Dual-GPU graphics cards have always been kind of strange. Technically, one of them is probably the fastest video card on the planet at any given time. For over a year now, for instance, the Radeon HD 5970 has held that title. Yet dual-GPU cards don’t garner loads of attention, for a host of reasons: they’re subject to the same performance and compatibility pitfalls as any multi-GPU configuration, they tend to have rather extreme power and cooling needs, and they’re usually really freaking expensive, to name a few.
Nevertheless, these odd creations are here to stay. Much of the credit for that fact goes to AMD. The company has been carefully refining its dual-GPU products over the past few years, ever since it decided to stop producing as-big-as-you-can-make-them GPUs. Instead, recent flagship Radeons have been based on pairs of mid-sized chips.
The Radeon HD 5970 was odd in that it wasn’t the absolute pinnacle of extremeness that one would expect out of this class of product. The card’s sheer length and price tag were both plenty extreme, but its 725MHz default clock speed gave it performance closer to a pair of Radeon HD 5850s than to a pair of higher-end 5870s. The limiting factor there was power draw. AMD had to tune the card conservatively to ensure that it didn’t exceed its 300W—its rated power draw and the max capacity provided by its 6- and 8-pin auxiliary power connectors—even in absolute peak cases. To skirt this limitation somewhat, AMD practically encouraged 5970 owners to venture into 400W territory by overclocking their cards, even going so far as to screen the chips to ensure they would reach clock speeds similar to the Radeon HD 5870’s. It was innovation, of a sort, born of necessity.
Now the 5970’s successor has arrived in the form of a product code-named Antilles: the Radeon HD 6990, an all-new card based on a pair of the “Cayman” GPUs that power the Radeon HD 6970. Cayman is something of an incremental improvement over the Cypress chip that powered the Radeon HD 5970, so one might expect the 6990 to be an incremental step up from the 5970, as well. That’s not quite the case, for a couple of reasons.
First, AMD has endowed the 6990 with a pair of 8-pin auxiliary power connectors and raised the card’s max power rating to 375W. That gives the card quite a bit more headroom. Second, and more critically, AMD built a power-capping feature into the Cayman known as PowerTune that allows the GPU to monitor its own power draw and ramp back clock speeds if needed to stay within its prescribed power envelope. Although PowerTune doesn’t often limit performance dramatically in typical gaming workloads, we’ve found that it will kick in when synthetic tests push the GPU past its normal bounds. That ability to prevent problems in worst-case scenarios has freed AMD to push for higher default clock speeds without fear of creating problems.
As a result of these and other changes, AMD has set the Radeon HD 6990’s clock speed at 830MHz while leaving all of Cayman’s execution units enabled. Each GPU on the card also has 2GB of GDDR5 memory clocked at 1250MHz, for an effective transfer rate of 5 GT/s. Those numbers put the 6990’s theoretical peak performance right in between what one would expect from a couple of Radeon HD 6950s and a couple of Radeon HD 6970s—not too shabby, to say the least.
AMD apparently wasn’t satisfied with that achievement, though. As you may know, all Radeon HD 6900-series cards have a dual-position switch on the top of the card near the CrossFire connector, ostensibly to enable one to switch to a recovery firmware in the event of a failed video BIOS flash attempt. On the 6990, however, moving that switch from its default position (2) to the other one (1) enables access to a hopped-up BIOS. AMD calls it the “Antilles Unlocking Switch for Uber Mode” or—yes, this is happening—AUSUM. Several things happen when your 6990 cards goes into the umlaut-impaired uber mode. The base GPU clock rises to 880MHz, same as the 6970, and the core GPU voltage rises from 1.12V to 1.175V. Also, the board’s PowerTune limit is raised to 450W. You’re essentially overclocking your card when you switch it into uber mode; AMD doesn’t guarantee proper operation for everyone in every system. However, going AUSUM worked just fine with our 6990 sample on our Intel X58-based test system, much like the 5970 did for us at its “suggested” overclocked speed.
