There's a big hole in between AMD's $450 Radeon HD 7950 and its $160 Radeon HD 7700. Today, the company introduces Radeon HD 7850 and 7870 to fill that gap, and they push a lot more performance than we expected. But are they really ready for prime time?
If you follow the graphics industry, you probably already knew today’s AMD Radeon HD 7800-series launch was happening. The company told us to expect this back in December when it introduced Radeon HD 7970. It reminded us in January when the Radeon HD 7950 debuted, and again in February during the Radeon HD 7770 and 7750 briefings.
Consequently, we’re also not surprised that, as the Radeon HD 7870 and 7850 became public knowledge, supply of the Radeon HD 6970 and 6950 dried up. After all, it’s hard to recommend last generation’s card, still priced competitively, when it’s no longer for sale. The new models replace those older cards with similar suggested prices of $349 for the 7870 and $249 for the 7850.

Another Paper Preview Launch

Unfortunately, although we’re able to talk about the performance of Radeon HD 7870 and 7850, you won’t be able to buy the card until at least March 19^th, AMD tells us. The company says that, because CeBIT and the Game Developers Conference fall so close to the official launch, it’d be difficult to keep information about the cards under wraps. Regardless of the explanation, it sucks to wait another two weeks for street pricing and a better feel for availability.
Cynicism aside, we’re still technology geeks at heart, and we’re desperately curious to know how the 7800-series cards measure up to the Radeon HD 6900s, which earned our affections in the past for their conservative power use, excellent CrossFire scaling, and rich display connectivity compared to Nvidia’s GPUs with two independent pipelines.

This is the third distinct Graphics Core Next-based ASIC from AMD in as many months. Code-named Pitcairn, we finally have a solution in between the flagship Tahiti and the entry-level Cape Verde. The graphics processor is composed of 2.8 billion transistors in a 212 mm² die manufactured on TSMC’s 28 nm lithography node. That’s nearly twice as large as the GPU at the heart of Radeon HD 7770, but around two-thirds of Tahiti, which drives the Radeon HD 7970.
Pitcairn hosts 20 compute units, each with four vector units made up of 16 stream processors and a single texture unit. In an uncut GPU, those numbers multiply out to 1280 total shaders and 80 texture units. AMD shears off four compute units from Pitcairn to create the Radeon HD 7850, resulting in 1024 shaders 64 texture units.
The GPU’s back-end is made up of eight render partitions, each with four full-color ROPs, totaling 32 ROPs. Four 64-bit memory controllers yield an aggregate 256-bit memory bus. AMD doesn’t make any changes to this part of the chip; both the Radeon HD 7870 and 7850 sports identical back-ends.

	Radeon HD 7870	Radeon HD 6970	Radeon HD 7850	Radeon HD 6950
Stream processors	1280	1536	1024	1408
Texture Units	80	96	64	88
Full Color ROPs	32	32	32	32
Graphics (Shdr) Clock	1000 MHz	880 MHz	860 MHz	800 MHz
Texture Fillrate	80 Gtex/s	84.5 Gtex/s	55 Gtex/s	70.4 Gtex/s
Memory Clock	1200 MHz	1375 MHz	1200 MHz	1250 MHz
Memory Bus	256-bit	256-bit	256-bit	256-bit
Memory Bandwidth	153.6 GB/s	160 GB/s	153.6 GB/s	160 GB/s
Graphics RAM	2 GB GDDR5	2 GB GDDR5	2 GB GDDR5	1-2 GB GDDR5
Die Size	212 mm2	389 mm2	212 mm2	389 mm2
Transistors (Billion)	2.8	2.64	2.8	2.64
Process Technology	28 nm	40 nm	28 nm	40 nm
Power Connectors	2 x 6-pin	1 x 8-pin, 1 x 6-pin	1 x 6-pin	2 x 6-pin
Maximum power (TDP)	175 W	250 W	130 W	200 W
Price	$349 MSRP	$350-$410 (EOL)	$249 MSRP	$250 (1 GB, EOL) $270-$300 (2 GB, EOL)

In general, AMD’s Radeon HD 6900 family features more shaders and texture units, while the 7800s operate at higher clock rates. Remember also that both line-ups also employ different architectures optimized for different things. VLIW4 was the product of much evolution, and hence quite mature. Meanwhile, GCN is much newer. We’ve seen it completely tank in DirectX 9-class applications, and then deliver stellar performance in more modern titles.
All of the cards in the above chart feature 256-bit memory interfaces feeding 32 ROPs, but the 6900 series’ higher memory clock rates facilitate a little more bandwidth.
Interestingly, the Radeon HD 6950 only proffers about 10% fewer shaders than the Radeon HD 6970. In contrast, the Radeon HD 7850 comes armed with a roughly 25% deficit compared to Radeon HD 7870. AMD seems to have decided that it needed to spread its mid-range cards further apart from each other. I’ve often wondered if the Radeon HD 6950 cannibalized Radeon HD 6970 sales because they were so similar (surely it didn’t help that 6950s were getting updated via firmware to look even more like 6970s) That could be what AMD is trying to avoid this time around.
Although it’s tempting to look at raw specifications and draw comparisons, you really can’t, on account of the dissimilar architectures. We can, however, point out that all of AMD’s GCN-based cards put a special emphasis on maintaining exceptional performance at lower maximum power. The Radeon HD 7750 doesn’t even need a dedicated PCIe power connector, and yet its performance is often comparable to the Radeon HD 5770/6770. With maximum TDPs at 175 and 130 W, the Radeon HD 7870 and 7850 sport power ceilings that are 75 and 70 W lower than their predecessors, respectively. The Radeon HD 7870 requires two six-pin PCIe power connectors, and the Radeon HD 7850 only needs one. Lower power means less heat. Less heat translates to more conservative cooling. And that leaves the door open for gaming enthusiasts to enjoy quieter systems that go easier on the power bill. These cards might not even necessitate a power supply upgrade, if your machine is reasonably modern.

