When the Radeon HD 7970 launched at $550, it looked like a reasonable alternative to the GeForce GTX 590 and Radeon HD 6990. Both dual-GPU boards are measurably faster, but they’re also $700+, power-hungry, and in the case of the 6990, embarrassingly loud. Even still, the 7970's asking price is still pretty steep.
And that’s why a card like the Radeon HD 7950 is such a welcome addition to AMD’s portfolio. The company is, as of this writing, unwilling to comment on the 7950’s anticipated price tag. However, we’ve already run the benchmarks. We know how it stacks up to the Radeon HD 7970 and GeForce GTX 580. So, we know what we’d pay for this new board. If our target is close, we’d be looking for something under $500—perhaps $480 or $490.
What makes the Radeon HD 7950 worth a few bucks more than Nvidia's GeForce GTX 580? Well, let’s have a closer look at the card itself…
Update: Before publication, but after our launch coverage was finalized for international translation, AMD let us know that the Radeon HD 7950 should sell for $450. That's well below where I thought the company would target, given its competition. Clearly, AMD is pricing the 7950 to out-value Nvidia's GeForce GTX 580 (or force its competitor to adjust downward) rather than exist in a price structure defined by the company's single-GPU flagship. Advantage: AMD.

That's a Radeon HD 7970 up top and a Radeon HD 7950 down below. In the right light, they'd pass as twins.
Meet AMD’s Radeon HD 7950
Physically, the Radeon HD 7950 is identical to AMD’s already-available Radeon HD 7970—save one distinguishing feature: a second six-pin auxiliary power connector. That’s a telltale indication of a sub-225 W maximum board power (75 W from the slot, plus up to 75 W from each plug). In fact, AMD rates the 7950 right at 200 W. In comparison, the Radeon HD 7970’s power ceiling is 250 W, necessitating its eight- and six-pin power connectors.
A 10.5” PCB is extended out an additional half of an inch by a metal base plate and plastic shroud. So, plan accordingly when you pick a chassis. This card is fairly long.
As with the Radeon HD 7970, AMD employs a centrifugal fan mounted on one end of the Radeon HD 7950, which blows across the length of the card and exhausts heated air out the back of your chassis. This is the design we prefer. It wasn’t possible to cool the Radeon HD 6990 or GeForce GTX 590 the same way. In both examples, a center-mounted fan exhausted some air from a rear I/O slot and everything else was recirculated.
Because it relies on effective exhaust, one of the card’s two slots is grated for unrestricted air flow. The other slot is populated with four display outputs: one dual-link DVI connector, one full-sized HDMI port, and a pair of mini-DisplayPort outputs.
Board partners will almost certainly bundle a variety of adapters, so be sure you’re getting the components you need before making a purchase. The two Sapphire Radeon HD 7970s we bought came with DVI-to-VGA, mini-DisplayPort-to-DisplayPort, mini-DisplayPort-to-single-link DVI, and HDMI-to-DVI adapters. Meanwhile, the XFX R7950 Black Edition card we received only included an HDMI-to-DVI adapter.
More notable, though, is that all four outputs can be active at the same time, supporting extensive display configurations that you simply cannot achieve on a single Nvidia-based board.
Tahiti Pro: Same GPU, But On A Diet
Radeon HD 7950 centers on the same 4.31 billion-transistor Tahiti GPU as AMD’s faster, more expensive flagship, manufactured on TSMC’s 28 nm node.
However, instead of sporting 32 Compute Units, this scaled-back model comes equipped with 28 Compute Units. As you know, each CU plays host to four Vector Units, each with 16 shaders, ALUs, Stream Processors, or whatever else you’d like to call them. That’s a total of 64 SPs per CU. A quick little multiplication (64*28) gives you a grand total of 1792 SPs on this chip.
And because each of those four missing CUs also included four texture units, that specification drops from 128 to 112.
To help differentiate the Radeon HD 7950 even further, AMD dials back its core clock rate to 800 MHz (down from 925 MHz on the reference Radeon HD 7970). Peak compute performance correspondingly drops to 2.87 TFLOPS from 3.79 TFLOPS.
The render back-ends are independent of the CUs, and AMD leaves all eight ROP partitions enabled, yielding up to 32 raster operations per clock cycle. Six 64-bit memory controllers feed the partitions through a crossbar. An aggregate 384-bit data path populated with 3 GB of GDDR5 memory operating at 1250 MHz adds up to 240 GB/s of bandwidth. That’s a slight drop from the Radeon HD 7970’s 264 GB/s, but still a very substantial increase over the Radeon HD 6970.
| Radeon HD 7950 | Radeon HD 7970 | Radeon HD 6970 | GeForce GTX 580 | |
|---|---|---|---|---|
| Stream processors | 1792 | 2048 | 1536 | 512 |
| Texture Units | 112 | 128 | 96 | 64 |
| Full Color ROPs | 32 | 32 | 32 | 48 |
| Graphics Clock | 800 MHz | 925 MHz | 880 MHz | 772 MHz |
| Texture Fillrate | 89.6 Gtex/s | 118.4 Gtex/s | 84.5 Gtex/s | 49.4 Gtex/s |
| Memory Clock | 1250 MHz | 1375 MHz | 1375 MHz | 1002 MHz |
| Memory Bus | 384-bit | 384-bit | 256-bit | 384-bit |
| Memory Bandwidth | 240 GB/s | 264 GB/s | 160 GB/s | 192.4 GB/s |
| Graphics RAM | 3 GB GDDR5 | 3 GB GDDR5 | 2 GB GDDR5 | 1.5 GB GDDR5 |
| Die Size | 365 mm2 | 365 mm2 | 389 mm2 | 520 mm2 |
| Transistors (Billion) | 4.31 | 4.31 | 2.64 | 3 |
| Process Technology | 28 nm | 28 nm | 40 nm | 40 nm |
| Power Connectors | 2 x 6-pin | 1 x 8-pin, 1 x 6-pin | 1 x 8-pin, 1 x 6-pin | 1 x 8-pin, 1 x 6-pin |
| Maximum Power | 200 W | 250 W | 250 W | 244 W |
| Price (Street) | $549 | ~$350 | ~$480 |
Tessellation
I like to use the HAWX 2 benchmark sequence as a way to gauge tessellation performance, since the feature can be turned on and off independently and I’ve seen what wireframe models of the geometry actually look like.
This was a metric that AMD very much did not like back when its Radeon HD 5870 was its best basis for comparison. The Radeon HD 6800-series cards improved the company’s handling of excessive geometry, and the 6900-series boards were even better.


