Announced late last month and shipping 3 weeks ago, AMD kicked off the 28nm generation with a bang with their Radeon HD 7970. Combining TSMC’s new 28nm HKMG process with AMD’s equally new Graphics Core Next Architecture, AMD finally took back the single-GPU performance crown for the first time since 2010 with an all-around impressive flagship video card.

Of course AMD has always produced multiple video cards from their high-end GPUs, and with Tahiti this was no different. The second Tahiti card has been waiting in the wings for its own launch, and that launch has finally come. Today AMD is launching the Radeon HD 7950, the cooler, quieter, and cheaper sibling of the Radeon HD 7970. Aimed right at NVIDIA’s GeForce GTX 580, AMD is looking to sew up the high-end market, and as we’ll see the Radeon HD 7950 is exactly the card to accomplish that.

AMD GPU Specification Comparison
	AMD Radeon HD 7970	AMD Radeon HD 7950	AMD Radeon HD 6970	AMD Radeon HD 6950
Stream Processors	2048	1792	1536	1408
Texture Units	128	112	96	88
ROPs	32	32	32	32
Core Clock	925MHz	800MHz	880MHz	800MHz
Memory Clock	1.375GHz (5.5GHz effective) GDDR5	1.25GHz (5GHz effective) GDDR5	1.375GHz (5.5GHz effective) GDDR5	1.25GHz (5GHz effective) GDDR5
Memory Bus Width	384-bit	384-bit	256-bit	256-bit
Frame Buffer	3GB	3GB	2GB	2GB
FP64	1/4	1/4	1/4	1/4
Transistor Count	4.31B	4.31B	2.64B	2.64B
PowerTune Limit	250W	200W	250W	200W
Manufacturing Process	TSMC 28nm	TSMC 28nm	TSMC 40nm	TSMC 40nm
Price Point	$549	$449	$350	$250

As has been the case for AMD since the 5000 series, AMD has gone with a two-pronged approach to binning and cutting down their flagship GPU for their second-tier card. The first change is an across-the-board reduction in clockspeeds, with the core clock being dropped from 925MHz to 800MHz and the memory clock being dropped from 5.5GHz to 5GHz. The second change is that the shader count has been reduced from a full 2048 SPs to 1792 SPs, accomplished by disabling 1 of the GPU’s 8 CU arrays and allowing AMD to use Tahiti GPUs with a defective CU array that would have never worked in the first place.

No other changes have been made, a particularly important consideration since it means all 32 ROPs and the 6 64bit memory channels are still in place. Altogether this gives the 7950 86% of the ROP throughput, 75% of the shader and texture throughput, and 91% of the memory bandwidth of the 7970. This should put the 7950 in direct competition with NVIDIA’s GeForce GTX 580, which typically trails the 7970 by a similar degree. Otherwise compared to the 6000 series, this makes the core performance gap between the 7950 and 7970 a bit bigger than between the 6970 and 6950, while the memory bandwidth gap is identical.

The tradeoff of course on a second-tier part is that while performance has been reduced so has power consumption. Just as with the 7970, the 7950 takes after its 6000-series predecessor, shipping with a 200W maximum board power limit. With the 7000 series AMD has not been publishing any kind of typical power numbers and thereby making the board power limit the only number they publish, but also making for a far more accurate TDP than past estimated TDP numbers as it’s an absolute limit. For gaming scenarios you’re almost always looking at less than 190W power consumption, though the spread between typical power and the PowerTune cap is not as wide on the 7950 as it was the 7970. Meanwhile for idle power consumption AMD is not providing an official number there either, but with the use of power islands the difference in idle power consumption between various core configurations has been virtually eliminated. Idle TDP should be 15W, while long idle is 3W.

In a bit of an unusual move for AMD, for the 7950 they are doing away with reference designs entirely. All 7950s will be custom to some degree—the first run will use a partner’s choice of cooler alongside a new PCB from AMD specifically for the 7950, while in the future partners will have the option of going fully custom. Furthermore partners will be shipping factory overclocked parts from right out of the gate, and at this point we’re not even sure just how many models will actually be shipping at stock clocks; neither MSI or Sapphire have a stock clocked card as part of their lineup. Overall at the low-end we’re seeing overclocked cards shipping as low as 810MHz, while 900MHz is particularly common at the high-end.

The use of customized factory overclocked cards is not unusual for AMD’s lower-end cards, but this is the first time we’ve seen AMD’s partners launch factory overclocked parts out of the gate like this, and it’s the first time we’ve seen AMD launch a part over $200 without a reference cooler. As a result the 7950 will be a true Your Mileage May Vary situation, with the gaming performance and physical performance characteristics depending heavily on how a partner has configured their card.

Radeon HD 7950 Partner Specification Comparison
	AMD Radeon HD 7950 (Stock)	Sapphire HD 7950 Overclock Edition	XFX R7950 Black Edition Double Dissipation
Stream Processors	1792	1792	1792
Texture Units	112	112	112
ROPs	32	32	32
Core Clock	800MHz	900MHz	900MHz
Memory Clock	1.25GHz (5GHz effective) GDDR5	1.25GHz (5GHz effective) GDDR5	1.375GHz (5.5GHz effective) GDDR5
Memory Bus Width	384-bit	384-bit	384-bit
Frame Buffer	3GB	3GB	3GB
FP64	1/4	1/4	1/4
Transistor Count	4.31B	4.31B	4.31B
Manufacturing Process	TSMC 28nm	TSMC 28nm	TSMC 28nm
Warranty	N/A	2 Years	Lifetime
Price Point	$449	$479	$499

For the launch of the 7950 AMD shipped us a pair of internal reference cards built on the 7970 PCB and cooler. Since no one will actually be shipping a card like this—although they technically could if they wanted to—we also went looking for partner cards, which XFX and Sapphire provided. The XFX R7950 Black Edition Double Dissipation and Sapphire HD 7950 Overclock Edition are far more representative of what we’re actually going to see on the market; factory overclocks aside, both use open air coolers, just as with every other 7950 card we’ve seen the specs for ahead of today’s launch. Given the lack of any cards using fully exhausting blowers, it would appear that AMD and their partners have become particularly comfortable with open air coolers for 200W cards.

