nutshell, the current solution is to have one primary GPU aided by one auxiliary GPU (or linked GPU) where the link features a bridge chip supplying each GPU with 16 lanes of second generation PCIe and interfaces with the system through another 16 lanes. This way, total bandwidth to each link is 5GB/s in each direction. In case this is not enough, AMD relies on the same principle used on the SMP K7 – K10 processors, namely a sideport with the only difference that the protocol used here is not HyperTransport but PCIe 2.0. Suffice it to say that the adaptation of this concept to feature the new RV770 GPUs also needed a few slight changes in the bus interface, including the internal interconnects to the latest PCIe 2.0 standard to provide enough data bandwidth for the combined 1600 Stream processors. In a nutshell, the current solution is to have one primary GPU aided by one auxiliary GPU (or linked GPU) where the link features a bridge chip supplying each GPU with 16 lanes of second generation PCIe and interfaces with the system through another 16 lanes. This way, total bandwidth to each link is 5GB/s in each direction. In case this is not enough, AMD relies on the same principle used on the SMP K7 – K10 processors, namely a sideport with the only difference that the protocol used here is not HyperTransport but PCIe 2.0.

The reason is rather obvious, since PCIe is required anyway for the connection to the system, the addition of another interface would be somewhat inefficient. Since each sideport is the equivalent of a 16 lanes PCIe interface sans arbitration (the connection is hardwired for a strict point to point topology), this adds another 5GB/s data bandwidth in each direction. If this is still not enough, the Crossfire Bridge Interface (CFBI) provides the dot on the i with an additional 0.9GB/sec bandwidth available for crosstalk between the two GPUs and/or a sister card in a Crossfire X system configuration.

A slight deviation from the PCIe protocol in its pure form is the support of multiple device selects by the bridge chip, which allows the broadcasting of writes to both GPUs instead of sequential writes with interleaved device selection.

Otherwise, the internal data management has not changed, the individual memory domains are still exclusive, that means, one GPU cannot access the frame buffer of the other GPU. The workaround is simple, albeit a bit costly, each card has no less than 1 GB of local memory available, using, who would have thought, nothing but the latest generation of DRAM technology in the shape and form of GDDR5 running at 900 MHz clock and 3600 MHz data rate.

By now, essentially every manufacturing partner of ATI/AMD has launched their RADEON 4870 X2, currently all cards are still manufactured by AMD and distributed to the partners. Today’s sample was provided by Diamond MultiMedia, a name that brings back fond memories of the Monster Fusion we reviewed almost exactly 10 years ago.

Tech Specs

• • Graphics Engine: Dual ATi Radeon HD4870 (RV770)
• • Interface: PCI-Express 2.0
• • Engine Clock: 750MHz
• • Memory Data Rate: 3.6GHz
• • Memory Type/Size: GDDR5/2GB
• • Memory Interface: 2 x 256-bit
• • Max Resolution: 2560x1600
• • VGA Output: Yes with adapter
• • TV-Out: HDTV/S-Video/Composite Out
• • Built in HDMI and 7.1 surround audio
• • ATi AVIVO HD video and display technology
• • Unified Video Decoder (UVD) for Blue-Ray and HD Video
• • DVI: two ports
• • Native CrossfireX support