Now, we can see that, in relative terms, the Radeon HD 7900s aren’t particularly great performers in HAWX 2. The 7950 trails Nvidia’s GeForce GTX 570 with tessellation turned on and off.
However, we’re not interested in the frame rate result. Rather, we want to know how much performance each card manages to retain when switching tessellation on.

Here, the Radeon HD 7900s shoot to the top as they post 82% of their tessellation-off frame rate. That’s better than any of the Nvidia cards, and a 10% improvement over the previous-generation flagship.
Audio Output
AMD was the first company to facilitate bitstreaming high-definition lossless audio formats like Dolby TrueHD and DTS-HD Master Audio. Prior to that, your DVD playback software would have been responsible for decoding the compressed audio into linear PCM and sending that 16-bit/48 kHz signal over HDMI to your receiver. With the right Blu-ray, however, bitstreaming can get you 24-bit resolution and 96 kHz sampling. Now, it’s highly unlikely you’d ever be able to tell the difference. But for the folks who take comfort in knowing their equipment is running at its peak potential (yes, I’m personally OCD about this), this is a feature of interest.
Granted, the Radeon HD 7950 is too high-end of a card for most HTPCs. It’s long, its TDP is too high, and it makes too much noise under load. Nevertheless, I got it running in my own home theater environment and confirmed that bitstreaming works just as well on the Radeon HD 7900 series as it did in the two generations prior (using CyberLink’s PowerDVD 11).
Some of the card’s other audio-oriented features, like discrete digital multi-point audio, remain untested, as we’re not using any speaker-equipped displays in the lab. However, AMD’s driver framework for supporting multiple, independent audio streams appears to be in place.
Anticipating today’s launch, XFX shot over its R7950 Black Edition, which deviates from AMD’s reference design on a number of different levels.
Right off the bat, we noticed that this wasn’t based on a reference PCB. Also, XFX’s cooler is dramatically different. Instead of the centrifugal fan exhausting out a single-slot grate, the card employs a feature XFX calls Double Dissipation. Two axial fans blow over aluminum fins, directing some exhaust out the back, but not all.
XFX also doesn’t rely on AMD’s reference clock rates. The R7950 Black Edition ships with a 900 MHz core and 1375 MHz memory—a notable step up from the default 800/1250 MHz.