Last but not least of course, is pricing. AMD is continuing their conservative pricing strategy of trying to price their cards against existing competitive cards, rather than using the cost savings of the 28nm process to bring down prices across the board. As a result the 7950 is priced at $449, $100 below the 7970 and almost directly opposite the cheapest GeForce GTX 580s, making the 7950 a de facto GTX 580 competitor. This pricing strategy seems to have worked well for the 7970—cards are still selling at a brisk pace, but the shelves are rarely completely bare—and it looks like AMD is going to continue following it while they can. Meanwhile the fact that the 7950 is an entirely semi-custom lineup means that pricing is going to be equally variable, with high-end factory overclocked cards such as the Sapphire and XFX going for $479 and $499 respectively.

Winter 2011 GPU Pricing Comparison
AMD	Price	NVIDIA
	$750	GeForce GTX 590
Radeon HD 6990	$700
Radeon HD 7970	$549
Radeon HD 7950	$450+	GeForce GTX 580
Radeon HD 6970	$350	GeForce GTX 570
Radeon HD 6950 2GB	$250
	$240	GeForce GTX 560 Ti
Radeon HD 6870	$160

 

Getting the Most Out of GCN: Driver Improvements

With the launch of any new architecture there’s still a lot of room for improvement on the part of driver developers, and GCN is no exception. On January 20^th AMD released the first driver update for the 7000 Series, which brought with it an interesting mix of bug fixes, new features, and performance improvements. On the feature side AMD enabled support for Analytical Anti-Aliasing and Super Sample Anti-Aliasing for DX10+ games, an overdue feature that we’re very happy to see finally make it to AMD cards. Meanwhile on the performance side the new drivers improved the performance of the 7000 series in several games. Game performance typically rises slowly over time, but as this is one of the first post-launch driver releases, the gains are larger than what we’re used to seeing farther down the line.

To get an idea of where performance has improved and by how much, we reran our entire benchmark suite on the 7970.

As to be expected, at this point in time AMD is mostly focusing on improving performance on a game-by-game basis to deal with games that didn’t immediately adapt to the GCN architecture well, while the fact that they seem to be targeting common benchmarks first is likely intentional. Crysis: Warhead is the biggest winner here as minimum framerates in particular are greatly improved; we’re seeing a 22% improvement at 1920, while at 2560 there’s still an 11% improvement. Metro:2033 and DiRT 3 also picked up 10% or more in performance versus the release drivers, while Battlefield 3 has seen a much smaller  2%-3% improvement. Everything else in our suite is virtually unchanged, as it looks like AMD has not targeted any of those games at this time.

As one would expect, a result of these improvements the performance lead of the 7970 versus the GTX 580 has widened. The average lead for the 7970 is now 19% at 1920 and 26% at 2560, with the lead approaching 40% in games like Metro that specifically benefited from this update. At this point the only game the 7970 still seems to have trouble pulling well ahead of the GTX 580 is Battlefield 3, where the lead is only 8%.

Meet the Sapphire HD 7950 Overclock Edition

Since our reference 7950s are built on the 7970 PCB and cooler, we’re going to jump right into our vendor cards starting with the Sapphire HD 7950 Overclock Edition.

As with all of the 7950 cards launching today, Sapphire’s HD 7950 Overclock Edition uses the AMD 7950 PCB. This is a slightly shorter PCB measuring 10.25” long, saving .25” over the 7970 PCB by eliminating a few components that the lower board power of the 7950 makes unnecessary. The PCB is otherwise very similar to the 7970 PCB, utilizing 12 GDDR5 memory chips organized around the Tahiti GPU, while at the top you’ll find the 2 CrossFire connectors, a pair of 6pin PCIe power sockets, and the BIOS selection switch. The latter will be of particular interest to unlockers, as the switch should make it possible to safely attempt to unlock the 7950 into a 7970.

Moving on, as this is a semi-custom card the real differentiation is in the factory overclock and the cooler. On the performance side of things Sapphire will be shipping the 7950 Overclock Edition at 900MHz core and 5GHz memory, representing a 100MHz (12.5%) core overclock and no change on the memory clock.

Meanwhile for the cooler Sapphire is using what they’re calling the Dual-X cooler. The Dual-X is yet another double-wide dual-fan open air cooler, with 2 fans providing copious airflow over an aluminum heatsink running virtually the entire length of the card. Sapphire’s fan cutouts are just a bit bigger than most other dual-fan coolers and placed a bit higher, and as a result the Dual-X cooler is a bit taller than the PCB by about 15mm at its highest point. Meanwhile the cooler is also a fair bit longer than the PCB, putting the total card length at 11”.

Moving below the fans and the heatsink we’ll find the heatpipe assembly, which is responsible for carrying heat from the GPU to the heatsink. The Dual-X uses 5 copper heatpipes of varying radius that run from one end of the heatsink to the other. The 5 heatpipes converge at the base of the assembly, where a copper baseplate provides contact with the GPU. Meanwhile cooling for the VRM MOSFETs and RAM is provided by a black aluminum plate, which is placed over those components with heat transfer provided by the use of thermal pads. There is no connection between the plate and the heatsink, so the only heat dissipation from the plate is provided by whatever airflow from the fans reaches the plate.

