“The fastest graphics card in the world.” That’s what AMD says about their dual-GPU beast of a card, the HD 6990 that releases today. Will it ring true? With two of their top-of-the-line Cayman GPUs on board, it’s hard to think it wouldn’t. They’re not individually the fastest GPUs in the world, but can they reign supreme as a team? Read on to find out!
Ok, so it may not be as important as something top secret, but the packaging around this card was different than anything else we’ve had sent. The thing came in its own metal case.
It’s a pretty cool way to present a graphics card, that’s for sure. As you can see, AMD is promoting Dragon Age II, which is compatible with Eyefinity. There were a couple Eyefinity screenshots in our preview teaser from last week. I’m not sure if retail cards will come in such a case, but seeing it was pretty neat.
Specifications and Features
Getting right to it, here you can see the 6990 has double the texture units, double the stream processors, double the ROPs and double the RAM of the 6970, so with the exception of memory & CPU clock speeds this is the real deal: two 6970’s on a single PCB.
But there is a little bit lacking – 50 MHz GPU speed and 125 MHz video RAM speed to be exact. In order to maintain a 375 W TDP, they had to reduce those; but there is hope! AUSUM (yep, awesome) – Antilles Unlocking Switch for Uber Mode.
So this awesome little switch changes from the stock, benign 375 W TDP to very close to a full fledged double-6970-on-on-PCB card.
Absorbing power at an unprecedented 450 W TDP, with the second BIOS position you unlock a full 880 MHz on both GPUs. The memory speed still remains short by a little bit but with 4 GB of GDDR5 memory total on-board, it’s hard to fault them for that.
Accompanying the 6990 release will be a new version of Catalyst, 11.4. There is also a re-branding of CCC as well. If you have a non-AMD processor with an AMD GPU, you’ll still use “Catalyst” but those with either an AMD APU or an AMD CPU + GPU will be using “AMD VISION Engine”.
For overclockers the bright side of this story is that they’re unlocking the maximum available frequencies.
In our conference call before receiving the cards, I asked product manager Evan Groenke if there were any plans of introducing voltage control to go along with the increased frequency control. It was a shot in the dark, and it missed. No such thing will come from AMD; any voltage control will be coming from partners.
A logical question would be ‘how do you power such a beast?’ Two words: Volterra and symmetry.
Here’s hoping that ‘digital programmable’ leads to quick implementation of software voltage control (we’re looking at you MSI Afterburner).
Moving on to actual use, this card is especially suited for Eyefinity. It will probably be an absolute powerhouse for regular gaming and benchmarking, but with (an insane) 4 GB of video RAM, Eyefinity is definitely a huge focus for AMD this time around. To aid in easy implementation, you get no less than five video outputs; one DVI and four mini Display Port. Also shipping with all 6990s are three display adapters.
Unfortunately we don’t have the ability to test Eyefinity for you, so we’ll just have to relay AMD’s numbers. With FPS like these, you can see how nicely you can fare with five monitors staring you in the face.
There are two more items we won’t be testing today. Below you’ll find two graphs. The one on the left tests the difference between the two BIOSes. Considering the option is free and there for all (as well as our audience), we didn’t bother testing the slower BIOS.
On the right is the more important graph that we couldn’t test – two 6990’s in CrossfireX.
While it’s not independent testing, if those CrossfireX results are anywhere near applicable in the real world, color me impressed. AMD has gotten it right with the current generation’s CrossfireX scaling, that’s for sure.
HUGE! is the first impression anyone will have when they see this thing. In person or on a computer screen, it’s just plain giant. There is precedent though; it’s exactly the size of its dual-GPU predecessor, the 5970.
Even better than HUGE, it’s huge and pretty nice looking too. The current generation AMD series’ stock coolers have been pretty easy on the eyes and the 6990 doesn’t disappoint there either.
You already know about the five video connectors, but here’s a real photo of them. To the right you can see the required dual eight-pin power connectors.