Hoping that XFX’s aftermarket cooling and unique layout would make for a good overclocking foundation, we tweaked our sample up as high as it’d go using AMD’s Overdrive interface, eventually settling on a stable 1025 MHz core and 1425 MHz memory frequency.
As you can see in the following few charts, the impact of the speed-up is generally sufficient for catching a stock Radeon HD 7970.






Of course, curious about the efficacy of XFX’s modifications, we reset everything to default and took some power, thermal, and noise measurements.


XFX’s box says that the R7950’s dual fans help reduce noise and operating temperatures. Just as interesting, though, is the fact that idle power consumption drops by about 7 W compared to the stock Radeon HD 7950, and long-idle power use falls by 8 W.


An early version of the R7950’s video BIOS employed aggressive fan speed settings that had a major impact on thermals, dropping load temperatures from 75 degrees Celsius down to 68. However, the production firmware relaxes that profile, resulting in less noise, but giving up all of the benefit in our temperature measurement. As a result, we see the same 75 degrees during a 10-loop run of the Metro 2033 benchmark.
At idle, XFX’s R7950 Black Edition does shave off two degrees compared to AMD’s reference design.


And it does so without making any additional noise. At idle, our XFX sample achieves the same whisper-quiet 37.3 dB(A) reading from one meter away as our AMD Radeon HD 7950 cards.
Under load, the production fan profile turns what was originally a very loud pair of axial fans into a solution that makes less noise than AMD’s own centrifugal implementation. The company does say that enthusiasts who really want the more aggressive firmware will be able to get it from XFX’s site.
| Test Hardware | |
|---|---|
| Processors | Intel Core i7-3960X (Sandy Bridge-E) 3.3 GHz at 4.2 GHz (42 * 100 MHz), LGA 2011, 15 MB Shared L3, Hyper-Threading enabled, Power-savings enabled |
| Motherboard | Gigabyte X79-UD5 (LGA 2011) X79 Express Chipset, BIOS F8 |
| Memory | G.Skill 16 GB (4 x 4 GB) DDR3-1600, F3-12800CL9Q2-32GBZL @ 9-9-9-24 and 1.5 V |
| Hard Drive | Intel SSDSC2MH250A2 250 GB SATA 6Gb/s |
| Graphics | AMD Radeon HD 7950 3 GB |
| AMD Radeon HD 7970 3 GB | |
| AMD Radeon HD 6990 4 GB | |
| AMD Radeon HD 6970 2 GB | |
| AMD Radeon HD 6950 2 GB | |
| Nvidia GeForce GTX 590 3 GB | |
| Nvidia GeForce GTX 580 1.5 GB | |
| Nvidia GeForce GTX 570 1.25 GB | |
| Power Supply | Cooler Master UCP-1000 W |
| System Software And Drivers | |
| Operating System | Windows 7 Ultimate 64-bit |
| DirectX | DirectX 11 |
| Graphics Driver | AMD 8.921.2 RC11 (For Radeon HD 7970 and 7950) |
| AMD 11.12 CAP3 (For CrossFire Configurations) | |
| AMD Catalyst 11.12 | |
| Nvidia GeForce Release 285.62 | |
We've transitioned our test platform for graphics to a Sandy Bridge-E-based Core i7-3960X overclocked to 4.2 GHz. You'll notice that, in some cases, that's still not enough processing power to let some of our more demanding two- and four-GPU configurations really stretch their legs. I also made the call to swap from an Asus motherboard to a Gigabyte platform after discovering, last year during a Z68 Express motherboard round-up, that certain settings in Asus' BIOS would alter Turbo Boost behavior in an undesirable way.
| Games | |
|---|---|
| Battlefield 3 | Ultra Quality Settings, No AA / 16x AF, 4x MSAA / 16x AF, vsync off, 1680x1050 / 1920x1080 / 2560x1600, DirectX 11, Going Hunting, 90-second playback, Fraps |
| Crysis 2 | DirectX 9 / DirectX 11, Ultra System Spec, vsync off, 1680x1050 / 1920x1080 / 2560x1600, No AA / No AF, Central Park, High-Resolution Textures: On |
| Metro 2033 | Very High Quality Settings, AAA / 4x AF, 4x MSAA / 16x AF, 1680x1050 / 1920x1080 / 2560x1600, Built-in Benchmark, Depth of Field filter Disabled, Steam version |
| DiRT 3 | Ultra High Settings, No AA / No AF, 8x AA / No AF, 1680x1050 / 1920x1080 / 2560x1600, Steam version, Built-In Benchmark Sequence, DX 11 |
| The Elder Scrolls V: Skyrim | High Quality (8x AA / 8x AF) / Ultra Quality (8x AA, 16x AF) Settings, FXAA enabled, vsync off, 1680x1050 / 1920x1080 / 2560x1600, 25-second playback, Fraps |
| 3DMark 11 | Version 1.03, Extreme Preset |
| HAWX 2 | Highest Quality Settings, 8x AA, 1920x1200, Retail Version, Built-in Benchmark, Tessellation on/off |
| World of Warcraft: Cataclysm | Ultra Quality Settings, No AA / 16x AF, 8x AA / 16x AF, From Crushblow to The Krazzworks, 1680x1050 / 1920x1080 / 2560x1600, Fraps, DirectX 11 Rendering |
| SiSoftware Sandra 2012 | Sandra Tech Support (Engineer) 2012.SP1c, GP Processing and GP Bandwidth Modules |
| CyberLink MediaEspresso 6.5 | 449 MB 1080i Video Sample to Apple iPad 2 Profile (1024x768) |
| LuxMark | 64-bit Binary, Version 1.0 |
| MotionDSP vReveal 3 | 1080i Video Sample Playback, Apply One-Click Fix |