At the front of the card we’ll find the display ports, which as this is an AMD PCB the card utilizes the standard AMD 7000 series port configuration of 1 DL-DVI port, 1 HDMI port, and 2 mini-DisplayPorts. Filling out the second slot is the grating for ventilation, though even with the ventilation slot the usual precautions for an open-air cooler apply: you’ll need a case with enough airflow to handle the roughly 200W of heat the card is capable of dumping inside of your case.

Rounding out the package is the usual collection of dongles and materials. Sapphire includes 2 molex-to-6pin PCIe adaptors, an HDMI to DVI dongle, a miniDP to DisplayPort dongle, a DVI to VGA dongle, and a 1.8m HDMI cable. Along with the dongles Sapphire packs a quick start guide and a driver installation CD.

The only thing you won’t find packed in the box is TriXX, Sapphire’s in-house overclocking utility. TriXX has been around since the 6900 series, but as this is the first high-end Sapphire card we’ve reviewed since it was released, this is the first time we’ve had it available for a review.

Fundamentally TriXX is a fairly well designed, albeit barebones overclocking utility. Along with an info readout similar to GPU-Z, TriXX provides overclocking and fan control support for Sapphire’s cards, including support for custom fan profiles and more importantly voltage control. With TriXX it’s possible to overvolt most of Sapphire’s performance and high-end cards, and as Sapphire uses AMD reference PCBs it also works with any other cards using AMD’s PCBs.

Beyond these features there’s little more to TriXX. It’s not an all-encompassing video card utility like MSI’s Afterburner, which means it comes up short if you need more functionality but it's exactly what you need if you just want to overclock. To that end it’s a clear step up compared to most other manufacturer’s poorly designed utilities, and from a design perspective its only real sin is the hard to read blue-on-black text. Otherwise it’s a competent overclocking utility that does exactly what it’s supposed to and provides voltage control for those who need it.

Finally, Sapphire will be selling the 7950 OE for $479, $30 over the baseline 7950 MSRP. Meanwhile the warranty on their card is their standard 2 year warranty.

Meet the XFX R7950 Black Edition Double Dissipation

Our second partner card of the day is XFX’s R7950 Black Edition Double Dissipation. Like the 7970 Black Edition Double Dissipation we reviewed earlier this month, the 7950 BEDD is a factory overclocked card (Black Edition) using XFX’s Double Dissipation cooler.

Starting with the overclock, XFX is shipping the 7950 BEDD with a core clock of 900MHz and a memory clock of 5.5GHz. This represents a 100MHz (12.5%) core clock overclock and 500MHz (10%) memory clock overclock, putting it just shy of the 925MHz core clock the 7970 ships at.

In terms of construction like all of the other 7950 cards launching today XFX is using AMD’s 7950 PCB. This means the PCB measures 10.25” long and features 2 6pin PCIe power sockets towards the rear of the card, while at the front the card uses the AMD standard port configuration of 1 DL-DVI, 1 HDMi, and 2 mini-DisplayPorts. The one notable deviation here from the Sapphire card is that XFX has not included a BIOS selection switch, so the card lacks any kind of ability to easily recover from a bad BIOS flash, and if unlocking proves viable it would not be a good candidate for the process.

Meanwhile cooling is provided by XFX’s Double Dissipation cooler. This is the same heatsink and fan assembly we saw with the 7970 BEDD, which makes this an open air cooler using a pair of fans to push air along an aluminum heatsink running almost the entire length of the card. Because it’s the same assembly, the shrouding for the card sticks out over the end of the PCB, negating the benefit of the shorter 7950 PCB and making the card 10.65” long just like the 7970 BEDD.

Top: 7950 BEDD. Bottom: 7970 BEDD

Do note that while it uses the same fan and heatsink assembly, Double Dissipation does not mean it uses the same vapor chamber assembly to transfer heat from the card. Where the 7970 BEDD used a fairly large vapor chamber, the 7950 BEDD uses a much smaller vapor chamber that only makes contact with roughly half of the heatsink, meaning that heat isn’t being transferred to the extremities of the heatsink nearly as well on the 7950 BEDD. Furthermore the aluminum plate covering the RAM and MOSFETs is poorly sized, leaving parts of the RAM chips (and their thermal pads) exposed. We’ll see how this plays out when we get to our testing, but the 7950 BEDD is clearly not as well built as the 7970 BEDD.

Rounding out the package is the same collection of extras that we saw in the 7970 BEDD. Inside you’ll find the usual driver CD and quick start guide, along with a metal XFX case badge, a mid-length CrossFire bridge, and a passive HDMI to SL-DVI adaptor. All of this is packed in one of XFX’s pleasantly small boxes, which doesn’t use much more space than the card itself.

The MSRP on the 7950 BEDD is $499, $50 over the MSRP for a regular 7950 and making it one of the more expensive 7950s launching today. XFX is offering a base 2 year warranty on the 7950 BEDD, which can be extended to a lifetime warranty by registering the card within 30 days of purchasing it.

The Test

AMD’s shipping drivers for the 7950 are 8.921.2-120119a, the same driver version released as the first 7970 driver update two weeks ago. We have updated our 7970 results with these new drivers, everything else remains unchanged.

