Several months late and supposedly only a couple of weeks ahead of Nvidia's own dual-GPU flagship launch, AMD's Radeon HD 6990 has no trouble establishing performance superiority. But does speed at any cost sacrifice too much of the user experience?
In drag racing, they say ‘a chase is a race.’ In other words, if you floor it and the guy next to you follows suit, that’s a race, and you’d better be prepared to pay up at the finish if it’s a money contest.
Both AMD and Nvidia have ridiculous dual-GPU hot rods they’ve been tweaking and tuning for months. Understandably, they want to stay secretive about their respective power plants. But neither one seems willing to mash the pedal and risk an embarrassing second-place finish. It’s a good thing that these two companies don’t live their lives a quarter-mile at a time. I can just see Vin, shaking his head in disappointment.
But come on already, guys! The AMD Radeon HD 6990 was supposed to be a 2010 model, and here we are in March wondering if AMD overpromised during its press briefing last October. We even heard rumors that the 6990 was canceled.
Au contraire, Pierre. It looks like AMD is making the first move with its blown Charger, daring Nvidia to throw-down with a twin sequential turbo-charged Supra...you probably know it as the rumored GeForce GTX 590. We received a single Radeon HD 6990 4 GB one week ago, beta drivers a couple of days later, and updated Catalyst Application Profiles a couple of days after that. Needless to say, the benchmarking marathon that went on in our Bakersfield, CA lab made the 24 Hours of Le Mans look like kart racing at an amusement park.
Meet Radeon HD 6990 4 GB
It just sounds majestic, doesn’t it? 6990. 4 GB. Unlike anything we’ve ever seen from AMD on the desktop. But don’t let naming trickery disarm you like the beautiful rosso corsa of Ferrari’s race cars.
The Radeon HD 6990 follows in the pedigree of Radeon HD 4870 X2 and Radeon HD 5970. It’s a dual-GPU card with graphics processors running, by default, at slightly reduced clock speeds compared to the company’s fastest single-chip board. Its 4 GB of memory are divided between both ASICs. So, you’re essentially looking at two 2 GB configurations on a single PCB, running in CrossFire.
Although it was previously referred to by the code name Antilles, Radeon HD 6990 centers on two of the Cayman-based GPUs found in Radeon HD 6970 and 6950 graphics cards. If you remember from Radeon HD 6970 And 6950 Review: Is Cayman A Gator Or A Crock?, Cayman employs a slightly modified architecture, designed to extract more performance per square millimeter of die space. There are situations where this VLIW4 architecture could underperform AMD's older VLIW5 design, but the company says those situations are rare.
Bottom line: the highest-end Cayman configuration offers fewer ALUs than the most complex Cypress processor (found in the Radeon HD 5800-series cards). However, Cayman’s ALUs are more capable. For a deeper background on Cayman’s architecture, check the second page of our launch coverage.
Each Cayman GPU serves up 1536 ALUs spread across 24 SIMDs. SIMDs are tied to four texture units, totaling 96. Radeon HD 6990 utilizes Cayman in its uncut form, so you get 3072 ALUs and 192 texture units between the pair of GPUs. As mentioned, the 4 GB frame buffer is divided up, 2 GB of GDDR5 per processor, connected via a 256-bit bus.
AMD unifies the two Cayman GPUs using the exact same 48-lane PCI Express 2.0 switch from PLX found on the Radeon HD 5970. Sixteen of those lanes serve the slot interface, 16 go to GPU 1, and 16 go to GPU 2.