AMD’s Radeon HD 7950 starts off strong in 3DMark 11, edging past Nvidia’s GeForce GTX 580.
The company’s dual-GPU Radeon HD 6990 remains the fastest board you can buy for this benchmark. However, as we showed in AMD Radeon HD 6990 4 GB Review: Antilles Makes (Too Much) Noise, there’s a hefty price to pay with regard to noise (not to mention a price tag that sits around $730).







Another strong showing puts the Radeon HD 7950 ahead of Nvidia’s GeForce GTX 580 at all three tested resolutions, with and without multi-sample anti-aliasing applied.
It’s worth noting that AMD’s Radeon HD 6990 remains at the top without anti-aliasing. However, once it’s applied, Nvidia’s GeForce GTX 590 pulls into the lead.



When we organize by putting the longest bars first, the Radeon HD 7950 drops to the very bottom of our charts, behind even AMD’s own Radeon HD 6950.
Is the story really this bad for AMD in Crysis 2?
Actually, no. We knew from our Radeon HD 7970 launch coverage that this architecture does not handle DirectX 9-based apps particularly well. Whether that’s an inherent compromise or a lack of driver maturity will become more evident over time.
What is immediately apparent, though, is that switching into DirectX 11 mode, which hammers every other architecture, turns out to be a simple movement sideways for the Radeon HD 7970 and 7950. In fact, at 2560x1600, both boards perform better using DirectX 11.
Looking at the DirectX 11 numbers, we see the Radeon HD 7950 pull itself together, falling between the GeForce GTX 570 and 580. Although we know that DirectX 11 mode applies an unnecessarily extensive amount of geometry, we’d hesitate to blame the loss here to Crytek’s implementation of tessellation. After all, we saw in the HAWX 2 scaling numbers that this card is able to maintain more of its performance with tessellation applied than any competing product.



Its popularity aside, Skyrim isn’t a great game to test with because it’s so predominantly platform-limited. Even at 4.2 GHz, our Core i7-3960X is the primary bottleneck at 1680x1050 and 1920x1080.
Nevertheless, AMD’s Radeon HD 7950 lands right about where we’d expect it at both the High and Ultra quality presets, ahead of Nvidia’s GeForce GTX 580, but behind the Radeon HD 7970 and both dual-GPU boards.