CPU:	Intel Core i7-3960X @ 4.3GHz
Motherboard:	EVGA X79 SLI
Chipset Drivers:	Intel 9.​2.​3.​1022
Power Supply:	Antec True Power Quattro 1200
Hard Disk:	Samsung 470 (256GB)
Memory:	G.Skill Ripjaws DDR3-1867 4 x 4GB (8-10-9-26)
Case:	Thermaltake Spedo Advance
Video Cards:	AMD Radeon HD 7970 AMD Radeon HD 6990 AMD Radeon HD 6970 AMD Radeon HD 6950 AMD Radeon HD 5870 AMD Radeon HD 5850 AMD Radeon HD 4870 AMD Radeon HD 3870 NVIDIA GeForce GTX 590 NVIDIA GeForce GTX 580 NVIDIA GeForce GTX 570 NVIDIA GeForce GTX 470 NVIDIA GeForce GTX 285 NVIDIA GeForce 8800GT
Video Drivers:	NVIDIA ForceWare 290.53 Beta AMD Catalyst Beta 8.921.2-120119a
OS:	Windows 7 Ultimate 64-bit

 

Crysis: Warhead

Kicking things off as always is Crysis: Warhead. It’s no longer the toughest game in our benchmark suite, but it’s still a technically complex game that has proven to be a very consistent benchmark. Thus even 4 years since the release of the original Crysis, “but can it run Crysis?” is still an important question, and the answer continues to be “no.” While we’re closer than ever, full Enthusiast settings at a 60fps is still beyond the grasp of a single-GPU card.

AMD’s first round of driver optimizations have given the 7900 a very solid footing in Crysis, putting the 7950 off to a great start. The 7950 is 19% ahead of the GTX 580 at 2560 and 14% ahead at 1920, putting the card in a comfortable position that for single-GPU cards is second only to the 7970. In spite of Crysis being shader-bound most of the time the 7950 is generally within 15% of the 7970, so it’s doing better than the theoretical performance gap between the two cards would predict. Meanwhile compared to AMD’s last generation offerings it’s not much of a contest: the 7950 is 20-25% ahead.

As for our factory overclocked Sapphire and XFX cards, they further close the gap between the 7950 and 7970. The 12.5% core overclock on these cards puts them between 7% and 10% faster than the stock clocked 7950, with the XFX card edging out the Sapphire due to its memory overclock. These cards do so well here than the reference 7970’s lead is reduced to just 5%.

The minimum framerates in Crysis are also looking good on the 7950, with the 7950 turning in a 10-22% better minimum framerate than the GTX 580 depending on whether we’re talking about 1920 or 2560. As like we saw with the 7970, the biggest lead is at the highest resolutions, which has typically been the case for AMD cards for some time now. The overclocked partner cards add to this, tacking on an extra 5-10% in performance.

Metro: 2033

Paired with Crysis as our second behemoth FPS is Metro: 2033. Metro gives up Crysis’ lush tropics and frozen wastelands for an underground experience, but even underground it can be quite brutal on GPUs, which is why it’s also our new benchmark of choice for looking at power/temperature/noise during a game. If its sequel due this year is anywhere near as GPU intensive then a single GPU may not be enough to run the game with every quality feature turned up.

The story in Metro is much the same as it was in Crysis, with the 7950 sailing past the GTX 580. At 2560 this is a 20% performance lead, enough to push the 7950 above 30fps, while at 1920 it’s still holding on to a 15% advantage. The factory overclocked cards further widen the gap by 7-9% for the Sapphire and XFX cards respectively. Overall the 7950 generally trails the 7970 by 15%, while the factory overclocked cards close this gap by nearly 2/3rds.

DiRT 3

For racing games our racer of choice continues to be DiRT, which is now in its 3rd iteration. Codemasters uses the same EGO engine between its DiRT, F1, and GRID series, so the performance of EGO is has been relevant for a number of racing games over the years.

With DiRT 3 AMD’s strong start finally starts to wobble. Even with their first round of driver optimizations the 7950 can only outrun the GTX 580 at 2560, while at 1920 the two cards are effectively tied. DiRT games have traditionally done better on NVIDIA’s cards so this is not unexpected, but it’s also a sign that the 7950 won’t have a clearcut lead over the GTX 580 like the 7970 did. Meanwhile the factory overclocked cards can keep the 7950 clear of the GTX 580, picking up about 8% over the 7950, but of course the tradeoff is going to be their higher cost.

The story of DiRT 3’s minimum framerates closely mirrors the average framerates. The 7950 has a 5% lead at 2560, but trails the GTX 580 by 2% at 1920.

Total War: Shogun 2

Total War: Shogun 2 is the latest installment of the long-running Total War series of turn based strategy games, and alongside Civilization V is notable for just how many units it can put on a screen at once. As it also turns out, it’s the single most punishing game in our benchmark suite.

With Shogun 2, the 7950 once again shifts into a comfortable lead over the GTX 580. At 2560 the 7950 is ahead by 7%, while at 1920 and its lower settings this grows to 17%. Relative to the 7970 the 7950 is consistently behind by 18%, indicating that Shogun 2 is largely shader bound, and as a result giving us one of the largest gaps between the 7970 and 7950 in our benchmark suite. This also means the factory overclocked cards can’t get quite as close to the 7970 as they can in other games, but at the same time it means performance generally scales very closely with clockspeed, with the XFX and Sapphire cards picking up 11% and 10% respectively.

Batman: Arkham City

Batman: Arkham City is loosely based on Unreal Engine 3, while the DirectX 11 functionality was apparently developed in-house. With the addition of these features Batman is far more a GPU demanding game than its predecessor was, particularly with tessellation cranked up to high.