A Sizable Gap

We’re curious about the space between AMD’s $250 Radeon HD 7850 and its $160 Radeon HD 7770. Like it or not, the Radeon HD 6800s are gone for good. AMD isn’t taking orders on the 6870, 6850, or 6790. The company won’t comment on future products, but we find it hard to believe it’d abandon the $180-$230 market currently inhabited by the GeForce GTX 560, 560 Ti, and Radeon HD 6870. That leaves two likely possibilities for the future: a cheaper 1 GB version of the Radeon HD 7850, or perhaps a crippled version of the 7850 that might populate a lower position in the hierarchy.

Radeon HD 7870 and 7850

Officially, AMD gives its Radeon HD 7870 the same GHz Edition suffix as the Radeon HD 7770, indicating only that the core breaks the completely arbitrary gigahertz barrier.

As far as appearances go, the Radeon HD 7870 is quite similar to the Radeon HD 6870. It’s 9.5” long with two six-pin auxiliary power inputs. The biggest functional difference between old and new is the loss of a DVI display connector. As with the flagship Radeon HD 7970, AMD says it got rid of the second DVI output to improve exhaust airflow. Two mini-DisplayPort connectors and a single HDMI port remain intact. Of course, the card’s style is a little different, adhering to the tapered end we’ve seen from other 7000-series models.

Keep in mind that the Radeon HD 7870 actually replaces the 6970, though, and not the 6870. The higher-end 6970 is an inch longer, and it requires eight- and six-pin power inputs to feed an additional 75 W.
We like the 7870’s more compact dimensions, but enthusiasts will be disappointed to learn about two notable omissions: a BIOS selector switch and dual CrossFire connectors. Unlike the Radeon HD 6900s, which scaled up to four cards in a CrossFire array, the Radeon HD 7870 and 7850 can only be used in pairs. Moreover, the lack of a backup firmware makes flashing the BIOS riskier.
The reference Radeon HD 7850 seems identical to the 7870 with one exception: a single six-pin power input (rather than two), reflecting the 7850’s lower 130 W typical board power rating.
But the reference card isn’t necessarily representative of what you’ll see at retail, and some of the models AMD showed us are actually quite a bit smaller. 

Almost all of the partner boards share the same combination of one DVI-D output, a single HDMI connector, and twin mini-DisplayPorts. Based on the pictures AMD provided, the lone exception is XFX’s model with two DVI-D outputs.

AMD's new FirePro V3900 is the company's low-profile, entry-level workstation graphics card. It's priced to compete against Nvidia’s Quadro 400. Today we're putting it up against Nvidia’s Quadro 400 and five other professional and desktop graphics cards.
Low-cost workstation graphics cards certainly aren't designed to be gaming racehorses. Rather, they're meant to be good enough for a number of mainstream professional tasks. After the recent launch of AMD's FirePro V3900, we were given an opportunity to find out what such basic cards can do, and how well AMD's entry-level card stacks up against its competition.

The FirePro V3900 is aggressively priced compared to what you're probably used to seeing for a workstation-oriented board, currently selling for about $110. According to AMD, the card competes against Nvidia’s Quadro 400, which is offered at a similar price. We decided to benchmark not only those two cards, but also Nvidia's Quadro 600, a Radeon HD 6570, a GeForce GT 430, a GeForce GT 440, and the brand-new Radeon HD 7750. We know those other cards weren't designed to contend in the professional space. However, we're curious as to how they'll do, given similar specifications.

Good Old Friends in New Clothes

AMD's FirePro V3900 features a Turks-based GPU and is quite similar to the Radeon HD 6570. The previously-launched FirePro V4900 sports a more complex Turks-based GPU and is more easily compared to the Radeon HD 6670. According to AMD’s roadmap, Turks will continue to persist until 2015, which is all the more reason to put this little V3900 through its paces.

Of course, we know that the Turks GPU is a descendant of Barts, which powers the Radeon HD 6800-series cards, and which itself is a descendant of Cypress, the GPU driving AMD's older Radeon HD 5800-series cards. Turks, however, only has six SIMD engines, each of which consists of sixteen thread processors. Each thread processor has five stream processors (ALUs). The SIMD blocks have four texture units. And thus, a Turks chip has a grand total of 24 texture units and 480 ALUs. The DRAM is attached via two 64-bit memory interfaces, for an aggregate 128-bit bus. Moreover, the two rendering back-ends of the chip sport four color ROPs, for a total of eight.

	AMD FirePro V3900	AMD Radeon HD 6570	Nvidia Quadro 400	Nvidia Quadro 600
Stream Processors	480 (96 5D)	480 (96 5D)	48 (1D)	96 (1D)
ROPs	8	8	4	8
GPU Clock	650 MHz	650 MHz	450 MHz	640 MHz
Memory Clock	900 MHz	800 MHz	770 MHz	800 MHz
Interface	128-bit	128-bit	64-bit	128-bit
Memory Bandwidth	28.8 GB/s	28.8 GB/s	12.3 GB/s	25.6 GB/s
Video Memory Size	1 GB GDDR3	1 GB GDDR3	512 MB GDDR3	1 GB GDDR3
Shader Model	5.0	5.0	4.0	4.0
TDP	approx. 50 W	approx. 50 W	approx. 32 W	approx. 40 W
Street Price	$110	$50	$95	$145

On paper, the V3900's smaller Turks chip looks more impressive than the GT216 on Nvidia's Quadro 400. It even looks a little better than the GF108 on the Quadro 600. We'll test to see if this translates into the performance of real-world tasks.
Also, it's no secret that the higher performance of workstation cards versus their gamer-oriented brethren in professional applications is mostly a result of optimized drivers. So, a comparison with the FirePro's desktop equivalent is particularly apropos as well.