Almost ironically, given its status as an AMD Gaming Evolved title, DiRT 3 puts the Radeon HD 7950 just behind the GeForce GTX 580, even if, practically, you’d never be able to tell them apart.
Incremental performance improvements demonstrate a notable jump up from AMD’s previous single-GPU flagship, the Radeon HD 6970. However, we like the step up offered by AMD's Radeon HD 7970 more for gamers interested in high-resolutions like 2560x1600.



Without anti-aliasing applied, the Radeon HD 7950 trails to Nvidia’s GeForce GTX 580 just slightly in World of Warcraft. Adding 8x MSAA, however, hits the new card’s performance in a big way, dropping it below the much less expensive GeForce GTX 570.
Perhaps the most interesting results in this test come from AMD’s Radeon HD 6990. Processing overhead keeps the dual-GPU board from well. There’s clearly plenty of GPU muscle left in reserve. We can see this by looking at the scores at 1680x1050 and 1920x1080. When anti-aliasing is bumped up to 8x, the 6990 takes hardly any hit at all. At 2560x1600, it climbs up to second place when the high resolution impacts the competition more severely.



The Radeon HD 7950 re-assumes its position between the Radeon HD 7970 and GeForce GTX 580, serving up playable performance at up to about 1920x1080 using Very High detail settings and DirectX 11.
The Radeon HD 6990 and GeForce GTX 590 rule at the top of this test. But AMD’s single-GPU Radeon HD 7970 comes very close to usurping the Nvidia card, especially.

Our first Sandra 2012 chart measures 32-bit floating-point and 64-bit double performance. More ALUs and higher clock rates earn the Radeon HD 7970 a first-place finish here, followed by the dual-GPU Radeon HD 6990. It’s quite impressive, then, that the less expensive, cooler, and more power-friendly Radeon HD 7950 does as well as we observe. The 7950’s double-precision performance is roughly three times higher than Nvidia’s GeForce GTX 580.

The bandwidth chart measures two things: throughput between the GPU and memory (blue bars) and transfer rate between the GPU and host over PCI Express.
Presumably, Sandra is counting the bandwidth of both GPUs in aggregate as it reports the numbers for AMD’s Radeon HD 6990 and Nvidia’s GeForce GTX 590. Both Radeon HD 7900-series cards follow closely in third and fourth place, though.
The data transfer bandwidth results are a little harder to discern next to the big memory bars. However, it’s quite clear that the cards with second-gen interfaces are stuck around 6 GB/s or thereabout, while both Radeon HD 7900-series boards are able to exceed 9 GB/s by virtue of PCI Express 3.0.

This was perhaps the most disappointing chart we encountered. CyberLink generally does a great job of supporting new GPU-based technologies, and we’ve used special versions of its MediaEspresso software in the past to analyze the visual quality of AMD’s accelerated output.
However, we tried three versions of the software: the latest demo, the latest press room-based review copy, and a third with certain features that we can’t disclose yet. None of them made hardware-accelerated encoding available on the AMD cards. Decoding was available, but as you can see, it’s faster to disable AMD's UVD 3 block and use Intel’s Core i7-3960X for the job instead.
Although MediaEspresso also defaults to using accelerated encode-only on the Nvidia cards, we turned on decoding as well. The resulting speed-up is noticeable compared to the software-only approach.
The biggest bummer is that AMD still hasn’t enabled the Video Codec Engine purportedly built into its Radeon HD 7900-series cards. This fixed-function hardware is supposedly not any faster on its own than encoding collaboratively with the programmable hardware. However, in light of the fact that we can’t use the shader hardware on its own, even with the latest Media Codec Package installed, AMD’s performance story in MediaEspresso ends with a whimper.

AMD’s OpenCL performance under LuxMark is much more promising. This tool, derived from the LuxRenderer engine, is a real-world representation of OpenCL-based performance that reports its result in samples per second.
Don’s Radeon HD 7970 launch coverage mentioned that GCN was architected to greatly improve compute performance, and these numbers validate that claim by blowing right past Nvidia’s dual-GPU GeForce GTX 590. The single-GPU Radeon HD 7970 nearly matches pace with the dual-GPU Radeon HD 6990.

