Single page Print

Additional goodness: TrueAudio, displays, and XDMA
AMD has built several new technologies into its Hawaii chip, and some are complex enough I can't do them justice in the time I have to finish this review. Only two of AMD's current graphics chips, Hawaii and Bonaire, have these next-level capabilities built in. I suspect we may see more GPUs from this same family in the coming months.

The most notable of the new features is probably the TrueAudio DSP block for accelerated processing of sound effects. There's much to be said on this subject, and I intend to address TrueAudio in more detail in a separate article shortly. For now, you might want to check out my live blog from the GPU14 event for some additional details on this feature. We don't yet have any software to take advantage of the TrueAudio hardware, but I suspect we'll spend quite a bit of time with TrueAudio once the first games that support it arrive.

AMD has also freshened up the display block in its latest GPUs. You can see the connector payload above. Both of the DVI ports are dual-link, and the DisplayPort output is omni-capable: it supports multi-stream transport (MST) and can sustain the pixel rates needed to drive a single-tile 4K display at 60Hz, once such mythical beasts become available. Furthermore, Radeons will support the DisplayID 1.3 standard, written by an AMD engineer, that allows for auto-configuration of tiled 4K displays—provided those displays also support this standard. The current 4K monitors from Sharp and Asus do not, but AMD intends to recognize those panels and take care of them automagically in its drivers.

Perhaps the biggest change here is the elimination of the requirement that multi-monitor Eyefinity configs include at least one DisplayPort connection. With the 260X and 290X, users can finally connect three monitors via the HDMI and DVI links alone. Huzzah.

You may have noticed the distinct lack of CrossFire connectors on the 290X. There are, uh, vestiges where the "golden fingers" connectors ought to be, but no actual fingers. That's because AMD has replaced the CrossFire bridge connector with a new solution called XDMA. Rather than pass data from GPU to GPU over a dedicated bridge, XDMA incorporates a direct memory access (DMA) engine into the CrossFire image compositing block. This DMA facility can transfer data directly from GPU to GPU via PCI Express, without a detour into system memory.

XDMA is purportedly compatible with AMD's frame pacing tech, which reduces the micro-stuttering problems associated with multi-GPU teaming. Even more importantly, the firm claims XDMA can handle resolutions above four megapixels, including Eyefinity multi-display configs and 4K monitors. Since current CrossFire configs have serious problems with such setups, this new data sharing method could bring a very notable improvement.

AMD insists XDMA carries no performance penalty compared to a dedicated CrossFire bridge. They make a good argument when they point out that the situation can't get much worse than it is now with CrossFire at resolutions above four megapixels. The Radeon driver software shifts frame data from the secondary GPU into system memory and then to the primary GPU. The end result? At 4K resolutions, the transfers are too slow, and pretty much every other frame is dropped completely, never to be displayed. Quicker, more direct GPU-to-GPU data transfers can only help.

The firm is confident that lower-bandwidth CrossFire configs aren't any worse off without the dedicated bridge, either. In fact, they wanted to be sure before deciding to go with XDMA as their only solution, which is why 290X boards retain those phantom fingers. Early boards included bridge connectors for comparative testing. Once AMD was convinced the solution was solid, the fingers were, uh, snipped off.

So how well does XDMA work? We're dying to try it, and we've asked AMD for a second 290X card explicitly for the purpose of testing CrossFire with XDMA at 4K resolutions, but we don't yet have a second card. We're hoping to rectify that problem shortly.

We have lots of questions about what sort of PCI Express configurations will prove to be suitable for high-resolution CrossFire configs. XDMA seems well-suited for systems based on Intel's X79 chipset, with 16 lanes of PCIe 3.0 bandwidth running to two expansions slots, or for dual-GPU cards like the Radeon HD 7990 with PCIe switch chips oboard. Those are notable configs for high-end CrossFire setups. But will XMDA play well in systems with less bandwidth, like those based on Haswell or Richland processors with dual x8 PCIe links? Or systems with higher PCIe latency, like AMD's 990FX platform? What happens with three- and four-card setups? We'll have to push the limits in order to find out.