Enthusiasts want to know about Nvidia's next-generation architecture so badly that they broke into our content management system and took the data to be used for today's launch. Now we can really answer how Kepler fares against AMD's GCN architecture.

Nvidia is fond of drawing parallels. With its prior-generation graphics cards, the company likened each model to a different role on a virtual battlefield. GeForce GTX 480 was the tank—big performance and, to a fault, a big price, big power, and big heat, as well. GeForce GTX 460 came to be referred to as the hunter, incorporating a better balance of speed, efficiency, and cost more apropos to gamers. Finally, GeForce GTS 450 was dubbed the sniper for its focus on enabling playable frame rates at 1680x1050, according to Nvidia.

As silly as that trio of categories seemed, they make it easier for us to put a finger on the pulse of GeForce GTX 680. Though its name (and price) suggests a successor to Nvidia’s current single-GPU flagship, this is decidedly the hunter—a gamer-oriented card that almost completely de-emphasizes the once-emphatic message of improving general-purpose compute performance. But hey, it does that whole gamer thing really well, just like the GeForce GTX 460.

Regardless of the role it was designed to fill, competition is the biggest influencer of positioning. AMD may have higher-end cards on its roadmap that we haven’t seen or heard about yet. However, in the context of AMD’s six Radeon HD 7000-series boards that are already available, Nvidia knows exactly what it’s up against.

Had Radeon HD 7970 been 30 or 40 percent faster than it is, there’s a good chance we wouldn’t be looking at a GeForce GTX 680 today. Maybe it would have been called GTX 660 or 670. But because of where AMD’s flagship sits, Nvidia sees justification in crowning its new hunter as a successor to its old tank—all the while making it pretty clear that another piece of heavy armor is in the works.  

What we have on our hands, then, is a $500 card based on Nvidia’s GK104 graphics processor, designed specifically for gamers (if you’re interested in compute potential, you’ll have to keep waiting). GeForce GTX 680 addresses some of last generation’s most glaring competitive disadvantages, and it adds a handful of interesting features, too. 

At 10” long, the GeForce GTX 680’s PCB is half an inch longer than AMD’s Radeon HD 7800 boards and half an inch shorter than the Radeon HD 7900s.

Looking at the card head-on, we see that it employs a centrifugal fan, which pushes hot air out of the card’s rear bracket. There’s only about half of one slot available for passing exhaust back there. But as we’ll see in our thermal and acoustic tests, the GTX 680 does not have a problem contending with heat.

The rest of the dual-slot bracket plays host to four display outputs: two dual-link DVI connectors, one full-sized HDMI port, and a DisplayPort output. All of them are useable concurrently, addressing one of our sharpest critiques of all prior-gen Fermi-based boards. At long last, we can consider multi-monitor gaming tests to replace 2560x1600 in our single-card reviews (and indeed, multi-monitor benchmarks will follow in a story on which we're already working)! Like AMD, Nvidia claims that this card support HDMI 1.4a, monitors with 4K horizontal resolutions, and multi-stream audio.

Up top, GeForce GTX 680 features twin SLI connectors, enabling two-, three-, and four-way SLI configurations. In comparison, AMD’s Radeon HD 7970 and 7950 similarly support up to four-way arrays.

We also get our first piece of physical evidence that Nvidia’s GK104 processor was designed to be a viable option in more mainstream environments: GeForce GTX 680 employs two six-pin auxiliary power connectors. Those two inputs, plus a PCI Express slot, facilitate up to 225 W of power delivery. Nvidia rates this card for up to 195 W. However, it also says typical power use is closer to 170 W. Keep those numbers in mind—the available headroom between 170 W and the 225 W specification ceiling come into play shortly.

Nvidia claims it put a significant effort into three aspects of its dual-slot cooler’s design, contributing to an impressive acoustic footprint, even under load. We can peel back the GTX 680’s shroud for a closer look…

First, the company cites a trio of horseshoe-shaped heat pipes embedded in the GPU heat sink, which quickly draw heat away from GK104. Thermal energy is transferred from those pipes into a dual-slot aluminum sink. 

Optimizations to the sink itself also rank on Nvidia’s list of improvements. For instance, the fin stack is angled where air exits, creating more space, internally, between the cooler and exhaust grate. Apparently, in prior designs, heat was getting trapped between the bracket and fins, affecting cooling performance. This new approach is said to yield lower temperatures than the older implementation.

Finally, Nvidia says it added acoustic dampening material to the fan motor—a step it also took with the GeForce GTX 580, which contributed to getting that card’s noise level down compared to its maligned predecessor.

Who needs sleep when you have caffeine? We take a second GeForce GTX 680 and run it in SLI against two Radeon HD 7970s. Then we add 5760x1080 benchmark results. Then we overclock our single-GPU flagships for a third comparison. Does our story change?

In the last week, I ate terribly, drank horrible, sugary things, and slept just enough to get myself into the lab to run benchmarks. Yesterday, as I polished off the last bit of text for GeForce GTX 680 2 GB Review: Kepler Sends Tahiti On Vacation, I truly thought I was in for a relaxing early weekend. But then the FedEx guy knocked. Oh, FedEx guy. The rumble of your truck is a bittersweet symphony. 