If that’s not enough AUSUM-ness for you, AMD has given 6990 users more than enough leeway to get into real trouble. The Overdrive controls in the AMD control panel will allow GPU overclocks as high as 1200MHz, with memory overclocking as high as 1500MHz (or 6 GT/s).
| Peak bilinear
| Peak bilinear
| Peak shader
|GeForce GTX 560 Ti||26.3||52.6||52.6||1263||1644||128|
|GeForce GTX 570||29.3||43.9||43.9||1405||2928||152|
|GeForce GTX 580||37.1||49.4||49.4||1581||3088||192|
|Radeon HD 6850||24.8||37.2||18.6||1488||775||128|
|Radeon HD 6870||28.8||50.4||25.2||2016||900||134|
|Radeon HD 6950||25.6||70.4||35.2||2253||1600||160|
|Radeon HD 6970||28.2||84.5||42.2||2703||1760||176|
|Radeon HD 5970||46.4||116.0||58.0||4640||1450||256|
|Radeon HD 6990||53.1||159.4||79.7||5100||3320||320|
|Radeon HD 6990 AUSUM||56.3||169.0||84.5||5407||3520||320|
With or without the AUSUM switch enabled, the 6990’s specifications are downright staggering. On paper, at least, it’s far and away the fastest consumer graphics card ever. Of course, we’re just adding up the capacities of its two individual GPUs in the table above and assuming the best—perfect scaling—will happen. That’s not always how things work out in the real world, of course, but the 6990 has more than enough extra oomph to overcome less-than-ideal outcomes.
Yep, the Radeon HD 6990 is long—just a sliver shy of a full 12″, in fact, inviting all sorts of remarks that are surely beneath us. You will want to check the clearance in your case carefully before ordering up one of these puppies. Even the ridiculously lengthy 5970 is a tad shorter.
Beneath the 6990’s massive cooling shroud lies a brand-new board design that, interestingly, places a single blower in the center of the card, above the voltage regulators and the PCIe bridge chip that acts as an interconnect between the two GPUs and the rest of the system. Those VRMs, incidentally, are digital programmable units from Volterra that are unique to Antilles. AMD says they enable lower temperatures and lower power draw than conventional VRMs.
The blower is flanked by a pair of heatsinks with copper vapor chambers at their bases. AMD claims that, although this card fits into roughly the same form factor as the 5970, it moves 20% more air with this arrangement. In addition, the firm tell us the thermal interface material between the heatsinks and the GPUs is a special, phase-change variety that offers 8% better performance than the standard gray goo. Take note: if you disassemble your card, you’ll likely have to use regular thermal paste when reassembling it, sacrificing some of its cooling efficiency. We’ve avoided taking ours apart, so far, because we want our power, noise, and temperature readings to track with what you’d see from retail products.
An array of compact Mini-DisplayPort connectors allows the 6990 to sport a rich mix of display outputs while leaving room for a full slot cover of exhaust venting. The 6990, obviously, can drive up to five displays natively. Since it supports DisplayPort 1.2, it can even drive multiple displays simultaneously off of a single output with the help of a DisplayPort hub.
AMD clearly leads the industry on the display output front. The only drawback is the need for adapters to support “legacy” displays with HDMI or DVI inputs. Fortunately, every Radeon HD 6990 will ship with a trio of adapter dongles to convert those Mini-DP ports to serve other standards: one passive HDMI type, one passive single-link DVI type, and one active single-link DVI type. Out of the box, the 6990 should be able to drive a trio of single-link DVI monitors, then. The reason that third adapter is of the “active” variety is that the GPU has a limited number of timing sources for its display outputs. If you’d like to drive more than three “legacy” displays with a 6990, you’ll need additional active adapters. Similarly, driving a second or third dual-link DVI display, such as a 30″ panel, will require additional active, dual-link-capable dongles.