Arguably a less scientific metric, we’re using vReveal to demonstrate the impact of GPU acceleration on a workload that would have been processor-bound in the past.
We’re playing back a 1080i video clip and applying the One-Click Fix feature. You can see that, with GPU acceleration applied, all of the tested cards achieve a full 30 FPS (that is to say vReveal does its work in real-time). An overclocked processor tackling the workload on its own cannot maintain the same frame rate.

But just because the GPUs all deliver 30 FPS doesn’t mean they are equally effective at offloading our metric, though. The AMD cards relieve the overclocked processor just a little bit more than Nvidia’s.
While it may not be as sexy as 3D performance, 2D rendering is still important. While there is a clear trend towards rendering 2D content using Microsoft’s more modern Direct2D API, it’s a safe bet that more than 90 percent of all applications in use today still rely on the drawing functionality provided by the older GDI (Graphics Device Interface) and GDI+. Most user interface elements, such as frames, buttons, and toolbars, are rendered using these components, though. Meanwhile, older programs created for very specific purposes rely completely on this rendering method for all of their 2D objects. That’s why we decided to test 2D performance as well.
Text Output
As in our Radeon HD 7970 coverage, the two Tahiti-based cards are the only ones that stumble with regard to direct (hardware-accelerated) output to the display, and not buffered and unaccelerated output in the form of a DIB (device-independent bitmap). It looks like nothing has changed in this metric since we tested with the launch driver.
Again, it’s unlikely that this score will result in a tangible real-world performance hit. However, it’s interesting that the older Radeons do a lot better. Our guess is that hardware acceleration for direct text rendering is still immature, since that score is even slower than the non-accelerated software solution using a DIB. Obviously, a little driver tweaking is in order, even if we're using the third driver revision seen since the 7970's debut.
The Radeon HD 7950 only achieves half of the text output of a Radeon HD 6970, putting it on par with the integrated GeForce 7025 GPU found in Nvidia's aging nForce 610i chipset.

Image Manipulation
Surprisingly, the Radeon HD 7950 is faster than the 7970 in our stretching test's direct output mode, reaching scores similar to those of the Radeon HD 6970. At the same time, we see that the cards perform better in software mode across the board.
Meanwhile, simple copy operations (blitting) show very little variation between cards, and only the GeForce GTX 580 is faster taking the direct route than using the detour through the buffer (a clear sign that hardware acceleration is being used more efficiently). Once again, AMD’s software implementation of this previously hardware-accelerated feature still needs some work.


Geometry Performance
Both the Radeon HD 7970 and 7950 fall behind by a small margin when drawing lines. The remainder of our benchmarks show all of the contenders performing quite similarly, though.
Both splines and rectangles are apparently accelerated quite well when they are rendered sequentially, since the direct output path is faster than the software version in either case. This performance delta is especially apparent in the triangles test. The exact opposite applies when it comes to drawing polygons, where buffered output is much higher.





AMD has certainly improved its drivers since the first time we took a closer look at 2D performance, though the Radeon HD 7900s continue falling behind previous-generation boards when tasked with hardware-accelerated text output, achieving half the performance of older cards. Stretching and drawing lines could also be faster on the new cards.
Now, in context, it’s unlikely that you’d notice any visible slow-downs in everyday tasks. The lower performance could become apparent when you move longer pieces of floating text in programs like Corel 14. The 2D situation is much-improved compared to what we saw when the Radeon HD 5870 launched, but it still needs some work.

Two Radeon HD 7950s in CrossFire maintain their lead over a pair of GeForce GTX 580s. A couple of Radeon HD 7970s are also pretty awesome-looking though, especially when you consider the only two faster configurations employ four GPUs each!





Hold onto your collective rear-ends. A second Radeon HD 7950 adds 96% to the single-card result without anti-aliasing and 98% when you apply 4x MSAA, making Battlefield 3 more than playable at 2560x1600 with Ultra quality settings.

Knowing what we know about the 7950’s weakness in Crysis 2 under DirectX 9, you’re going to want to stick with DirectX 11. A second Radeon HD 7950 increases performance by 92%. That’s enough of a speed-up to wiggle past a pair of GeForce GTX 580s and absolutely destroy the four processors wielded by two Radeon HD 6990s.
We couldn’t come up with any good reason to explain the quad-GPU failure under DirectX 11, but a far better DirectX 9 result makes it pretty clear that the problem is related to API choice.