Sure enough, I had a second GeForce GTX 680 on my hands. And I had already purchased a pair of Radeon HD 7970s. Really, there would be no excuse for not comparing the four boards.
I wanted to test 5760x1080 performance in my original piece, but simply ran out of time after the first three resolutions. Nvidia posted a new driver immediately after the launch, so I could also take the opportunity to verify that its NVEnc/MPEG-2 bug was fixed (it's not) (Update: Nvidia tells us that NVEnc is working, but that its CUDA cores are yielding better performance, masking the impact of its fixed-function hardware). Many folks asked to see a comparison between overclocked Radeon HD 7970 and GeForce GTX 680 cards, and that’d become possible as well.
And oh, there’s the issue of availability. We’ve been quick to slam AMD in the past for paper-launching its products, sending samples out to press with a note to expect hardware weeks later. Nvidia made it a point that GeForce GTX 680 would be available on launch day, and an early leak from Newegg provided confirmation enough. Unfortunately, the boards disappeared within hours of going on sale, and there are no longer any GeForce GTX 680s to buy. Nvidia says that another wave will hit our shores in early April (Update: Nvidia also claims that a quantity of boards are arriving to AIBs every day, so keep checking online. The "early April" guesstimate referred to what the company calls "going virtual," where its partners will start shipping their own customized designs and volume ramps up substantially. We'll be keeping an eye on the availability story in the meantime; as of 3/25, there are none available). But until then, everyone who didn’t set their alarm for the 6:00 AM embargo is just as stuck as the folks who had to wait weeks for Radeon HD 7970s, and that’s no fun.
Lastly, I made a mistake in yesterday’s story. In calculating the performance per watt index of the GeForce GTX 680, a single cell in Excel was reversed, and the 680’s performance ended up getting divided by the 7970’s power. The result was an overstatement of Kepler’s efficiency compared to Fermi, which has already been fixed in the launch story. GK104’s performance per watt outcome compared to GeForce GTX 580 is still significantly better than Tahiti’s—but not to the degree previously reported.
At any rate, today’s update should give you an even more complete picture of the GeForce GTX 680’s behavior and how it compares against AMD’s fastest single-GPU card. Let’s have a look at the card from EVGA enabling our SLI-based benchmarks, and then get on with the numbers!

Did you spend some time this week wondering what Nvidia was planning for today's announcement? Everyone who guessed GeForce GTX 690, you're right. Powered by two GK104 graphics processors, we have the scoop on Nvidia's soon-to-be flagship.
Update (5/3/2012): Nvidia's GeForce GTX 690 just launched. You can read all about the card and its performance in GeForce GTX 690 Review: Testing Nvidia's Sexiest Graphics Card.
I had my head so thoroughly buried in the Ivy Bridge launch that I completely missed Nvidia’s “Something is Coming” tease last week.
Well, now we know that “Something” is actually the company’s GeForce GTX 690. It’s not here yet; you’ll have to wait until next week for a comprehensive look at the card and its performance. But we do have all of the speeds and feeds thanks to tonight’s webcast from the Nvidia's GeForce LAN 2012 in Shanghai, China.

Inside of GeForce GTX 690

GeForce GTX 690 is a dual GK104 part. Its GPUs aren’t neutered in any way, so you end up with 3072 cumulative CUDA cores (1536 * 2), 256 combined texture units (128 * 2), and 64 full-color ROPs (32 * 2).

Gone is the NF200 bridge that previously linked GF110 GPUs on Nvidia’s GeForce GTX 590. That component was limited to PCI Express 2.0, and these new graphics processors beg for a third-gen-capable connection. Rather than engineering its own solution, though, Nvidia looked to PLX for one of its PEX 874x switches. The PCIe 3.0-capable switches accommodate 48 lanes (that’s 16 from each GPU and 16 to the bus interface) and operate at low latencies (down to 126 ns). What of NF200’s features incorporated to augment the performance of multi-GPU applications, Broadcast and PW Short? According to Nvidia, those capabilities were wrapped up into the PCIe 3.0 specification, so GeForce GTX 690 gets them through the PLX component.

Each graphics processor has its own aggregate 256-bit memory bus, on which Nvidia drops 2 GB of GDDR5 memory. The company says it’s using the same 6 GT/s memory found on its GeForce GTX 680, able to push more than 192 GB/s of bandwidth per graphics processor. Core clocks on the GK104s drop just a little, though, from a 1006 MHz base down to 915 MHz. Setting a lower guaranteed ceiling is a concession needed to duck in under a 300 W TDP. However, in cases where that power limit isn’t being hit, Nvidia rates GPU Boost up to 1019 MHz—just slightly lower than a GeForce GTX 680’s 1058 MHz spec.

Power is actually an important part of the GeForce GTX 690’s story. Nvidia’s GeForce GTX 590 bore a maximum board power of 365 W. One 75 W slot and two 150 W eight-pin plugs pushed the 590 mighty close to its ceiling. But because the Kepler architecture takes such a profound step forward in efficiency, Nvidia is able to tag its GTX 690 with a 300 W TDP. And yet it still includes two eight-pin inputs able to deliver up to 375 W, including the slot.
Consequently, representatives are claiming that there’s significant headroom to push the GeForce GTX 690 beyond the performance of twin GeForce GTX 680s. At its stock power and clock settings, though, we’re expecting it to be a little slower than two single-GPU boards on their own. The good news is that one GeForce GTX 690 should use markedly less power and operate more quietly than a pair of GTX 680s in SLI. Nvidia says it is deliberately binning low-leakage parts to minimize heat, and its own measurements put a GeForce GTX 690 at 47 dB(A) under load compared to 51 dB(A) for the 680s.

GeForce GTX 680 cards are nowhere to be found, and the Radeon HD 7970 recently dropped to a much more attractive price. We thought it was time to round up a handful of Tahiti-based cards to see how board partners are improving upon AMD's original recipe.