With Batman the 7950 continues to lead the GTX 580, though not by as great a distance as we’ve seen in other games. At 2560 it enjoys a 6% lead, while at 1920 that lead has been cut in half to 3%. Interestingly, Batman is the opposite of Shogun when comparing the 7950 to the 7970: the 7950 is closer than average here, indicating that Batman is limited by something derived from the core clock, likely the tessellator judging from what we saw in our 7970 review. As a result overclocking is also particularly effective here, with the XFX 7950 trailing the 7970 by only 2% thanks to its identical core and memory clocks.

Portal 2

Portal 2 continues the long and proud tradition of Valve’s in-house Source engine. While Source continues to be a DX9 engine, Valve has continued to upgrade it over the years to improve its quality, and combined with their choice of style you’d have a hard time telling it’s over 7 years old at this point. Consequently Portal 2’s performance does get rather high on high-end cards, but we have ways of fixing that…

Portal 2 ends up being the first game that the 7950 fails to surpass the GTX 580 regardless of the resolution. At 2560 it only trails the GTX 580 by 2%, but at 1920 this widens to 6%, and when we enable SSAA this gap doubles to a surprisingly large 12%. The use of SSAA is especially hard on shader throughput as evidenced by the 20% gap between the 7970 and 7950, though this makes the gap between the 7950 and GTX 580 all the more curious since the 7950 has otherwise proven itself quite capable. Overclocking helps quite a bit here, with the factory overclocked cards gaining 9-13%, but it’s still not enough to push past the GTX 580 once SSAA is involved.

Battlefield 3

Its popularity aside, Battlefield 3 may be the most interesting game in our benchmark suite for a single reason: it’s the first AAA DX10+ game. It’s been 5 years since the launch of the first DX10 GPUs, and 3 whole process node shrinks later we’re finally to the point where games are using DX10’s functionality as a baseline rather than an addition. Not surprisingly BF3 is one of the best looking games in our suite, but as with past Battlefield games that beauty comes with a high performance cost

Battlefield 3 was a game the 7970 struggled with at launch, and even with AMD’s driver optimizations they haven’t been able to do a great deal about it so far. As a result the 7950 trails the GTX 580 the entire time by anywhere between 3% and 10%, and unfortunately for AMD BF3 is a very demanding game, making it one of the worst titles to fall behind at. As 2560 is not going to be playable, we’re realistically looking at 1920, where the 7950 is fast enough to crack 60fps, but is where that 10% performance gap is found.

Thankfully for Sapphire and XFX, overclocking is the great equalizer here. With their factory overclocks their cards generally erase the 10% performance gap, taking a very slight lead over the GTX 580 at both 2560 and 1920. Though there continues to be very little difference between the two cards themselves—XFX’s memory overclock is rarely worth more than a frame or two per second.

Starcraft II

Our next game is Starcraft II, Blizzard’s 2010 RTS megahit. Much like Portal 2 it’s a DX9 game designed to run on a wide range of hardware so performance is quite peppy with most high-end cards, but it can still challenge a GPU when it needs to.

As we start to reach the end of our benchmark suite, we’re once again looking at a game that the 7970 had trouble with, which doesn’t bode well for the 7950. The 7950 can take a 5% lead over the GTX 580 at 2560, but that becomes an 8% gap at 1920. Thankfully for AMD by the time we reach 1920 the framerates are over 90fps, but from a technical perspective it’s odd to see AMD struggling so much at these lower resolutions.

Civilization V

Our final game, Civilization 5, gives us an interesting look at things that other RTSes cannot match, with a much weaker focus on shading in the game world, and a much greater focus on creating the geometry needed to bring such a world to life. In doing so it uses a slew of DirectX 11 technologies, including tessellation for said geometry, driver command lists for reducing CPU overhead, and compute shaders for on-the-fly texture decompression.

As we saw in our 7970 review, Tahiti’s performance under Civ V is quite interesting—it’s well ahead of the 6000 series, more so than we would have expected. The 7950 continues this trend, keeping it relatively competitive with the GTX 580, trailing by only 5% at 2560. AMD has continued to stay mum on why their Civilization V performance on Tahiti is so good, and in light of the 14% gap between the 7970 and 7950, it looks like shader performance—particularly compute shader performance—seems to be the deciding factor. Meanwhile the factory overclocked XFX and Sapphire cards once more tilt the results in the 7950’s favor, with their performance turning the tables on the GTX 580 to take a 4-5% lead.

Compute Performance

Moving on from our look at gaming performance, we have our customary look at compute performance. Since compute performance is by definition shader bound, the 7950 is at a bit of a disadvantage here compared to gaming performance. Whereas ROP performance scales with the core clock, shader performance is hit by both the reduction in the core clock and the disabled CU array.

Our first compute benchmark comes from Civilization V, which uses DirectCompute to decompress textures on the fly. Civ V includes a sub-benchmark that exclusively tests the speed of their texture decompression algorithm by repeatedly decompressing the textures required for one of the game’s leader scenes. Note that this is a DX11 DirectCompute benchmark.

AMD’s greatly improved compute performance continues to shine here, though in the case of Civilization V it’s largely consumed by just closing the previously large gap between the GTX 500 series and the Radeon HD 6000 series. As a result the 7950 falls ever so short of the GTX 580, while the factory overclocked Sapphire and XFX cards give the 7950 enough of a push to come within 5% of the 7970.

Our next benchmark is SmallLuxGPU, the GPU ray tracing branch of the open source LuxRender renderer. We’re now using a development build from the version 2.0 branch, and we’ve moved on to a more complex scene that hopefully will provide a greater challenge to our GPUs.