The back of the card is covered with a metal shroud. While I didn’t take the card apart (heh, I was mentioned specifically in the conference call as one who loves to do so – and quickly instructed not to), from their cooler blowup diagram, we can see that there are memory modules on the back of the card that this helps cool. At least it seems like something is back there, because that back plate gets mighty warm under extended benchmarking.
On the right you can see the BIOS selector switch, which is in the same location as those of the 6970 and 6950 that came before.
Last, but not least, a comparison – of size. The 6970 was already a big card; larger than an aftermarket-designed 5870. Yet the 6990 simply eclipses both of them. The first impression is worth repeating: HUGE!
As I mentioned before, AMD requested reviewers not remove the cooler before testing. That’s because of their apparently very special “phase change TIM”.
It’s similar to the 5970 that came before, with some improvements. The fan itself is very far from quiet. To its credit, that was quieter than the 6970 at the same speed, and at a lower pitch which helped make it more tolerable. Let’s face it, squirrel cage fans just aren’t quiet by their very nature. But how did it perform?
To measure idle temperatures, the GPU was left sitting for at least five minutes. After taking the idle measurement, Furmark was run for three minutes, which allowed temperatures to rise and level off. Temperatures were measured with the fan operating at 75% and are normalized to 25° C ambient.
|Temp Idle:||33° C||38° C|
|Temp Load:||51° C||69° C|
Not too bad, considering the amount of heat dissipation area is effectively cut in half for two of the same GPU. There’s some room for clocking here but not a lot once you crank up the volts. To pour on more voltage, you’ll likely need either really low ambient temperatures or aftermarket cooling.
One very important note – due to the way this card has to be cooled, it shoots warmed air forward and backward, meaning half of the heat (plus that from the back plate) is going to be dissipated right into your case. Those of you that purchase one of these will want to make absolutely sure you have great case airflow. As you can see, it’s not exactly a cool running card, so remember to maintain good airflow lest it heat up the rest of your system or circulate heated air in on itself.
We’ll have to depend on AMD for the only photo of a naked 6990. If I get around to an aftermarket cooling solution or decide to volt-mod it, you’ll get plenty of photos. Until then though, these images from AMD will have to do.
They also gave us some nice macro images just this evening.
For those interested, here’s the higher resolution photo of the full PCB and here is the high resolution photo of the voltage regulators. This is definitely one brute of a card.
We all know this card isn’t going to be hugging any trees. It was made for power and as such, power it will suck out of your PSU. Without belaboring the point, let’s get to the numbers. The GPU and CPU were loaded courtesy Furmark and Prime 95, respectively.
|GPU & CPU Load||371W||X||557W|
While it does pull more power than the Sparkle X580, it surprisingly stayed within 50 watts under load. Pretty impressive considering there are two GPUs on-board. You don’t want to get this GPU without a minimum 600-650 W power supply, and I’d go with 750 W to be safe, especially if you have a high end X58-based i7 to go with it.
It’s also important to note these numbers were only put up with Furmark. AMD didn’t design the card for using Furmark all the time; they designed it to be an FPS monster. These numbers were absolute peak wattage pulls. Even in heavy 3D benchmarking, system wattage very rarely went over 440 W. When gaming, the load will likely be even less than that.
Like the 6970 when it first came out, there is no readily available software that can adjust voltage yet. With the voltage regulator, I’m hoping that will be remedied quickly, but as such the headroom without it isn’t massive.
The first thing to remember for anyone that hasn’t used a dual-GPU board before: be sure to change frequencies and power settings on both GPUs. Changing them on just the screen that pops up in Overdrive and not clicking the drop-down menu to toggle GPUs will give you only one overclocked GPU.
After testing for stability through all benches, this card settled in at a decent 940 Mhz GPU 24/7 stable (except for 3DMark03, which hates me and had to be reduced to 935 MHz). With the limited time available to play with the card, not much effort could go into finding the maximum memory frequency. It was bumped to 1265 MHz and left alone.
Considering there was no voltage increase to speak of, that’s not too bad for an already powerful card, gaining 60 MHz on each GPU. With such a solid base to start from, anything over is just icing on the cake.