| Radeon HD 6990 | Radeon HD 6970 | Radeon HD 6950 | GeForce GTX 580 | |
|---|---|---|---|---|
| Manufacturing Process | 40 nm TSMC | 40 nm TSMC | 40 nm TSMC | 40 nm TSMC |
| Die Size | 2 x 389 mm² | 389 mm² | 389 mm² | 520 mm² |
| Transistors | 2 x 2.64 billion | 2.64 billion | 2.64 billion | 3 billion |
| Engine Clock | 830 MHz | 880 MHz | 800 MHz | 772 MHz |
| Stream Processors / CUDA Cores | 3072 | 1536 | 1408 | 512 |
| Compute Performance | 5.1 TFLOPS | 2.7 TFLOPS | 2.25 TFLOPS | 1.58 TFLOPS |
| Texture Units | 192 | 96 | 88 | 64 |
| Texture Fillrate | 159.4 Gtex/s | 84.5 Gtex/s | 70.4 Gtex/s | 49.4 Gtex/s |
| ROPs | 64 | 32 | 32 | 48 |
| Pixel Fillrate | 53.1 Gpix/s | 28.2 Gpix/s | 25.6 Gpix/s | 37.1 Gpix/s |
| Frame Buffer | 4 GB GDDR5 | 2 GB GDDR5 | 2 GB GDDR5 | 1.5 GB GDDR5 |
| Memory Clock | 1250 MHz | 1375 MHz | 1250 MHz | 1002 MHz |
| Memory Bandwidth | 2 x 160 GB/s (256-bit) | 176 GB/s (256-bit) | 160 GB/s (256-bit) | 192 GB/s (384-bit) |
| Maximum Board Power | 375 W | 250 W | 200 W | 244 W |
Of course, we’re ecstatic that AMD is using fully-functional 40 nm Cayman GPUs—the kind you’d find on a Radeon HD 6970. But that product is already rated for up to 250 W maximum board power. Keeping the 6990’s thermal output manageable meant turning down the clocks from 880 MHz (Radeon HD 6970) to 830 MHz (Radeon HD 6990). AMD also uses a lower memory clock (1250 MHz rather than 1375 MHz). The resulting compute power adds up to 5.1 TFLOPS of single-precision math or 1.27 TFLOPS double-precision.
But AMD also arms this card with a couple of surprises that "break the rules" in the name of more muscle.
- AMD’s Dual-Cayman Board Mashes The Gas
- Radeon HD 6990: Power, Cooling, And Size--All Extreme
- Display Outputs And AMD's Tessellation Coup
- Test Setup And Benchmarks
- Benchmark Results: 3DMark 11 (DX11)
- Benchmark Results: Metro 2033 (DX11)
- Benchmark Results: Lost Planet 2 (DX11)
- Benchmark Results: Aliens Vs. Predator (DX11)
- Benchmark Results: Battlefield: Bad Company 2 (DX11)
- Benchmark Results: F1 2010 (DX11)
- Benchmark Results: Just Cause 2 (DX11)
- Benchmark Results: World Of Warcraft (DX9)
- Benchmark Results: Dual-GPU Performance (CrossFire And SLI)
- Benchmark Results: Quad-CrossFire!
- The Big Reveal: Power And Noise
- Conclusion



Out of spec for default seems kind of weird though.
We don't have two cards here to test, unfortunately. The logged load results for a single card are on the same page, though!
and tested at 7680 x 1600
that will see just how well it does.
That thing is an absolute monster of a card.
They really should have made it 32nm. then the power draw would have fallen below 300w and the thing would be cooler.
STILL NICE WORK AMD
But omg this thing is freakin loud. What's the point of having a quite system now with Noctua fans
Badly
That maybe possible when they get 28nm ready on bulldozer, they are just raping the rewards of old tech.
Edit: I should say its either to much or to little and always in the wrong way for what you pay for. I also dislike the the 375W TDP. We have specs/rules for a reason.
This is isanity!
Well, I guess if you were thinking of running 3, or even 6 displays (which would require at least one hub or daisy chain monitor), this is the card you would want, perhaps even two of them. I'm guessing if you put two of them in you really really want it water cooled.
LOL. Look at the power AND noise graphs, scrum
The big (and really inexcusable) problem is the noise, and to a lesser extent power consumption. It's by far the loudest single card stock cooler ever conceived, and that's taking into account the former champions, the GTX480 and HD5970. The load temps aren't great, but they're acceptable in my opinion. I'm not sure what people were expecting, but this is an extreme high-end dual GPU card, and load temps in the upper 80's C aren't uncommon in this performance segment. The problem is once again the excessive noise that's generated in order to keep the 2 GPU's running at those already high temps.
I totally agree with the reviewer, the HD6990 seems rushed, the drivers are buggy, and if running a fan at 3k+ RPM is the only way to keep a card operating, it probably needs a little more tweaking before release.
Bulldozer will be manufactured using Global Foundries 32nm process, not 28. The node shrink you're referring to for next-gen GPU's will be manufactured using a completely different 28nm process at TSMC. AMD uses Golbal Foundries only for its CPU's at this time.