After a handful of price drops on AMD's highest-end graphics cards, the Radeon HD 7970 is now more attractive than it was immediately after the introduction of Nvidia's $500 GeForce GTX 680. Particularly because Nvidia's flagship card is still very difficult to find more than a month after its launch. 

With that in mind, added to the fact that AMD's board partners have had ample time to make their own improvements to the company's top-end model, we have a great opportunity to revisit the Radeon HD 7970. Five different vendors sent models for us to consider, resulting in an impressive round-up of Tahiti GPU-based graphics cards.

Note that HIS has two models represented on the chart. We were sent the company's "Turbo X" model for testing, and we believe it has the highest factory core overclock of any Radeon HD 7970 in the world at 1120 MHz. Unfortunately, though, this product is not expected to be available in the U.S. (only in Europe and Asia). Instead, we'll be getting the "Turbo" model, with no X and a 1050 MHz core clock. Thus, we benchmarked the card a second time with HIS' more conservative specification to demonstrate what North America can expect.

In lieu of HIS' Turbo X, MSI's 7970 Lightning sports the highest-clocked core available for sale to our U.S. audience, running at 1070 MHz. Sapphire's HD 7970 OC enjoys the distinction of shipping with the fastest memory compared to all other Radeon HD 7970s. However, its core operates a more modest 1000 MHz.

Operating frequencies aren't the only variables affecting our decision of what Radeon HD 7970 to recommend, though. Software bundles, acoustics, thermals, price tags, support, and overclocking headroom all come into play. Let's break down each model and explore what these five vendors offer.

We already know that Nvidia's GeForce GTX 690 sports two GK104s and is priced at $1000. But hardware like this is fun to read about. Oh, you actually want to buy one? Expect performance just shy of two GTX 680s in SLI, and good luck tracking one down!
Bugatti’s Veyron Super Sport. Aston Martin’s One-77. Lamborghini’s Reventón. They’re all million-dollar-plus automobiles that most of us love to read about, probably won’t see on the road, and almost certainly will never own.
And yet, they’re still sexy.
When it comes to drawing parallels between high-end hardware and really nice cars, I’m just as guilty as many others in the technology press. Really, though, the comparisons are hardly realistic. For the price of four Veyron tires, you could buy 35 Core i7-3960X processors.  A thousand bucks for a CPU sounds ludicrous, but that’s certainly more accessible than many of life’s other luxuries.


So, while it’s tempting to review Nvidia’s new GeForce GTX 690 as if it were the gaming world’s Bugatti—to describe its features as if it were an untouchable kilogram of PCI Express 3.0-capable unobtanium—the fact of the matter is that this thing could become a viable option for enthusiasts with a thousand bucks in their hands.
Sure, you’ll also need an expensive platform and at least one 30” display in order to enjoy its capabilities. But we can’t just take the GTX 690 for a lap around the track, call it an amazing-looking piece of equipment, and assume you’ll never have to choose between this and a couple of GeForce GTX 680s or Radeon HD 7970s.
No, GeForce GTX 690 has to pass muster—same as any other card that finds its way into our little Nürburgring for GPUs.
We received enough detail about GeForce GTX 690 from Nvidia’s announcement at the GeForce LAN 2012 in Shanghai to write an early preview (Nvidia GeForce GTX 690 4 GB: Dual GK104, Announced). But I’ll try to expound just a little, borrowing from the text used in my news piece. And then of course we’ll move on to the benchmarks at 1920x1080, 2560x1600, and 5760x1080.

GeForce GTX 690 4 GB: Under The Hood

From that preview:
GeForce GTX 690 is a dual-GK104 part. Its GPUs aren’t neutered in any way, so you end up with 3072 cumulative CUDA cores (1536 * 2), 256 combined texture units (128 * 2), and 64 full-color ROPs (32 * 2).
Gone is the NF200 bridge that previously linked GF110 GPUs on Nvidia’s GeForce GTX 590. That component was limited to PCI Express 2.0, and these new graphics processors beg for a third-gen-capable connection.

We now know that the GeForce GTX 690’s GPUs are linked by a PEX 8747, one of PLX’s third-gen, 48-lane, five-port switches manufactured at 40 nm. The switch communicates with each GK104 over 16 lanes of PCI Express 3.0, and then creates a 16-lane interface to the host. Developed expressly for graphics applications, the 8 W switch is rated for 126 ns latency, so you shouldn’t have to worry about its insertion between the GPUs negatively impacting performance. And again, a number of the capabilities uniquely integrated into NF200, like multi-cast, are supported by PEX 8747, too. Continuing…
Each graphics processor has its own aggregate 256-bit memory bus, on which Nvidia drops 2 GB of GDDR5 memory. The company says it’s using the same 6 GT/s memory found on its GeForce GTX 680, able to push more than 192 GB/s of bandwidth per graphics processor. Core clocks on the GK104s drop just a little, though, from a 1006 MHz base down to 915 MHz. Setting a lower guaranteed ceiling is a concession needed to duck in under a 300 W TDP. However, in cases where that power limit isn’t being hit, Nvidia rates GPU Boost up to 1019 MHz—just slightly lower than a GeForce GTX 680’s 1058 MHz spec.