All of this to-do about multiple display is, of course, only an issue because AMD has been pushing its Eyefinity multi-monitor gaming feature so enthusiastically in the past year and a half—and because the 6990 looks like the perfect device for driving large numbers of megapixels. Given the 6990’s five-way output array, AMD has pointed out how naturally this card would support an interesting display configuration: a five-display-wide wall of monitors in portrait orientation. That sounds like a whole lotta bezel area to me, but it’s certainly a bevy o’ pixels.
Before we move on to our test results, where you can see exactly how the 6990 performs, there’s just a couple more details to which we should attend. Although the 6990 is being unveiled today, you likely won’t see it selling at online retailers until some time later this week or perhaps early next. When it does arrive, if you’d like to make one your very own, you need only hand over something close to its list price to your favorite online retailer. That price? $699.99.
That’s a lot, but given that the Radeon HD 6970 is selling for about 340 bucks a pop, this single card that has essentially two of ’em onboard isn’t priced at any great premium, believe it or not.
Our testing methods
As ever, we did our best to deliver clean benchmark numbers. Tests were run at least three times, and we’ve reported the median result.
Our test systems were configured like so:
|North bridge||X58 IOH|
|Memory size||12GB (6 DIMMs)|
|Memory type||Corsair Dominator CMD12GX3M6A1600C8
DDR3 SDRAMat 1600MHz
|Memory timings||8-8-8-24 2T|
|Chipset drivers||INF update 184.108.40.2065
Rapid Storage Technology 220.127.116.114
with Realtek R2.58 drivers
|Radeon HD 5970 2GB
with Catalyst 11.4 preview drivers
|Dual Radeon HD 6950 2GB
with Catalyst 11.4 preview drivers
| Radeon HD 6970 2GB
with Catalyst 11.4 preview drivers
|Dual Radeon HD 6970 2GB
with Catalyst 11.4 preview drivers
| Radeon HD 6990 4GB
with Catalyst 11.4 preview drivers
|MSI GeForce GTX 560 Ti Twin Frozr II 1GB +
Asus GeForce GTX 560 Ti DirectCU II TOP 1GB
with ForceWare 267.26 beta drivers
|Zotac GeForce GTX 570 1280MB
with ForceWare 267.24 beta drivers
|Zotac GeForce GTX 570 1280MB +
GeForce GTX 570 1280MB
with ForceWare 267.24 beta drivers
|Zotac GeForce GTX 580 1536MB
with ForceWare 267.24 beta drivers
|Zotac GeForce GTX 580 1536MB +
Asus GeForce GTX 580 1536MB
with ForceWare 267.24 beta drivers
|Hard drive||WD RE3 WD1002FBYS 1TB SATA|
|Power supply||PC Power & Cooling Silencer 750 Watt|
|OS||Windows 7 Ultimate x64 Edition
Service Pack 1
Thanks to Intel, Corsair, Western Digital, Gigabyte, and PC Power & Cooling for helping to outfit our test rigs with some of the finest hardware available. AMD, Nvidia, and the makers of the various products supplied the graphics cards for testing, as well.
Unless otherwise specified, image quality settings for the graphics cards were left at the control panel defaults. Vertical refresh sync (vsync) was disabled for all tests.
We used the following test applications:
- Aliens vs. Predator benchmark
- Battlefield: Bad Company 2
- F1 2010
- Sid Meier’s Civilization V
- Metro 2033
- StarCraft II
- Fraps 3.3.1
- GPU-Z 0.5.1
Some further notes on our methods:
- Many of our performance tests are scripted and repeatable, but for some of the games, including Battlefield: Bad Company 2 and Bulletstorm, we used the Fraps utility to record frame rates while playing a 60- or 90-second sequence from the game. Although capturing frame rates while playing isn’t precisely repeatable, we tried to make each run as similar as possible to all of the others. We raised our sample size, testing each Fraps sequence five times per video card, in order to counteract any variability. We’ve included second-by-second frame rate results from Fraps for those games, and in that case, you’re seeing the results from a single, representative pass through the test sequence.