Test System and Methodology
There are four cards on the testbed for this review, three from AMD (well, one is ATI) and one from NVIDIA. Thankfully the one NVIDIA competitor is among the best it has to offer, so we’ll see top-of-the-line vs. top-of-the-line here. All cards were tested on Intel’s Sandy Bridge platform.
|Processor||Intel i7 2600K|
|Motherboard||ASRock P67 Extreme 6|
|RAM||2x4G G.Skill Ripjaws DDR3-2133|
|Video||AMD 6990 |
vs. Sparkle X580
vs. AMD 6970
vs. ASUS Matrix 5870 Platinum
|Power Supply||Corsair TX650|
|Operating Systems||Windows 7 x64|
The results graphed below were obtained by operating the CPU & RAM at their stock speeds (3.4 GHz and DDR3-2133/9-11-9-24). The 6990 was run both stock and overclocked. Stock benchmarks were run three times and the results were averaged. Stock game tests and all overclocked results were run once.
The graphs you will see are all normalized to the stock 6990 results. The 6990 at stock is always 100% and all other results are expressed and graphed as percentages relative to its score/FPS. The actual benchmark results are in parenthesis below the percentage.
Performance Stock and Overclocked
This card was exciting to bench. It’s not every day you get to experiment with a card flaunting itself as “the fastest graphics card in the world”. Let’s jump right in.
Many overclockers (yours truly included) end up coming to our hobby by way of gaming. When you’re in a game, you need every bit of FPS you can squeeze out of your machine, lest your head end up as a blood splatter on the wall behind you. We’ll test the 6990’s FPS acumen on three game benchmarks today.
Stalker: Call of Pripyat
Stalker is a nice bench to test various levels of processing power, otherwise known as eye candy. Run at 1920 x 1080 (1080P) at the Ultra Detail, DX11 setting, there were three configurations –
- MSAA and Tessellation both turned off
- MSAA set to 4x and Tessellation off, and
- MSAA at 4x and Tessellation on
– each more difficult than the last.
The 6990 starts off well and just keeps getting more impressive. The more processing power that was needed, the farther it separated itself from the competition. Eyefinity would shine on this benchmark. It’s very easy to picture playable FPS with three (or five) monitors.
Aliens vs. Predator DirectX 11 Benchmark
Sticking with the DirectX 11 theme -it’s what this card was designed for after all- we now bring you some AvP, first with the benchmark’s default settings and then with everything cranked as high as it would go.
AvP really shows off the 6990’s capabilities, with over a 40% improvement from the closest competition in both configurations benched.
Heaven is the newest HWBot 3D benchmark, complete with its own custom HWBot wrapper. This is one long benchmark and it is quite strenuous, with massive detail and effects levels.
Again, more than 40% over the closest competition. The amount of power in this GPU is just astounding. Interestingly, the card lost a little bit overclocked in the DX9 bench. Considering it was a single run, that’s not anything to worry about. You can begin to see a nick in the 6990’s armor – older DirectX versions.
Aquamark 3 and 3DMark01
Starting with the oldest first, Aquamark 3 and 3DMark01 are both almost all CPU now-a-days. There is something to be had from overclocking the bejezus out of a GPU for them, but CPU is where it’s at; so if these scores seem a bit low, it’s mostly because the rest of the system is running at stock.
What’s really funny is that increasing GPU power is inversely proportional to score on these older benchmarks. The graphs go in the opposite direction. Considering the current generation of GPUs is absent form the top 20 Aquamark 3 results and the top 20 3DMark01 results, that actually makes sense. Nothing to stress over here.
3DMark03 and 3DMark06
Ironically, while 3DMark03 is older, it is still very dependent on GPU processing power and scales well with GPU strength. In fact, it’s so tough on GPU overclocks, it alone had to have the overclocked frequency bumped down 5 MHz to complete the bench. 3DMark06 is another story altogether; with each new generation it becomes more and more CPU-dependent.
For the third time (in 06, 01 and Heaven DX9) the score decreased with overclock. I even ran 06 multiple times because it seemed counter-intuitive, yet the discrepancy remained. Graphics tests were almost identical to the stock run but the CPU score’s variation between runs took it just below the stock score.