Power is actually an important part of the GeForce GTX 690’s story. Nvidia’s GeForce GTX 590 bore a maximum board power of 365 W. One 75 W slot and two 150 W eight-pin plugs pushed the 590 mighty close to its ceiling. But because the Kepler architecture takes such a profound step forward in efficiency, Nvidia is able to tag its GTX 690 with a 300 W TDP. And yet it still includes two eight-pin inputs able to deliver up to 375 W, including the slot.
Although the GeForce GTX 690’s board-level features are certainly impressive, this story is arguably just as much about the effort Nvidia’s engineers put into building a card that’d operate quietly, more effectively dissipate heat, and look good in the process. The result is a sturdy, rigid piece of hardware, which demonstrates more showmanship than any reference card preceding it. Back to the preview:

The exterior frame is built of chromium-plated aluminum, rather than the plastic materials covering most other cards (including the GeForce GTX 680). The fan housing itself is a magnesium alloy, which purportedly aids heat dissipation and logically improves vibration dampening compared to plastic shrouds.
Dual vapor chambers cover each of the GPUs, similar to what we’ve seen from both Nvidia and AMD in the past. This time, however, a polycarbonate window over each fin stack allows curious enthusiasts to peer “under the hood.” An LED up top can actually be controlled through a new API Nvidia is making available to partners. So, it might respond to load, getting lighter and darker, as an example.
Currently, I should note, the LED simply lights up when the card is powered on. And here’s where we get to the part sure to turn some folks off…

As with dual-GPU boards from the past, Nvidia is using a center-mounted axial fan, which it claims is optimized for moving air without generating a lot of noise. The trouble with axial fans in this configuration is that they exhaust the heat from one GPU out the rear I/O panel, while the second chip’s thermal energy is jettisoned back into your chassis. Both AMD and Nvidia went the axial route last generation, so we have to surmise that it’s logistically the only approach that makes sense. At least the TDP on GeForce GTX 690 is lower than 590, indicating less maximum heat to dissipate.
Talk to some of the boutique system builders out there and they’ll tell you that shifting from centrifugal to axial-flow fans is akin to the manufacturer dealing with its own product’s thermal output, and then saying, “Here, now you figure out what to do with this.” If we’re to imagine that the most relevant application for GeForce GTX 690 is quad-SLI, you basically have the equivalent of two GeForce GTX 680s dumping hot air into your chassis instead of exhausting out. Finally…
The 690’s rear I/O panel plays host to three dual-link DVI outputs and a mini-DisplayPort connector, accommodating four screens total. A single SLI connector up top links the GeForce GTX 690 to one other card, enabling quad-SLI arrays.

Connecting three displays to two GeForce GTX 680s can be a real pain. Although each board offers four display outputs, there are recommended combinations, depending on whether your screens employ DVI, HDMI, or DisplayPort. With three DVI connectors on GeForce GTX 690, configuration becomes significantly easier.

Last week we reported that an engineering sample card with 768 shaders accidentally found its way into our lab instead of the HD 7850 we were expecting. This GPU may be meant for engineers, but it piqued our interest, since it happens to fill a large gap.

In last week’s news story AMD Pitcairn With 768 Shaders: What is This Mystery Chip?, we reported that, while testing a pre-production sample Radeon HD 7850 shipped to us in a rare single-slot form factor, we discovered a mysterious GPU that didn't match the Pitcairn we were expecting. Instead of coming equipped with the usual 1024 shaders, our engineering sample sported only 768 active shader cores.
We made a number of phone calls and exchanged a number of emails with AMD in an attempt to clarify our sample's origin and intention. What we found was that the 768-shader part did not, in fact, represent a piece of silicon destined to drive an upcoming super-secret derivative product. Rather, the GPU was only intended for board partners. It was never meant to go on sale, and it was certainly never supposed to land in our lab.
But if the part wasn't intended to sell or test, what was its purpose? There is a good explanation for its existence. For AMD’s add-in board partners, it is important to get their hands on GPUs as early as possible. Early access allows them to build boards in preparation for a launch. On the other hand, AMD is wary of leaks that would give away the performance profile of its GPUs in a very competitive graphics market. That’s where deliberately-hobbled engineering samples like this one come into play, allowing board partners to start developing their custom implementations with a chip that thermally and electrically simulates the final product. Meanwhile, AMD doesn’t need to worry about the performance attributes of its final effort getting outed by a company with unfettered access to its technology.
There remain a few points for us to clarify, though. When the single-slot Radeon HD 7850 that we tested appears for sale, we're told that it will feature a complete 1024-shader Pitcairn GPU. Obviously, the sample we're testing today isn't indicative of that final production product. To that end, the benchmark results we'll be presenting on the following pages are most easily thought of as experimental. We have a card here that nobody else has. It's properly recognized by AMD's Catalyst driver as a Radeon HD 7800-series board. And it effectively operates as a single-slot piece of hardware. AMD stresses that, at least for now, there are no plans to produce a Pitcairn GPU with a 768-shader configuration. As far as we are concerned, that’s actually too bad, since the benchmark numbers make such a chip a pretty interesting product. Also, judging by the overwhelmingly positive feedback our initial report generated, both on our own sites and around the Web, there is certainly a market for it. As such, this piece isn't a review so much as it's an exploration of a hypothetical component that doesn't exist (though we'd certainly welcome it if it ever did materialize).

But what makes this chip so interesting if it was never meant for public consumption? We don’t want to give too much away, but let’s take Crysis 2 as an example and look at the performance we get when we run the card through our benchmark using some settings that reflect what a board of this price range can be expected to face in real life.
Let's use a 1920x1080 resolution, 16x AF, 2x MSAA, Ultra details, and DirectX 11 mode for our example. A Radeon HD 7850, which can be found on Newegg starting around $250 (ignoring some rebate offers) takes those settings in stride, achieving a comfortable 41 FPS, on average. The next card down in AMD’s current portfolio, Radeon HD 7770, which sells for about $100 less, is not playable under the same load, averaging around 28 FPS. That leaves a gap of 13 FPS and $100 separating playability and disappointment. Doing a bit of math, the halfway-point between the fastest 7700- and entry-level 7800-series SKU gets us a price tag around $200 and a frame rate around 35 FPS. Lo and behold, the engineering sample in our lab comes in at a respectable 36.5 FPS!
AMD insists that there is no product planned in between the Radeon HD 7770 and 7850. The company is adamant that the 768-shader Pitcairn part was only intended for high-level validation and board characterization. And, at the end of the day, this might go down in history as a cool piece of technology that never makes it out of one site's testing lab.
However, we're certainly allowed to have the opinion that AMD has a gap in its line-up that we think it should be filling rather than ceding to the competition without a fight. And that’s why this GPU is so interesting (aside from the benchmark results that enthusiasts like us find fascinating in their own right). The comments on our news report showed that a card based on a chip like this would definitely appeal to gamers on a limited budget that currently can’t find the right card. While it’s not really our job to find and point out business opportunities to the companies whose products we review, we believe this card would find an audience, and we hope AMD takes note and listens to your comments!