We measured total system power consumption at the wall socket using a Yokogawa WT210 digital power meter. The monitor was plugged into a separate outlet, so its power draw was not part of our measurement. The cards were plugged into a motherboard on an open test bench.
The idle measurements were taken at the Windows desktop with the Aero theme enabled. The cards were tested under load running Battlefield: Bad Company 2 at a 2560×1600 resolution with 4X AA and 16X anisotropic filtering. We test power with BC2 because we think it’s a solidly representative peak gaming workload.
We measured noise levels on our test system, sitting on an open test bench, using an Extech 407738 digital sound level meter. The meter was mounted on a tripod approximately 10″ from the test system at a height even with the top of the video card.
You can think of these noise level measurements much like our system power consumption tests, because the entire systems’ noise levels were measured. Of course, noise levels will vary greatly in the real world along with the acoustic properties of the PC enclosure used, whether the enclosure provides adequate cooling to avoid a card’s highest fan speeds, placement of the enclosure in the room, and a whole range of other variables. These results should give a reasonably good picture of comparative fan noise, though.
- We used GPU-Z to log GPU temperatures during our load testing.
The tests and methods we employ are generally publicly available and reproducible. If you have questions about our methods, hit our forums to talk with us about them.
This game is based on the aging Unreal Engine, but it’s stressful enough on a GPU to still make a decent candidate for testing. We turned up all of the game’s image quality settings to their peaks and enabled 8X antialiasing, and then we tested in 90-second gameplay chunks.
Our single-GPU configs all struggled with this game, as did our pair of GeForce GTX 560 Ti cards in SLI. Those 560s had the least memory of any cards we tested, at 1GB each. Multi-GPU schemes like SLI and CrossFireX have some memory overhead, and we expect that’s what troubled our 560s in this case.
The 6990, meanwhile, goes toe to toe with a thousand-dollar option from Nvidia: a pair of GeForce GTX 580s in SLI. The Nvidia alternative in the same price range as the 6990 would be a pair of GTX 570s, but those are a bit slower. Then again, a couple of 6950s in CrossFireX perform very similarly to the 6990, and flipping the AUSUM switch doesn’t get you much here, either.
F1 2010 steps in and replaces CodeMasters’ previous effort, DiRT 2, as our racing game of choice. F1 2010 uses DirectX 11 to enhance image quality in a few, select ways. A higher quality FP16 render target improves the game’s high-dynamic-range lighting in DX11. A DX11 pixel shader is used to produce soft shadow edges, and a DX11 Compute Shader is used for higher-quality Gaussian blurs in HDR bloom, lens flares, and the like.
We used this game’s built-in benchmarking facility to script tests at multiple resolutions, always using the “Ultra” quality preset and 8X multisampled antialiasing.
The Radeons had a strong showing in the last game, but this is unexpected dominance from AMD. At the highest resolution where the GPU is the primary bottleneck, dual Radeon HD 6950s outrun a couple of GeForce GTX 580s. The 6990 is faster still, and the AUSUM switch nearly moves the 6990 into dual 6970 territory.
Civ V has a bunch of interesting built-in tests. Up first is a compute shader benchmark built into Civilization V. This test measures the GPU’s ability to decompress textures used for the graphically detailed leader characters depicted in the game. The decompression routine is based on a DirectX 11 compute shader. The benchmark reports individual results for a long list of leaders; we’ve averaged those scores to give you the results you see below.
Obviously, the green team takes this one. Not every compute shader is the same, but this one runs better on Nvidia’s Fermi architecture than on Cayman. Regardless of the GPU type, though, one thing holds steady: the performance gains from adding a second GPU are real but modest. That’s why the 6990 is in an unusual position, near the bottom of the pack.
In addition to the compute shader test, Civ V has several other built-in benchmarks, including two we think are useful for testing video cards. One of them concentrates on the world leaders presented in the game, which is interesting because the game’s developers have spent quite a bit of effort on generating very high quality images in those scenes, complete with some rather convincing material shaders to accent the hair, clothes, and skin of the characters. This benchmark isn’t necessarily representative of Civ V‘s core gameplay, but it does measure performance in one of the most graphically striking parts of the game. As with the earlier compute shader test, we chose to average the results from the individual leaders.