A second Radeon HD 7950 only adds 86% to the performance of just one GPU. That doesn’t change the overall finishing order, though. Two GeForce GTX 580s are faster, trumped by a pair of Radeon HD 7970s and then both of the quad-GPU configurations.

Although two dual-GPU Radeon HD 6990s have an edge on the Radeon HD 7950 CrossFire setup with 4x MSAA applied, the 7950s are actually faster using the Adaptive Anti-Aliasing mode. In the process, AMD’s latest skates right past two GeForce GTX 590s and 580s. Only the Radeon HD 7970s are faster.

Two Tahiti XT GPUs nearly match a quartet of Cayman processors in LuxMark. Scaled down to the Pro variant, you lose quite a bit of compute performance. However, a pair of Tahiti Pros still decimates four GF110s in this benchmark.
Power

When most folks set up their machines, Windows 7, by default, is configured to shut off attached displays after a certain period of time and then put the entire system to sleep at some point thereafter. When we test, we don’t want those variables affecting our results, though, so we turn it all off.
Thus, the power numbers in the chart above reflect idle power consumption sitting on the Windows desktop—and they’re impressive to say the least. AMD’s single-GPU boards—current-generation and last-generation—prove themselves to be the best in this metric.
But the company went one step further with the ZeroCore feature set we introduced in AMD Radeon HD 7970: Promising Performance, Paper-Launched. The result of ZeroCore is even lower power draw once output to the display is cut.

The Radeon HD 7950 leads the way, with an idle system power consumption just over 93 W (that’s a Core i7-3960X-powered platform with 16 GB running at 4.2 GHz, remember).
If you think that’s cool, it gets better on the next page.

Clearly, AMD’s Radeon HD 6990 and Nvidia’s GeForce GTX 590 are the highest power consumers. Comparatively, the Radeon HD 7950, which averages 315 W of system power use across this test to the 6990’s 450 W and the GTX 590’s 490 W, is downright eco-friendly.
Temperature

The two 28 nm Tahiti-based GPUs run significantly cooler than the 40 nm competition at idle.

Whereas most of the cards tested in a Metro 2033 loop are cooled in a way that keeps them from exceeding 86 degrees C, the Radeon HD 7900s want to stay in the 70s—and their fans blow hard enough to make sure that happens.
Noise

From exactly one meter behind each card’s rear I/O port, idle noise is fairly similar. In fact, although we went so far as to use a closed-loop liquid cooler for its quiet operation, the unit’s built-in pump was the loudest component on our test bench.

In an effort to keep each card within preset thermal limits, these boards all employ different fan speed profiles.
The Radeon HD 6990 is notoriously obtrusive. Despite AMD’s claims that we ended up with a defective sample, we have two of these things running in the lab, and they’re both noisy beasts.
Despite lessons we would have hoped AMD's engineers would have learned from the Radeon HD 6990, though, and even though this new architecture doesn’t get particularly hot under a normal load, the Radeon HD 7970’s fan is also disturbingly loud. It’s not as outright-obscene as the dual-GPU card. However, the company’s design guys could certainly take a hint or two from the effort that went into Nvidia’s GF110-based boards. Even the power-hungry GeForce GTX 590 runs quieter than the new Radeon HD 7950 on display today.
Power

Oddly, two Radeon HD 7970s use just slightly less power at idle than two Radeon HD 7950s. The point is, however, that they both draw a lot less than any other combination of cards.
And just look at those two GeForce GTX 590s sitting there, slurping down a whopping 84 W more than the 7970s on Windows' desktop!
If you flip back and forth between this page and the previous one, you'll notice that each 7900-series CrossFire setup is only using around the same amount of idle power as a single-card arrangement. That’s because they’re doing this:

That’s one card idling (in the back), its PCI Express 3.0 link throttling to 1.1 transfer rates to cut power, as the other card (up front) completely spins down.

As a result, you get even more impressive power results when output is cut to the attached display. The Radeon HD 7950 and 7970 CrossFire configs each shed an additional 6 W of consumption.

