Single page Print

Bending the envelope
When it came time to choose the 5970's clock speeds, AMD encountered a no doubt vexing tradeoff. In order to stay within the 300W power envelope dictated by the PCIe specification, they had to dial back clock speeds considerably from Radeon HD 5870 levels. (That's probably why this beast isn't called the Radeon HD 5870 X2.) In fact, the clock speeds they chose are right at Radeon HD 5850 levels: 725MHz for the GPU, and 1GHz (or 4GT/s) for the GDDR5 memory. Here's a look at how the 5970 stacks up, in peak theoretical performance.

fill rate
Peak bilinear
Peak bilinear
FP16 texel
Peak shader
arithmetic (GFLOPS)
Single-issue Dual-issue
GeForce GTX 285 21.4 53.6 26.8 166.4 744 1116
GeForce GTX 295 32.3 92.2 46.1 223.9 1192 1788
GeForce GTX 285 SLI 42.9 107.2 53.6 332.8 1488 2232
Radeon HD 4870 X2 24.0 60.0 30.0 230.4 2400 -
Radeon HD 5850 23.2 52.2 26.1 128.0 2088 -
Radeon HD 5870 27.2 68.0 34.0 153.6 2720 -
Radeon HD 5970 46.4 116.0 58.0 256.0 4640 -

These days, I feel like we should attach as many caveats to that table as one might to a federal budget resolution. These are simply theoretical peak values, and they are in some cases quite academic. You're rarely going to hit the peak pixel fill rate on one of these cards, for instance, since memory bandwidth will likely limit you first. The GeForces probably won't ever reach their peak dual-issue shader FLOPS numbers, and I doubt the Radeons will achieve more than 80% of their peak compute capacity when executing pixel shaders. On top of all that, we've just thrown in the numbers for the multi-GPU solutions, even though they'll rarely scale linearly.

With that said, you can see above what AMD's compromise on clock speeds has wrought. The 5970 is still undoubtedly the most capable single-card solution around, but it's closer to two 5850s than it is to two 5870s. (The GPUs on the 5970 don't have any functional units disabled, so the 5970 should be slightly faster than a pair of 5850s in terms of shader and texturing capacity.)

The more dramatic comparison, though, is the 5970 versus two GeForce GTX 285 cards in SLI. The 5970 has more than double the peak shader capacity, but it's pretty similar in terms of texture fill rates—and the GTX 285 SLI config has substantially higher memory bandwidth. Thanks to its 256-bit memory interface, Cypress was already balanced pretty heavily in favor of shader power rather than memory bandwidth. Doubling up on Cypress chips amplifies that fact, especially at these clock speeds, where the 5970 trails some competing solutions. The implications of these numbers should be fairly straightforward: in cases where performance is bound primarily by memory bandwidth or texture filtering, the 5970 may not look as impressive as one might have hoped.

The thing is, one could argue that the 300W PCIe spec for video cards is more of a speed limit than a credit limit. If you go over it, bad things might happen, but the system won't necessarily stop working immediately. In fact, you might just end up having more fun. Honest, officer. And we know the Cypress GPU and its GDDR5 memory are both good for higher clock speeds, as the Radeon HD 5870 attests.

AMD didn't want to tread formally beyond 300W, so it has decided to do so informally, instead, by enabling users to overclock the 5970 deep into 400W territory. The firm claims it has taken on considerable expense in the design and production of 5970 cards in order to give them ample headroom. The GPUs, it says, are screened for high speeds and low leakage. The DRAMs are rated for 5Gbps operation, just like on the 5870, even though they only operate at 4Gbps by default on this card. The electronics, including the capacitors and voltage regulators, are built to higher standards. And the cooler's "massive" vapor chamber can dissipate up to 400W.

So, you know, it's a Ferrari 599 GTB, and the local speed limit is 55 MPH. Wink, wink.

The 5970 laid bare

These Hynix GDDR5 DRAMs are rated for 5 Gbps operation

To enable its customers' potential delinquency, AMD has raised the caps on its Overdrive overclocking control panel for the 5970, pushing the limits to 1GHz for the GPU and 1.5GHz (or 6 GT/s) for the memory. The company also expects its board partners to offer voltage tweaking tools with their 5970 products, and it supplied us with a rudimentary example of such a tool.

Those sliders promise granularity, but it's a mirage. You get two settings here: default and peak. Fortunately, that's pretty much all you'll need to make good on what AMD strongly hints is the 5970's potential: the same clock speeds as the Radeon HD 5870, or 850MHz/1.2GHz GPU and DRAM, respectively.

We didn't get very far at all in our overclocking attempts without the voltage tweak, but at the higher voltages, we hit 5870 levels with very little drama. In fact, we've tested at those speeds and included full results in this article.

Do remember a couple of things, though. First, you've got to overclock both GPUs individually in the control panel in order to see any real performance gain. Second, although the clock speed settings in the Overdrive tool will persist after a reboot, the higher voltage settings did not, in our experience. This combination made for some interesting times, let me tell you.

In fact, we had so much fun sorting out that issue that we decided against pushing the 5970's clocks beyond 5870 levels. We'll leave that fun up to you.

Regardless of how easy AMD has decided to make it, this is still real overclocking, with no guarantees about likely clock speeds or what bad things may happen when you exceed stock settings. You'll have to engage in real overclocking to get a 5970 that runs at these speeds, too, because AMD plans to limit the fake "overclocking" conducted by board vendors, the sort where they set higher default speeds and back up the cards with full warranties. AMD even says it has a mechanism in place to cap default clock speeds, and it will prevent board vendors from pushing too far past the 300W limit. That may change eventually; plans are afoot to enable dual 8-pin power connectors on future 5970 cards, but don't expect to see such an animal until next year.