Not like it was ever really widely available anyway, right? The GeForce GTX 670 offers most of GK104's on-chip resources, doesn't give up much performance, and costs $100 less. Now, let's see if Nvidia can make enough of them to satisfy demand.

How’s this for perspective? Last generation, Nvidia’s dual-GPU flagship sold for somewhere around $700. Before that, the GeForce GTX 295 was a $500 board. If you want the pinnacle of graphics performance today, the GeForce GTX 690 will cost you at least $1000—more if your vendor of choice is marking it up, as many are right now.

It’s hard to have a straight-faced discussion about value with the prices of high-end graphics cards shooting off into space. But when Nvidia told me that it planned to sell its GeForce GTX 670 for $400, I was ready to get serious about the V-word.

Of course, an attractive price is only one variable in the equation that determines whether you want to buy something or not. Performance is another, as is availability (a more problematic factor for Nvidia as of late).

The natural question to ask becomes: how fast is GeForce GTX 670?

Benchmarks tell that tale. But specifications give us a good measure of what to expect. Nvidia’s newest card centers on the same GK104 GPU as the company's GeForce GTX 680. But whereas the 680’s graphics processor employed all eight of the chip's SMX clusters, the 670 utilizes seven. The eighth is disabled. Hardware-wise, everything else is the same.

That means GeForce GTX 670 weighs in with 1344 total CUDA cores (192 *7) and 112 texture units (16 * 7). It also ends up giving up a single PolyMorph engine, four warp schedulers, and eight dispatch units, though the overall effect of shutting down an SMX is intended to scale with the GPU’s other resources. We did run our usual tessellation scaling numbers, and found the GTX 670 just slightly worse-off than the GTX 680 before it.

GK104’s back-end remains intact, consisting of four ROP clusters that output eight 32-bit integer pixels per clock each, totaling 32. Similarly, four 64-bit memory controllers create a 256-bit aggregate interface.

Nvidia populates that bus with the same 2 GB of GDDR5 memory found on its GeForce GTX 680, even setting the same 1502 MHz frequency, yielding a 6008 MT/s data rate. To further differentiate the GeForce GTX 670 from the more expensive 680, Nvidia does drop the lower-end card’s core to a 915 MHz base, with the typical GPU Boost setting landing at 980 MHz.

As we’ll see, though, those moves don’t have as much cumulative effect as we might have suspected.

Nvidia’s GeForce GTX 680 is 10” long. Both its PCB and its cooling shroud are that long. A nice, sturdy aluminum frame encircles the whole card, adding rigidity for a beefy vapor chamber and acoustically-optimized centrifugal fan.

The reference GeForce GTX 670, on the other hand, is 9.5” long. But its PCB only accounts for 6.75” of that. Nvidia claims that the 670’s scaled-back power requirements allowed it to move voltage regulation circuitry to the other (left) side of the GPU, which itself is rotated to purportedly improve signal integrity.

Nvidia uses the same fan found on the GeForce GTX 680, which we like because it exhausts all of the card’s heated air out the rear I/O panel. But it employs a cost-reduced heat sink. You’ll see in the power and noise benchmarks that the result is slightly louder, slightly warmer operation under load. However, the GTX 670 is only marginally less attractive in that regard. 

Deactivating a single SMX and turning down the GeForce GTX 670’s core clock results in a typical board power of 141 W, according to Nvidia. That’s around 30 W less than the GeForce GTX 680, which bears a 170 W typical and 195 W maximum power rating. Since a 16-lane PCI Express slot only delivers 75 W of power, you still need two six-pin auxiliary connectors to drive the GTX 670. In the event that you forget to attach those leads, Nvidia added a pre-boot warning message to its Kepler-based cards instructing the end-user to plug them in.

The 670 offers the same four display outputs as Nvidia’s GeForce GTX 680: two dual-link DVI connectors (one DVI-I and one DVI-D), one full-sized HDMI output, and one full-sized DisplayPort connector. All four can be active simultaneously, partly addressing AMD’s Eyefinity technology, which we’ve seen enable up to six screens on one card.

Afox sent us its rendition of the new GeForce GT 640, based on Nvidia's GK107 GPU. This one slots in under the existing GeForce GTS 450 to do battle with AMD's Radeon HD 6670. Does the Kepler architecture deliver, or is a $100 price target too high?

There was once a time when Nvidia and AMD would fight viciously to deliver the very best performance at every price point. Regardless of which company had the fastest flagship, they both managed to excel in specific segments. It was truly a great time to be a PC gamer. This golden age of tit-for-tat lasted a very long time, starting with the very first Radeon, which challenged Nvidia’s relatively new (at the time) GeForce 256. 

Somewhere between Nvidia’s GeForce 200 and 400 generation, though, something changed. Even as the company maintained a strong portfolio above $150, Radeon cards began to dominate below the $100 price point. The potent GeForce 9600 GT was phased out in favor of the slightly slower GeForce GT 240, which in turn was replaced with a notably inferior GeForce GT 430. Radeon HD 5570 and 5670 cards held the upper hand in the entry-level market for a very long time, and we’ve seen so little competitive pressure that AMD now confidently sells the DDR3-equipped Radeon HD 6670 for the same price that the superior GDDR5-equipped Radeon HD 5670 used to cost. Performance per dollar is actually worse now than it was back then.