In 03 we see the potential of this powerhouse, scoring almost 20,000 (just under 20%) better than the X580.
3DMark Vantage and 3DMark 11
These two should show a bit more improvement, especially 3DMark 11, which is almost all GPU. Vantage is definitely dependent on CPU scores, but much less so than 06.
Vantage shows very similar gains to 03, netting just over 20% when overclocked. Even at stock, a card scoring over 30,000 points straight out of the box is very impressive. In 3DMark 11, the 6990 is clearly king-of-the-hill. A 40% score increase shows just how strong it is when the metric is extremely GPU dependent.
Pushing the Overclock
Now you’ve seen what the GPU can do versus others at stock CPU speeds. Let’s have some fun and see the true potential of this beast when you put a 5.4 GHz 3D benchable Sandy Bridge behind it.
So close to breaking 40,000 points in Vantage. 06 was getting near 40K too. With an i7 980X or 990X, breaking that barrier shouldn’t be much of a problem. These results aren’t bad at all considering the lack of voltage control so far.
3DMark 11 shows an increase of over 900 points, and most of it was from the GPU overclock. You know it’s a GPU dependent bench when a 2.0 GHz CPU overclock doesn’t even help net you a 10% increase in score. Still though, almost 11,000 points in that bench is impressive indeed.
Heaven is much the same story; so GPU dependent it gained less than 100 points. It’s nice to see 3D benchmarks coming back that are actually 3D benchmarks. The other ones are quite fun and are great to tinker with but if you really want to test out your graphics hardware, these two are great metrics.
One thing is certain, and my fellow benchmarkers will appreciate this – the 6990 needs moar volts! There is lots of potential under this hood; it just needs some juice to get there, and possibly a different cooler. Can’t wait to see who comes out with the first software that can do it.
Final Thoughts & Conclusion
“The fastest graphics card in the world.” Yep, it wins the title. Maybe not for housing the fastest individual GPUs (that still goes to the GTX 580), but it is definitely the fastest graphics card as a whole. Tailor made for DirectX 11, any gamer would do well to put one of these in their system, especially if you have or plan to get multiple monitors to run Eyefinity.
Benchmarkers are going to have fun with it too. Crossfire-on-a-stick is always fun to toy with, as evidenced by the popularity of the 5970 and even more so, the 4870×2 before it. Thanks to its symmetrical configuration, you should be able to strap two GPU pots side-by-side on this thing. Now that would be fun to play with!
The one thing that hasn’t leaked about these cards is the price; and if you want the best, you’re going to pay for it. MSRP comes in at a rather staggering $699.00. That’s the price you pay for the absolute top of the line.
More astute readers will notice that you can get 6970s now for as low as $339/each. That begs the question – why pay more for the same amount of card?
Analyzing HWBot results for two 6970s in Crossfire, the scores obtained above (with no voltage control) are actually higher than all but the top two in Vantage and 3DMark06. Three of the four that scored higher were run with i7 980X CPUs. The other (06) run was with a higher clocked i7 2600K (5.7 GHz), so it stands to reason those four results were higher and it hasn’t got a lot to do with the GPU. The heaven result above would make the top ten in Heaven (DX11) for Crossfire 6970s.
Presumably the decreased lag from having the two GPUs on the same board help it score a little better. That said, it’s really close to two 6970’s. With that in mind, perhaps $650 would be a better price point. Time will tell if this gets reduced. As any enthusiast knows, there’s always a premium to be paid for the latest and greatest.
People may ask how SLI GTX 580’s compare, but is it a comparison really? 580’s still retail for $500 and up. To get whatever gain you would get, you’re looking at $300+ more than this card. That’s not chump change, and the 6990 easily wins that price-for-performance battle, at least for ambient cooling. There are some absolutely monster GTX 580’s tearing up HWBot right now, so time will tell how this card will fare under extreme conditions.
Regardless, this is definitely one for the ages. This present-day Hercules, it truly is the fastest graphics card in the world.
– Jeremy Vaughan (hokiealumnus)