The 6990 comes out looking pretty good here, but why is the 5970 so much faster? My guess is that this pixel-shader-intensive test is causing the Cayman GPUs to heat up and invoke their PowerTune limits. Without PowerTune, the 5970 is slower in most real gaming scenarios, but it’s quicker here.
Another benchmark in Civ V focuses, rightly, on the most taxing part of the core gameplay, when you’re deep into a map and have hundreds of units and structures populating the space. This is when an underpowered GPU can slow down and cause the game to run poorly. This test outputs a generic score that can be a little hard to interpret, so we’ve converted the results into frames per second to make them more readable.
My, how the tables have turned! The GeForces take three of the top four spots. Why? I have another crackpot theory for you. There’s tremendous geometric complexity in this late-game scene, with a huge number of units in view at once. Nvidia’s Fermi architecture has some real advantages in geometry processing throughput, and I suspect they’re making themselves known here.
Up next is a little game you may have heard of called StarCraft II. We tested SC2 by playing back 33 minutes of a recent two-player match using the game’s replay feature while capturing frame rates with Fraps. Thanks to the relatively long time window involved, we decided not to repeat this test multiple times. The frame rate averages in our bar graphs come from the entire span of time. In order to keep them readable, we’ve focused our frame-by-frame graphs on a shorter window, later in the game.
We tested at the settings shown above, with the notable exception that we also enabled 4X antialiasing via these cards’ respective driver control panels. SC2 doesn’t support AA natively, but we think this class of card can produce playable frame rates with AA enabled—and the game looks better that way.
The demo we used for testing here is newer than any we’ve used before, and Blizzard has made a number of changes to SC2 over time. As a result, this test turns out to be more taxing than some of our past attempts. The GeForces end up looking very good here, not just in terms of higher average frame rates but also in terms of avoiding the lowest valleys. The frame rate minimums for the Radeons, even the AUSUM 6990, are in the teens.
Battlefield: Bad Company 2
BC2 uses DirectX 11, but according to this interview, DX11 is mainly used to speed up soft shadow filtering. The DirectX 10 rendering path produces the same images.
We turned up nearly all of the image quality settings in the game. Our test sessions took place in the first 60 seconds of the “Heart of Darkness” level.
We’ve been see-sawing back and forth between clear wins for AMD and Nvidia, but this game looks like an even match. Two GTX 570s in SLI perform about the same as a 6990 or a pair of 6970s in CrossFireX. Notice, also, the excellent ~40 FPS minimums produced by a single 6970 or GTX 570. Even those single-GPU cards handle Bad Company 2 pretty darn well.
We decided to test Metro 2033 at multiple image quality levels rather than multiple resolutions, because there’s quite a bit of opportunity to burden these graphics cards simply using this game’s more complex shader effects. We used three different quality presets built into the game’s benchmark utility, with the performance-destroying advanced depth-of-field shader disabled and tessellation enabled in each case.
We’ve seen the same trend in this game for quite a while. As the image quality rises, the Radeons become more competitive. At Metro‘s “Medium” settings, two GTX 570s in SLI are easily faster than the 6990. By the time we reach the “Very high” settings, the opposite is true.
Aliens vs. Predator
AvP uses several DirectX 11 features to improve image quality and performance, including tessellation, advanced shadow sampling, and DX11-enhanced multisampled anti-aliasing. Naturally, we were pleased when the game’s developers put together an easily scriptable benchmark tool. This benchmark cycles through a range of scenes in the game, including one spot where a horde of tessellated aliens comes crawling down the floor, ceiling, and walls of a corridor.
For these tests, we turned up all of the image quality options to the max, along with 4X antialiasing and 16X anisotropic filtering.
Wow, not much drama there as the resolution changes. The Radeons are looking relatively strong here, again, though, with the 6990 out ahead of dual GTX 570s.