Under SmallLuxGPU the 7970 enjoyed a large lead over the GTX 580, and this continues with the 7950. Even though the 7950 is well behind the 7970—to the tune of 24%—it’s still 33% ahead of the GTX 580 and the lead only grows from there. Meanwhile the XFX and Sapphire cards can catch up to the 7970 somewhat, but as this is truly a shader-bound test, you can’t make up for the lack of shaders units on the 7950.

For our next benchmark we’re looking at AESEncryptDecrypt, an OpenCL AES encryption routine that AES encrypts/decrypts an 8K x 8K pixel square image file. The results of this benchmark are the average time to encrypt the image over a number of iterations of the AES cypher.

In spite of being a compute benchmark, AESEncryptDecrypt is not particularly sensitive to GPU performance, showcasing the impact that setup times can have. The 7950 trails the 7970 by 10%, and overclocking doesn’t change this much. Unfortunately for AMD NVIDIA is still the leader here, showing that AMD’s compute performance still has room to grow.

Finally, our last benchmark is once again looking at compute shader performance, this time through the Fluid simulation sample in the DirectX SDK. This program simulates the motion and interactions of a 16k particle fluid using a compute shader, with a choice of several different algorithms. In this case we’re using an (O)n^2 nearest neighbor method that is optimized by using shared memory to cache data.

With the compute shader fluid simulation we once again shift back into a compute task that’s much more shader-bound. The 7950 only reaches 80% of the performance of the 7970, once more proving the real impact of losing a CU array. This is still enough to handily surpass the GTX 580 however, with the 7950 taking a 15% lead.

Power, Temperature, & Noise

As always, we wrap up our look at a new video card with a look at the physical performance attributes: power consumption, temperatures, and noise. With the 7970 we saw what is possible with Tahiti when AMD is trying to maximize their performance, but how will things stack up with the lower power 7950? Let’s find out.

Please note that we’ve decided to include our reference 7950 in these charts, even though none of AMD’s partners will be shipping a card in this exact configuration. Typically we wouldn’t do this, but as our partner cards are overclocked it’s the only stock performance card we have. Furthermore it provides an interesting contrast to the 7970, since they’re built on the same hardware. With that said, the Sapphire and XFX cards are far more representative of the cooling performance you’ll see with retail cards.

Radeon HD 7900 Series Voltages
Ref 7970 Load	Ref 7950 Load	XFX 7950 Black Edition DD	Sapphire HD 7950 Overclock Edition
1.17v	1.093v	1.093v	0.993v

Since the launch of the 7970 we’ve finally seen the release of tools that can read the VIDs of Tahiti GPUs. For our review we have four 7950s in our hands, and one of them is not like the others. Both of our AMD cards and our XFX card have a VID of 1.093v, however the Sapphire card is significantly lower. So low that we thought this was an error and went through 3 different tools just to be sure it wasn’t. Sure enough, our Sapphire 7950 does indeed ship at 0.993v, roughly 0.1v below our other 7950s and lower than what we thought Tahiti could operate at. The fact that the Sapphire is overclocked makes this all the more impressive, and we’ll see how this plays out in a moment.

Starting as always with idle power, there aren’t many surprises here.  Power islands not only give the 7000 series great idle power characteristics, but they also greatly diminish the difference between GPUs when it comes to idle power consumption. As a result the entire 7950 lineup is at 112W at the wall, consistently 1W below the 7970 and a good 20W below the GTX 580.

The situation with long idle power is much the same. As AMD’s ZeroCore power technology effectively powers down all the workhorse components of the GPU, the only thing left is a small constant load from the bus controller, giving us 103W on both the 7970 and 7950.

Meanwhile it’s under load that the physical performance attributes of the 7950 really begin to shine. Compared to the GTX 580, our 7950 equipped testbed draws 72W less at the wall; the performance advantage we saw in Metro earlier makes this all the more impressive. The 7950 also pulls less power than the 7970 here as we’d expect, leading to a difference of 38W at the wall.

As for our two partner cards, the Sapphire does particularly well here thanks to its very low VID, while the XFX does a bit worse than the reference 7950, likely due to its stock VID coupled with the factory overclock.

Under OCCT our 7950 cards begin to clump together in spite of their differences. With the reference 7950 at 289W, the Sapphire and XFX cards pull 5W less and 5W more respectively. It’s interesting to note here that as with the 7970 compared to the 6970, the 7950 is pulling more power than the 6950 even though both cards should have the same PowerTune limit. In practice it’s clear that the 7950 is more power hungry than the 6950 despite the PowerTune limit, so this needs to be taken into account accordingly.

As the use of power islands has brought down the idle power consumption, so has it brought down idle temperatures. The 7970 already did well here and the reference 7950 does similarly well, while the open air coolers on the XFX and Sapphire cards take this one step further, pulling temperatures down to 33C and 31C respectively.

Meanwhile load temperatures are a mixed bag among the 7950s, and it’s not all good. Starting with the Sapphire 7950, we have an open air card with a very low VID that’s performing as we’d expect. The Sapphire card is extremely cool here, and this kind of gap is consistent with other near-200W cards we’ve seen in the past. Elsewhere, in the middle is the reference 7950, which is using AMD’s fully exhausting blower; this pushes almost all the heat outside of our case, but it does keep temperatures higher, and in this case very close to the 7970.

Finally we have the XFX card, which is the odd man out. The R7950 BEDD is an open air cooler and we’d expect performance closer to the Sapphire card, not performance worse than the reference 7950.

Under OCCT the results closely mirror what we saw with under Metro. The Sapphire is well in the lead among the 7950s, followed by the reference and XFX cards.  With virtually every other card pushing 80C or higher, the 7950s are in good company here, particularly compared to the 88C GTX 580 due to the lack of comprehensive power throttling on NVIDIA’s part.

