Single page Print

An early peek at the Radeon HD 4870 X2

We take an early R700 engineering sample for a spin
— 11:51 PM on July 13, 2008

AMD has quite a hit on its hands in the Radeon HD 4850 and 4870, and pretty much everyone already knows it's developing another product code-named R700, a high-end graphics card based on two 4870s paired together. We have an early engineering sample of this beast, and we can offer a preliminary look at how the card—to be called the Radeon HD 4870 X2—performs. You might be expecting big things, and you might be right. Keep reading to see what we found.

R700 up close
Although this product's code name, R700, follows a naming convention similar to past high-end Radeon GPUs, it's not really a new GPU at all. Instead, it's just two RV770 graphics processors having a party together on one PCB, pretty much like the Radeon HD 3870 X2 was in the last generation. In fact, the new X2 looks an awful lot like the old one at first glance.

The Radeon HD 4870 X2

We don't have the full scoop on the 4870 X2 at this early date. This is just an engineering sample, not a retail product, and it may be subject to change by the time the boards ship to customers. That's scheduled to happen some time around the middle of next month, which is soon enough that we don't expect major changes to the card between now and then.

The board itself is a healthy 10.5" long, and it sports a couple of PCIe auxiliary power plugs, one 2x4 pin and one 2x3 pin.

Flip it over, and you can see the dual retention brackets for the heatsinks attached to each GPU. Here, if you're a hopeless geek, you might pause to reflect on the 4870 X2's likely specs. The RV770 GPUs on our early sample are clocked at 750MHz, the same speed as the regular 4870. As a result, the X2's 1600 total stream processors have a peak computational rate of 2.4 teraflops. That's, erm, considerable—beyond the obvious graphics applications, that's the sort of computing power that may one day enable men to figure out what women want.

Remove the card's cooling apparatus, and it looks like so:

In the center of the board sits a PLX PCIe switch chip, flanked by a pair of RV770 GPUs. We couldn't find this particular model of PCIe switch listed on PLX's website, but when installed, the card shows up as a PCI Express 2.0 x16 device. Each GPU shows up as being PCIe 2.0 x16 capable in AMD's driver control panel, as well. So our best guess is that we're looking at a PCI Express Gen2 switch that has 48 total lanes—16 routed to each GPU and 16 connected to the PCIe slot.

Rumor has it the R700 may include a faster GPU-to-GPU CrossFire interconnect in order to help with performance scaling in difficult cases, but we don't yet have definitive info on how much of a bandwidth boost it may offer.

We do know, however, that the two GPUs on the card don't share memory. The board has eight Hynix GDDR5 memory chips per graphics processor, four on the front and another four around back. Those chips are 1Gb each, so each GPU has a total of one gigabyte of memory to call its own. Cumulatively, 4870 X2's effective memory size is still 1GB, since data must be replicated into each GPU's memory space.

Although the Hynix chips on our engineering sample are rated for up to 4Gbps operation, on this board, they run at the same 900MHz base clock and 3600MT/s data rate as on the Radeon HD 4870. That ain't exactly shabby, though. All told, the R700 has an aggregate 512-bit path to memory that theoretically peaks at 230GB/s. To put that into perspective, its likely closest competitor, the GeForce GTX 280, has "only" 142GB/s of peak memory bandwidth.

With that in mind, the big question about the 4870 X2 is: How does it perform? If you've answered "About like two Radeon HD 4870s in a CrossFire setup," you're on the right track. Compared to a dual-card config, this puppy has the potential benefit of a faster CrossFire interconnect between the GPUs and twice the effective memory size (single 4870s currently have 512MB), but it has the possible disadvantage of those GPUs having to share PCI Express bandwidth to the rest of the system via that PLX switch. Which, of course, is why we test these things....