AMD shipped its original Radeon HD 7970 in December 2011, and for a brief period, this 4.3-billion-transistor monster held the graphics performance crown. Several months later, Nvidia launched the GTX 680, a power-efficient graphics card that retook the slot as the best single GPU card for gaming.

In some ways, though, the GTX 680 is an odd beast. Nvidia has been reluctant to supply what's called GPU compute performance, and it’s been suggested that Nvidia deemphasized the compute performance in search of gaming performance glory. The latest GTX 680 drivers don’t support OpenCL 1.2 (more on that later), though you can still run Nvidia’s own CUDA framework for GPU compute.

That’s proved a good product strategy, but now we’re starting to see a new generation of games that integrate GPU compute features into the game, either using DirectCompute or OpenCL. We’ll take a look at a couple of these games later. For now, let’s run down the key differences between the original and the new card.

AMD claims the core clock boost from 925MHz to 1GHz is due to better understanding of the manufacturing process and a few tweaks to that process. The reference card has a sticker on the back labeled “Rev 02”, though that probably refers to the card as a whole. Since a number of AMD partners already ship cards running in excess of 1GHz, officially supporting that clock speed isn’t a surprise.

AMD is also building the new version with faster memory—the same 6GHz effective memory that Nvidia uses on the GTX 680. But perhaps most interesting is the new “boost clock”—something that Nvidia implemented in the GTX 680. AMD tweaked its PowerTune technology, which manages the power envelope of the card, to allow for slightly higher core clock speeds if power demands aren’t too high.

It’s pretty simple: AMD simply added new P-states to PowerTune to allow for higher clock rates for short periods of time, when needed.

All our performance and power tests were run on our standard graphics test system:

The system was not overclocked. A Watts Up Pro USB power meter collected the system power data. All the game tests are reported at 2560 x 1600 pixel resolution with 4x multisampling antialiasing enabled.

The architecture hasn’t changed, and the AMD reference card looks nearly identical, right down to the display connectors and fan. However, PowerTune Boost, the higher core clock speed, and the faster memory does come at a cost—while idle power is actually a little better than the original, at 120W (versus 124W for the first Radeon HD 7970), full load power is a whopping 384W, over 18 percent higher than the 324W maximum power draw on the original HD 7970. The GTX is more power-efficient than either card, at 119W idle and 318W under load. Note that the HD 7970 GHz Edition still uses the same power connectors as the first 7970—one 8-pin PCIe power connector and one 6-pin connector. The new AMD progeny is the same 11-inch length as the original; both are a half-inch longer than the GTX 680.

These are power ratings for a system with a 1200W PSU.

What this also means is that the card may be louder, depending on the cooling system design. Certainly the reference GHz Edition card is noisier under full load than the original HD 7970.

Given the right case and cooling system, though, noise can be mitigated somewhat. How does the card peform? We ran through a number of our standard synthetic and game benchmarks, plus a pair of new game tests, which make heavier use of GPU compute.

Nvidia’s GTX 680 has an impressive tessellation engine, which is reflected in the 3DMark 2011 score. That’s the one benchmark that demonstrated a significant advantage for the GTX 680. The 3DMark Vantage test uses DirectX 10, so it’s a closer race in that benchmark.

But it is game performance that’s important here, so let’s see how our suite of games runs on the three GPUs. First, we’ll look at the older tests, which aren’t as GPU-compute-heavy.

The original 7970 has a slight edge over the GTX 680 in Crysis 2. The GHz Edition Radeon opens up a bigger lead.

The GTX 680 had a strong lead over the original HD 7970, but the GHz Edition 7970 wins out.

In Metro 2033, the two Radeons have little difference, but the GTX 680 seems to lag in this game test.

The faster memory, wider memory bus, and higher core clocks give the new GHz Edition 7970 a big boost in DiRT 3.

The newer Radeon performs noticeably better than the original, and outpaces the GTX 680 as well. So the higher core clock and faster memory (combined with the 384-bit memory interface) help substantially.

Now let’s turn to a couple of newer game titles that make heavier use of GPU compute. Sniper Elite V2 uses supersampling plus DirectCompute accelerated antialiasing. DirectCompute is also used for high-dynamic-range ambient occlusion shadow effects and depth-of-field calculations. DiRT Showdown is the first game shipped to use global illumination, plus other advanced lighting techniques.

Here, the differences are more dramatic. AMD’s focus on building a GPU that performs well in both graphics and GPU compute seems to have paid off. Note that we didn’t test GPU compute heavily, but it’s worth noting that the latest version of Winzip, Winzip 16.5, offers an option to enable OpenCL GPU acceleration to compress files. It does make a difference—an 8GB folder took about 10 percent less time. However, this option doesn’t appear when a GTX 680 is installed, even with the latest 304.48 beta drivers.

Competition is great for consumers. AMD will be shipping the Radeon HD 7970 GHz Edition in the next couple of weeks, at roughly $499—$100 cheaper than the original and in the same price ballpark as the GTX 680. On a purely performance basis, the new Radeon wins out over the GTX 680, albeit with a hefty power hit. However, the GTX 680 is quieter, consumes less power, and is more compact. That may affect your choice. But if you’re looking for the fastest single GPU card available, AMD seems to have retaken the crown.