by Alexey Stepin , Yaroslav Lyssenko, Anton Shilov
08/03/2008 | 07:01 PM
We have already discussed in detail ATI’s strategy: they decided not to focus on high-performance monolithic GPUs and instead bet on the relatively inexpensive mainstream graphics solutions. And it has totally paid back (for details see our article called ATI Radeon HD 4800 Graphics Architecture: Long Anticipated Revenge?). Today we should pay special attention to a completely opposite approach to winning the market suggested by Nvidia and find out if they made the right bet.
Until recently Nvidia has been dominating the high-performance single graphics solutions market and the only competitor it had there was probably ATI Radeon HD 3870 X2 Nevertheless, graphics architecture needed a refresh and the market longed for new graphics processors, especially since Nvidia’s major competitor didn’t keep hands in pockets and put the final touches to its “weapon of retribution”. Moreover, some games, such as Crysis or Call of Juarez, needed more than the industry could offer at that time. As usual, only next generation solutions could actually provide a new level of gaming performance. Besides, a mighty long time - a year and a half - has already passed since the arrival of Nvidia GeForce 8800 GTX, so the industry simply cried for more.
As we have already said, ATI gave up the idea of high-performance monolithic GPUs targeted at the high-end market segment. However, the reason for that decision had nothing to do with the resources availability.
The previous GPU generation already turned out pretty complex: the number of die transistors was long past 500 million, the die measured 484sq.mm and over 100W power consumption turned into a common standard. As a result, if they continued increasing the GPU potential the usual way, they would have ended up with a much more complex chip of larger size, which would evidently heat more and consume much more power. Of course, developing and finalizing “mega-chips” requires significant technical skills, financial investments and takes a lot of time, therefore, at some point they could eventually become unbearable even for the recognized leaders of the discrete graphics market. Moreover, the performance could actually increase not as dramatically, as it used to in the past years.
As a result, ATI decided to focus on designing the most advanced mainstream graphics processor. By combining a couple of those within a single product they successfully obtained a high-end solution. Of course, this strategy has a number of drawbacks. The main one is certainly the need to optimize drivers for multi-GPU configurations, as well as significant performance and price difference between single- and dual-chip graphics solutions.
Nvidia, however, remained loyal to its principles and wasn’t afraid to pursue “GPU performance at any price” slogan, despite their recent success in designing a dual-chip Nvidia GeForce 9800 GX2 solution. Although ATI’s approach certainly has its advantages, the traditional way also offers a number of positive things, the major being the absence of problems typical of contemporary multi-GPU configurations. Moreover, the use of monolithic chips gives a little more room for flexibility when lining up the entire family of solutions. By simply disabling some functional GPU units they can get less powerful and less expensive graphics adapter modifications (however, it also means that some solutions targeted for specific applications field will be more expensive to manufacture). And by polishing off the production technology and increasing the core frequency potential they will get more powerful and expensive solutions.
New Nvidia chip codenamed G200 was supposed to top the traditional approach to GPU design: the most complex, the most expensive, the most powerful and of course the fastest of all. Everything was at stake here: the company didn’t stop at anything trying to prove to the world who the real king of consumer 3D graphics was.
They claimed their absolute leadership on June 17 2008 and in our today’s article we will try to find out if that was a rightful claim. As usual, we are going to start with comparative analysis of the new GPU features and functionality of solutions based on it.
The new naming system Nvidia selected for their graphics cards on G200 chip, where the model number is placed after the suffix, seems pretty reasonable and gives a lot of flexibility with adding new models in the future. Unlike the old system that has reached its logical limit, the new one will allow to continue with new Nvidia GeForce generations, such as GTX 300, GTX 400, etc. The new Nvidia GeForce GTX 200 was at launch represented by two models: the flagship GeForce GTX 280 and less powerful GeForce GTX 260 with the recommended prices of $649 and $399 respectively. However, constantly growing pressure from ATI forced Nvidia to lower the first price to $499 so that this product could compete against ATI Radeon HD 4870 X2, and the second price point – to $299 in order to compete against ATI Radeon HD 4870.