The best value in Nvidia’s low-end line-up is the GDDR5-equipped GT 440, essentially an overclocked GT 430 with a lot more memory bandwidth. Unfortunately, that’s a relatively low-volume specialty model a few vendors offer as a well-kept secret. And it’s typically a bit slower than the cheaper DDR3-equipped Radeon HD 6670 (On a side note, Nvidia has recently re-branded the GeForce GT 440 as the GeForce GT 630, but despite the name-change this card is still based on the older Fermi architecture).

We’ve been scratching our heads over this state of affairs for a very long time. Low-end value-oriented GPUs aren’t as sexy as flagships, but they represent a lot of volume and, done right, profit. We don’t have a speculative explanation to help understand why Nvidia’s sub-$100 line-up is lacking. But today the company is quietly slipping out a glimmer of hope in its GeForce GT 640. This is the first Kepler-based card we’ve seen with a GPU other than GK104. Instead, it employs GK107.

Measuring 118 mm², GK107 is about 40% as large as the 294 mm² GK104, and its 1.3 billion transistors represent close to 37% of the higher-end gaming GPU.

In essence, the GeForce GT 640’s GK107 features one-quarter of a fully-functional GK104’s shader resources. Those 384 CUDA cores are divided up into a pair of SMXes, each with 16 texture units (totaling 32). Nvidia sets the chip to run at 900 MHz. Lacking the GPU Boost functionality of the GK104-based boards, GeForce GT 640 cannot dynamically operate at higher clock rates when thermal headroom exists.

This GPU’s back-end consists of two ROP clusters, each able to output eight 32-bit pixels per clock, adding up to 16. Two corresponding 64-bit memory interfaces aggregate to 128-bits. And with 2 GB of DDR3 operating at 891 MHz, the card gets access to a meager 28.5 GB/s.

The new architecture also enables a handful of features previously not available from Nvidia, including PCI Express 3.0 compatibility and Surround support across three monitors (the card actually supports up to four monitors in hardware, but we've been told that partners will typically include three outputs). In addition, it’s Nvidia’s first low-cost SKU based on a 28 nm GPU. To learn more about Kepler, read Chris Angelini’s GeForce GTX 680 launch article.

Nvidia rates the card with a 65 W thermal design power, which ducks in under the 75 W deliverable by a 16-lane PCI Express slot. As such, there are no auxiliary power connectors on the GeForce GT 640. Nevertheless, the reference design still employs a dual-slot form factor thanks to a thick GPU cooler.

More so than past derivative graphics processors from Nvidia, GeForce GT 640 looks like it could shake things up in the budget segment. The company is claiming its DDR3-equipped card is faster than the Radeon HD 6670 with GDDR5 and priced under $100. It doesn’t look like there’s currently a model with GDDR5 launching, but that could be an even more interesting proposition.

Unlike the last two Kepler-based launches, Nvidia is sliding this one out the door quietly. From what we’ve heard, supply of GK107 isn’t particularly robust and, as we’ve said, Nvidia doesn’t seem to make much noise about its entry-level offerings anyway. Our test sample, provided by Afox, is a good example of what we can expect from a typical GeForce GT 640 DDR3:

At first glance, you might mistake the half-height GeForce GT 640 DDR3 for a GeForce GT 430. The small PCB is about 6” by 2 ¾”, and should easily fit in all but the tightest enclosures. 

The Afox GeForce GT 640 is what is being referred to as a reference-clocked model running at 902 MHz. Here's the thing, though. The GeForce GT 640 isn't exactly new. Three different versions of the thing are listed on Nvidia's OEM product page, none of which matches what we have on the bench. Rather, today's launch involves what will likely become a retail product sold into the channel, while the OEM-specific versions go to larger integrators. At the low-end, Nvidia's OEM GeForce GT 640 128-bit DDR3 is most similar in name, though it employs a 797 MHz core, making it substantially slower than our 902 MHz sample.

As with the OEM version, retail GT 640s will come with either 1 or 2 GB of memory. We already know of one vendor with 2 GB cards; ours happens to be a 1 GB variant, though. Fortunately, the memory on all of the DDR3-based versions is at least clocked at a consistent 891 MHz.

Afox’s cooler is a simple radial design, reminiscent of the ones you might see on GeForce GT 430 and 440 cards.

Here’s something new: three display connectors on a low-end GeForce. The DVI, VGA, and HDMI outputs are capable of handling a trio of independent displays at the same time. And, unlike AMD's cards, you don't need a DisplayPort monitor or adapter to get the array up and running. Although we'd prefer a trio of digital outputs, it's understandable to find VGA on an entry-level add-in board. We tested triple-monitor Surround and had no trouble playing DiRT 3 at 5760x1080 using the lowest detail setting.

Afox's bundle is Spartan, including a user manual, driver CD, and half-height I/O bracket. The differentiator is a couple of temporary tattoos, which we've never seen before in a graphics card package.

Afox let us know that it expects the GT 640 to sell for $105, while the word from Nvidia is that it sell for $99 or lower. In light of aggressive competition, we think those numbers are a bit high. But, keep in mind that if supply is ample and costs are low enough, entry-level discrete graphics cards usually find their niche in the market within a couple of months.

Right now, word on the street is that supply of GK107 is limited. And although it'd seem logical that 28 nm-based GPUs would be cheaper to manufacture than 40 nm ASICs, there's still reportedly a premium on those wafers coming from Nvidia's foundry partner TSMC.