Last but not least we have our look at noise. The 7970’s cooler was already doing quite well at 40.2dB, but Sapphire takes this one step further with their open air cooler, bringing their idle noise down to 38.7dB. XFX on the other hand doesn’t fare as well, as their Double Dissipation cooler can’t rev down quite enough, coming in louder than the reference 7950 at 42.6dB.

It’s looking at load noise that particularly drives home the potential of the Sapphire 7950. Sapphire has exploited their open air cooler for all it's worth, and as a result it’s quieter under load with Metro than half our cards are at idle—41.3dB is only 0.3dB louder than the GTX 580 at idle, never mind the other 7950s. The XFX meanwhile fares really, really poorly here for an open air cooler. This doesn’t appear to be a matter of an aggressive fan profile as we’ve already seen that its temperatures are close to the reference 7950, so there’s more going on here.

Even under OCCT the Sapphire 7950 still only hits 42.8dB, a particularly impressive showing when you consider that AMD’s reference 6900 series cards idle at 42.7dB. It’s not a silent card, but for most practical purposes it’s damn close. Meanwhile the XFX 7950 again fares poorly, coming in at 56.8dB. This is still quieter than the GTX 580 in this case, but it’s clearly disappointing.

Wrapping things up, as we stated in our introduction the vast selection of 7950s means that there’s going to be a wide range of performance for 7950 cards, and these results drive home the point nicely. The Sapphire 7950 simply gushes potential—not only is it a factory overclocked card, but it gives off so little heat and noise compared to anything else in its performance class that it’s unrivaled. Much of this no doubt is due to the extremely low VID of our sample and I strongly suspect that this a better voltage than the bulk of Sapphire 7950s can actually hit, so our results are most likely better than average. Still, even with more voltage there’s still plenty of thermal and acoustic headroom for the Sapphire 7950 to work with, so it should still be a well performing card regardless.

Overall it’s results like what we’ve seen with the Sapphire that are a large part of why AMD’s partners are strongly favoring open air coolers. But with that said, that potential can only be reached in a case with good airflow, meaning that the Sapphire 7950 (or any other open air 7950 for that matter) is likely a poor choice for a cramped case. Furthermore all of these open air cards are poor choices for CrossFire if they need to be directly adjacent to each other. A lot of boards offer usable PCIe x16 slots farther away from their first x16 slot, but this is not the case for all boards. Ultimately the lack of 7950 cards using a fully exhausting blower is definitely going to limit the CrossFire potential of the 7950.

Finally, we have the XFX 7950 BEDD. The 7970 BEDD was a very impressive card; the 7950 BEDD is not. It’s not possible to completely isolate any specific reason for the 7950 BEDD’s poor performance, but the most likely factors are the GPU quality and the ability of the card to transfer heat between the GPU and heatsink. With regards to quality, if the Sapphire 7950 has among the best 7950 GPUs then the 7950 BEDD has an average GPU if not worse; 1.093v is the same as our reference cards, but there’s still room for quite a bit of variance. Whereas when we’re looking at heat transfer we have to look at the much smaller vapor chamber the 7950 BEDD uses compared to the 7970 BEDD, along with what at first glance looks like too much thermal paste.

Given our temperature and power consumption data we’re heavily favoring the heat transfer theory; the 7950 BEDD just isn’t very good at transferring heat from the GPU. This is not to say that the 7950 BEDD is a bad card—our results are reasonable for a 200W card—but we’ve seen much better in the 7970 BEDD, never mind the other 7950 cards.

Overclocking: Power, Temp, & Noise

In their marketing materials AMD is heavily pushing overclocking, and they have good reason to. With the 7970 we’ve established that Tahiti has quite a bit of overclocking headroom, and as the 7950 is clocked lower by default this opens up that headroom even further. Realistically AMD’s binning process means that the best clocking Tahiti GPUs are going to be allocated to the 7970 unless they have failed shaders, but even with that there’s quite a bit of potential on paper.

As with overclocking the 7970, our goal overclocking the 7950 is to see how much you can get for free; that is without any voltage adjustments. AMD’s reference PCBs are not particularly overbuilt for overclocking—cards like that will come later—so sticking to the reference voltage is the safest option, not to mention the easiest. With the 7970 we were able to get 200MHz (22%) overclocks without any voltage adjustment, and we’re hoping for the same out of the 7950.

With that said, we quickly ran into a wall on one card: the Sapphire 7950. Sapphire’s low VID of 0.993v may be great for temperature and noise at stock, but it’s not doing overclocking any favors. We only hit 950MHz at that voltage. As the Sapphire was the odd man out—every other card was at 1.093v—we did end up overvolting the Sapphire to 1.093v to see what it was capable of when put on similar footing as the rest of our cards.

After bringing up the voltage of our Sapphire card, all of our 7950s ended up overclocking to very similar levels. Our Sapphire and AMD cards topped out at 1025MHz core, a 225MHz (28%) overclock over a stock 7950 and a 125MHz (14%) overclock over the Sapphire’s factory overclock, while our XFX card reached 1050MHz, a 150MHz (17%) overclock beyond XFX’s factory overclock. Meanwhile the memory clocks on all of our cards topped out at 5.8GHz, beyond which we’d start seeing performance regressions from error correction on the memory bus.