As a result, the portrait of the new Nvidia GeForce GTX 200 family looks as follows:
Click to enlarge
Click to enlarge
The first thing that catches your eye right away is the size and complexity of the new core. Here Nvidia is an indisputable leader. They are in fact the first graphics chip developer to exceed 1 billion transistors. Together with not the most advanced 65nm production process the company strategy of designing highest performing monolithic GPUs paid back: G200 frequency potential turned out considerably lower than that of G92, while the chip size exceeded all possible limits.
A standard 300mm silicon wafer can accommodate less than 100 G200 chips, which significantly increases the production costs and lowers the yields. As a result, G200 based graphics cards cannot possibly be cheap. According to Rick Bergman, the head of AMD graphics division, G200 has every chance to become the last monolithic graphics core for the premium market segment, because further performance increase within Nvidia’s strategic concept will require enormous investments and will inevitably result in even higher end-product pricing.
While ATI used GDDR5 chips that can work at extremely high frequencies to achieve high memory bandwidth and at the same time keep the PCB layout simple, Nvidia chose a familiar solution and increased the memory bus width to 512bit. As a result, they managed to hit record-breaking results with their flagship model, but it inevitably increased the complexity of the PCB layout and raised the production costs. This is another reason why Nvidia GeForce G200 graphics cards can not be cheap and any price reduction will mean financial losses for the manufacturer.
G200 has 240 ALU, 80 TMU and 32 RBE, although Nvidia’s texture processors architecture makes the performance of these 80 TMU about half as fast. It is interesting that the dual-chip Nvidia GeForce 9800 GX2 seemed to be much better “armed” than GeForce GTX 280 in terms of both: the computational potential as well as texture processors performance. It may also turn out faster in games, which we are going to check later.
Trying to increase the texture performance of their new G200, Nvidia acted pretty wastefully. However, since Nvidia works very closely with game developers it could be their indirect attempt to influence them, since the tendency towards multi-platform gaming titles has been growing pretty strong for quite some time now. As a result, they use fewer graphics effects requiring fast TMU, since the graphics processors of contemporary gaming consoles are much weaker in this aspect than the PC GPUs.
As for the features of specific Nvidia GeForce GTX 200 models from the new family, they obviously differ by their clock speeds and the number of active GPU units. As a result, Nvidia may be making their GeForce GTX 260 solutions using not only some of the G200 chips that failed to meet the maximum frequency potential requirements, but also those that have several defective functional units. From the functional standpoint, this solution looks just as good as ATI Radeon HD 4870, at least first glance. Besides, they are currently priced very closely – somewhere around $300.
So, despite pretty surprising for 2008 combination of computational and texturing capacities, Nvidia G200 still has to boast relatively good potential, but only the benchmark results will be able to reveal that. Today we are not going to dwell on gaming performance, but will discuss in detail the new Nvidia GPU architecture and see how new it actually is.
The domain of previous Nvidia graphics processors with unified architecture consisted of 8 clusters, each including 16 unified execution units supporting all types of shaders. G200 not only has more ALU (the number of ALU has been increased from 128 to 240), but also has 10 clusters instead of 8 and uses a new way of grouping ALU into clusters.
Before that, each of the computational clusters included two shader processors (streaming processors in Nvidia’s terminology) each featuring 8 ALU. Now each cluster consists of three processors like that. Besides the L1 cache and instructions dispatcher, each still contains 8 ALU and a little local memory for exchanging data.
So, the computational capacity of the new G200 has increased significantly compared with G80 and G92. It is still far behind ATI RV770 with 800 ALU grouped in 160 shader processors, however, higher working frequency of Nvidia shader domain partially makes up for this difference. Since each G200 shader processor contains 24 ALU and 8 texturing units, the new core features 3:1 ALU:
Besides more execution units and new approach to grouping them into clusters, Nvidia also introduced a number of optimizations. Namely, they increased the number of threads a shader processor can process simultaneously from 768 to 1024; increased the number of general purpose registers and the internal buffers capacity; introduced support for dual-point precision floating-point calculations (FP64), which required a corresponding unit to be integrated into each of the 30 shader processors. The performance rate here is pretty modest though: Nvidia G200 in FP64 mode provides only about 90 Gflops, while ATI RV770 can theoretically hit 240 Gflops at dual-point precision calculations.