AMD initially suggested to us that the 6990’s idle power draw should be somewhat lower than the 5970’s, and so it is. That’s not a huge difference, but it is something. Heck, the entire system based on the 6990 only draws 8W more at idle than the same system equipped with a single GeForce GTX 580.
Under load, the 6990 remains reasonable, drawing less power than a pair of GTX 560s in SLI, though it generally outperforms them. There is a price for being AUSUM, though, and apparently it’s about 50 watts. Who knew? Still, the AUSUM 6990 config draws substantially less power than our GeForce GTX 570 SLI system.
Noise levels and GPU temperatures
The 6990 is the loudest solution we tested, both at idle and, more dramatically, when running a game. That difference is especially perceptible when the card is hitting over 58 on the decibel meter. You will notice the difference between the 6990 and the other solutions; it’s quite audible. The 6990 emits a fairly loud hiss, although its pitch and tenor aren’t especially offensive compared to some of the worst solutions we’ve seen over the years.
Dual-card setups have an acoustic advantage, as our results illustrate. With four slots occupied and two full-length coolers, there’s simply more surface area available for heat dissipation. With that said, AMD has apparently tuned the 6990’s cooler fairly aggressively; it has some of the lowest GPU temperatures of the bunch, and you’ll pay for maintaining those by adding a little extra noise.
With a total of just seven games tested, we can ruthlessly boil down the Radeon HD 6990 and its competition to a simple price-performance scatter plot, like so:
We’ve taken the results from the highest resolution or most intensive setting of each game tested, averaged them, and combined them with the lowest prevailing price at Newegg for each of these configurations. Doing so gives us a nice distribution of price-performance mixes, with the best tending toward the upper left and the worst toward the bottom right.
At present, in the suite of games we tested, AMD looks to have a performance advantage at several key price points. That may be a little jarring if your expectations were set several months ago, when we had something close to parity between red and green. We believe AMD has come by it honestly, delivering some impressive performance gains in recent driver releases. One of those changes, AMD tells us, is a revised resolve mechanism for multisampled antialiasing that improves frame rates generally when MSAA is in use—like in nearly all of our test scenarios—particularly in games that use deferred shading schemes. AMD’s driver developers have made some notable progress in CrossFireX multi-GPU performance scaling, too. SLI scaling has long been one of the hallmarks of Nvidia’s SLI, but AMD has closed that gap in recent months.
Of course, both of these changes benefit the Radeon HD 6990, which has no equal in a single-card package. This is the planet’s fastest single video card, supplanting the Radeon HD 5970 that came before it. The 6990 is even faster than two GeForce GTX 570 cards in SLI, which cost about the same amount, and the 6990 draws less power under load, even in AUSUM uber mode. Add in the 6990’s rich array of display outputs, and there’s no question Nvidia is severely outclassed at this lofty $700 price point. We just hope the 6990 isn’t quite as difficult to find over the next year as the Radeon HD 5970 was during much of its run. We do believe TSMC’s 40-nm supply problems are behind us, so we’re optimistic on that front.
Having said that, we can’t help but notice that AMD does offer a more attractive solution in terms of price, performance, and acoustics in the form of dual Radeon HD 6970 cards. You must really covet slot space—or have designs for a dual-6990, four-way CrossFireX rig—if you pick the 6990 over two 6970s. Not that there’s anything wrong with that.
We also can’t avoid noticing Nvidia still owns the title of the fastest dual-GPU solution on the market, in the form of two GeForce GTX 580s in SLI. And we have some clear indications that Nvidia may be cooking up an answer to the Radeon HD 6990 based on that same technology. The challenge Nvidia faces if it wants to dethrone the 6990 is, of course, power draw and the related cooling required. Given that two GTX 570s are slower than a single 6990 and draw more power, the GeForce team certainly has its work cut out for it. Besting the 6990 will have to involve some deep magic, or at least solid progress on multiple fronts.
Or, you know, a triple-slot cooler.