For their ability to deliver the fastest Metro 2033 results in CrossFire, beating quad-GPU Radeon HD 6990 and GeForce GTX 590 configurations, the Radeon HD 7970 and 7950 are also the two most power-friendly choices cards you could put together in multi-GPU arrangements.
And that’s under load. We’ve already seen how much better they are at idle. If you were to calculate out the efficiency of these cards versus older AMD and Nvidia boards, they’d almost certainly stomp everything else out there. Performance per watt is where it’s at.
Noise

The Radeon HD 7900s again show well when it comes to keeping idle acoustic output to a minimum. Nvidia’s cards are similarly silent. And although the Radeon HD 6990s come in last place, they’re still hardly a distraction.

As long as you have a full expansion slot separating two dual-GPU cards, the noise generated by a CrossFire or SLI configuration isn’t markedly higher than one board operating on its own in the real-world gaming environment we’re testing.
AMD is using a centrifugal fan design that exhausts all of its heated air, which we applaud.
The only odd result here is the Radeon HD 6990s, which operate more quietly in a quad-CrossFire arrangement than one card flying solo. Fortunately, the explanation is pretty simple. Compared to two GeForce GTX 590s, which are fully taxed in Metro 2033, the 6990s hover between 60 and 80% GPU utilization according to GPU-Z.
Our presumption is that host overhead is higher, so we’re seeing a bottleneck that inhibits performance, which would explain why one 6990 was able to outmaneuver a 7970 at 2560x1600, but even two 7950s beat a pair of 6990s at the same resolution.
AMD’s sweet spot strategy, implemented under Rick Bergman, was a very successful way for the company’s graphics team to compete aggressively with regard to performance and power consumption. However, it relied on an effectively-executed dual-GPU solution in order to capture the high-end market. Of course, that approach presented its own challenges—from issues with CrossFire scaling to the quick implementation of multi-GPU game profiles to, more recently, hot, loud cards like the Radeon HD 6990.
Transitioning to 28 nm manufacturing will likely help AMD design this generation’s dual-GPU card—and we can only hope acoustics receive plenty of attention.
More immediately, though, AMD’s single-GPU Radeon HD 7950 holds its own admirably in the high-end space, trading blows with Nvidia’s year-old GeForce GTX 580 in most of the games we tested, cruising right past the GF110-based board in some titles and giving up ground in others.
That the 7950 matches pace with Nvidia's GeForce GTX 580 doesn't impress us much, given the competing card's age. However, AMD does several other things very much right with the Radeon HD 7950. Compute performance, for example, appears to be significantly improved compared to the company’s prior-generation product (and indeed its competition). Innovations in power consumption translate to guilt-free multi-card configurations as second (and reportedly third and fourth) GPUs spin down when they’re idle. CrossFire scaling also looks to be commendable right out of the gate. Those are all attributes that help AMD's second-fastest single-chip solution stand above its most obvious nemesis in Nvidia's portfolio.
Is everything perfect on the beaches of Tahiti? Decidedly not. We weren’t able to get hardware-accelerated encoding working in one of the leading transcode apps, and despite having access to a preview build of an unnamed piece of software with VCE support, AMD’s latest drivers still don’t expose support (more than a month after the Radeon HD 7970 launched, even). Plus, even in the face of downright-chilly load temps, the Radeon HD 7950’s blower generates more noise than the dual-GPU GeForce GTX 590 running through loops of Metro 2033.
Ultimately, enthusiasm for the Radeon HD 7950 is going to be determined by the price at which AMD sells it (Ed.: which we now know should be somewhere around $450) and the amount of supply seen between now and whenever Nvidia is able to answer back. AMD is peddling one single-GPU card that’s faster than Nvidia’s best effort and another that’s comparable. If you want more performance, from an individual card, at least, you'll have to buy a Radeon HD 6990 or GeForce GTX 590. And gosh, I still don’t care for either of those two options.
The cheapest GeForce GTX 580 goes for about $470, while Radeon HD 7970s have crept up $10 to $560. I’d totally be willing to pay a $10 or $20 premium over the 580 for AMD’s new Radeon HD 7950, if only because of the compute potential and power consumption advantage. But I’d also want accelerated encoding to work, and I’d like to see the potential speed-up with VCE factored in.
Let’s see if AMD can get this card’s full suite of capabilities up and running before it has more compelling competition. At an anticipated price tag of $450, it's a cheaper, cooler, and more power-friendly alternative to GeForce GTX 580. In games, there's really no contest in a decision between the two.