Radeon HD 7950 Overclocking
	AMD Radeon HD 7950	Sapphire HD 7950 Overclock Edition	XFX R7950 BEDD
Shipping Core Clock	800MHz	900MHz	900MHz
Shipping Memory Clock	5GHz	5GHz	5.5GHz
Shipping Voltage	1.093v	0.993v	1.093v
Overclock Core Clock	1025MHz	1025MHz	1050MHz
Overclock Memory Clock	5.8GHz	5.8GHz	5.8GHz
Overclock Voltage	1.093v	1.093v	1.093v

As you can imagine, with such similar overclocks, gaming performance on all 4 cards ended up being very similar. So we’ll get to gaming performance in a minute, while we’ll start with power, temperature, & noise.

Even though we’re not increasing the voltage on our AMD and XFX cards, merely overclocking them and raising the PowerTune limit to avoid throttling does drive the power consumption up. As is typical with heavily overclocked cards, overclocking quickly drives up power consumption and the 7950s are no exception. After overclocking power consumption is almost identical to the stock 7970, so while you can get 7970 performance you still need to pay the price with 7970 power consumption. Meanwhile it’s interesting to note that even with the extra 0.1v we’ve given the Sapphire card its final power consumption is only ever so slightly higher than the other 7950s, proving that voltage is the great equalizer in this case.

With the increase in power comes an increase in temperatures. The Sapphire card still does very well here staying in the low 70s even under OCCT, while the reference and XFX cards hit the high 70s under Metro and mid 80s under OCCT. As we’ve yet to really ascertain what the thermal limits are for Tahiti, it’s not clear whether there’s too much thermal headroom left for the GPU, particularly under OCCT.

Last but not least we have load noise. The Sapphire card is once more a stellar performer, and we still can’t get it above 50dB even with OCCT. Unfortunately the XFX 7950 BEDD has its biggest fallout yet—it may be able to overclock well, but at 64dB under OCCT the performance isn’t going to be worth the immense amount of noise it creates to move enough air to keep the GPU cool.

Overclocking: Game & Compute Performance

We’ll keep the commentary thin here, but overall the overclocked performance of the 7950s looks very good. The XFX 7950 BEDD has the lead among the 7950s due to its slightly higher overclock of 1050MHz, while the 7950s as a whole enjoy anywhere between a 2% to 10% lead over the 7970. It’s clear that if you want a 7970’s gaming performance at a slightly lower price then the 7950 can deliver on that through overclocking—and there seems to be little reason not to pursue it—but you’re going to have to pay the price on power consumption to get there.

The compute performance gap on the other hand can’t be closed quite as easily. Overclocking can help, but if you need a 7900 series card for a heavy compute workload there’s no making up for the lost CU array; you’d still need a 7970.

Final Words

When AMD launched the Radeon HD 7970 last month there was a great deal of speculation that the Radeon HD 7950 would be their direct GeForce GTX 580 competitor, and indeed this has proven to be the case. While the 7970 sails past the GTX 580—and AMD has priced it based on that—the 7950 and the GTX 580 are trading blows on a game-by-game basis, similar to what we saw last year in comparing the GTX 500 series and the Radeon HD 6900 series. But when the 7950 wins it wins big, while the same cannot be said of the GTX 580; the only real weakness for the 7950 right now is Battlefield 3, and while that’s an important game it’s but one of several.

Ultimately it’s not a fair fight, not that AMD ever intended it to be one. Outside of a few corner cases the 7950 renders the GTX 580 irrelevant, and while it’s not quite as immense as what the 5850 did to the GTX 285 2 years ago the outcome is much the same. With the 7950 AMD can deliver performance similar to if not better than the GTX 580 while consuming significantly less power and enjoying all the temperature & noise benefits that provides, making it a very attractive card.

On that note the cooling situation makes the launch of the 7950 one of the more unusual high-end product launches in recent history. With high-end cards typically sticking to reference designs for the first phase of their lives the 7950 lineup is going to be much more varied than normal, not only in gaming performance due to factory overclocks but in cooling performance too. While we can speak in absolutes about the gaming performance of the 7950 there is no common thread on cooling performance—it needs to be evaluated on a per-product basis, so it will be important to do your research.

Meanwhile the $450 price tag is unfortunately not very aggressive on AMD’s part, but with their lead in rolling out their new lineup this is to be expected. Given its performance the 7950 only needs to be as cheap as the cheapest GTX 580 and that’s exactly what AMD has done. There will ultimately be a massive price shakeup at the high-end due to 28nm, but this looks like it won’t happen until AMD has some competition at 28nm or 7900 sales slow down significantly.

Finally, what about our retail sample cards, the XFX R7950 Black Edition Double Dissipation and the Sapphire HD 7950 Overclock Edition? These two cards clearly embody the type of variety we’re going to see from AMD’s partners; they have fairly large factory overclocks and large open air coolers, and with these customizations AMD’s partners are hoping to set themselves apart from each other while justifying a higher MSRP in the process.

Overall the Sapphire HD 7950 Overclock Edition is the clear winner among the two cards. While I believe our specific sample is well above the average card due to its extremely low VID, in terms of design Sapphire has clearly done their homework and it shows with an excellent cooler that is ridiculously quiet and equally as cool. The factory overclock isn’t anything that shouldn’t be achievable on your own, but if you’re serious about overclocking the cooler alone would be enough to justify the extra $30.

On the other hand the XFX R7950 Black Edition Double Dissipation ends up being a bummer, particularly compared to its 7970 based sibling. For what an open air cooler can do it’s simply too hot and too loud; the numbers we’re seeing would be acceptable for a blower, but not for an open air cooler. The gaming performance is great thanks to its best in class factory overclock, but this isn’t enough to overlook the obvious cooling troubles.

Wrapping things up, so far we’ve looked at single card performance, but what about CrossFire? Later this week we’ll be looking at 7970 and 7950 CrossFire performance, and what the plethora of open air coolers means for 7950 users. So stay tuned.