At the same time, new Nvidia chip has no DirectX 10.1 support, which seems to be more of a political move, rather than an implementation issue.
Nvidia G00 texturing processors haven’t changed that much since G84/G92, although Nvidia claims that they improved the control logic, which sped up texture sampling. There are 8 texturing processors with two addressing and two texture filtering units assigned to each of the 10 computational clusters of the chip.
Theoretically, this architecture of texture processors allows processing up to 80 pixels per clock, but with enabled anisotropic filtering or floating-point textures, this number drops down to 40. In other words, 80 texturing processors in reality often turn into 40. This Nvidia’s passion for growing texturing performance at any rate cost them a significant number of transistors.
Although computational part occupies a bigger area, texture processors till take up a lot of space. As a result, the G200 die got almost 600sq.mm big, and it became so complex that the developers had to move display controllers into a separate chip once again, the same way they did with G80 back in the days.
The basic functionality of G200 raster processors remained the same, however, the number of render back ends (RBE) increased to 8. In other words, the new chip has 32 ROP and can release up to 32 pixels per clock.
This is another way of using extensive approach to increasing performance. However, intensive approach is present too: RBE can shift pixels in RGBA8 format at full speed, compared with half the speed by G80/G92.
On the one hand, 8 RBE units translate into 8 64-bit memory controllers and 512-bit external access interface with unmatched bandwidth. But on the other hand, 32 RBE demanded their piece of transistor pie, which turned out not that small at all, as you can see on the diagram above. Moreover, the number of RBE is rarely the main factor limiting the gaming performance, although 32 raster processors may prove pretty handy in resolutions over 1920x1200 with enabled full-screen anti-aliasing. We are going to discuss it in detail in our Nvidia GeForce GTX 200 gaming performance review that will follow shortly.
I have to say that betting on multi-GPU solutions seems to make much more sense here: dual-GPU ATI solutions as well as Nvidia GeForce 9800 GX2 also have 32 RBE, but they do not require a lot of time and resources for designing a new high-end monolithic GPU.
As we have already said G200 has more RBE units and since each of them is connected with a 64-bit memory controller, the number of memory controllers has also been increased from 4 to 8. This also required a wider memory bus: 512bit instead of 256bit. The amount of memory has also increased. The top model of the family comes equipped with 1024MB of local graphics memory by default. This is where it has an advantage over competitors that are equipped with 512MB of memory by default, which may turn into a serious issue in high resolutions.
Since they introduced 512bit memory bus, they decided not to offer support for any new GDDR types. So, G200, like previous Nvidia chips supports only widely spread GDDR3. Today these chips can work at frequencies up to 1200 (2400) MHz and together with a 512bit bus it gives us a good reason to believe that G200 will not lack any memory subsystem bandwidth.
As for the other innovations, there is not much info about them available, but we know that G200 has got larger caches and more registers. They have optimized the internal interfaces between the functional GPU units and the thread dispatch processor logic. As a result, the computational capacities of the new GPU idle much less.
All this gives us hope that the new graphics solution generation will perform every well in real conditions, but as we have already mentioned, we are going to talk about Nvidia G20 gaming performance in our next article.
PureVideo HD video processor within G200 has come here from G84/G92/G94 as is, without any changes. All previously described innovations, such as support of two video streams decoding, compatibility with Vista Aero interface, dynamic contract and color balance adjustment were made on the software level back at launch of Nvidia GeForce 9600 GT.
The main disadvantage of PureVideo HD compared with ATI UVD/UVD2 is still the absence of fully-fledged support of hardware HD video decoding in a popular VC-1 format. In this case, such resource-hungry task as primary video stream processing including entropic algorithms decoding, is performed by the system CPU. Nvidia G200 also doesn’t have a sound core that has long been an inalienable part of all ATI GPUs with unified architecture since R600. This proves the hypothesis that they designed G200 very hastily, trying to make it a leader in the high-performance GPU segment at any rate.
When working on a new high-performance graphics processor, Nvidia couldn’t help paying special attention to power consumption management tools, because a large and complex chip like G200 cannot consume little power by definition. One of the most important goals was to make it more economical in idle mode and in applications that do not utilize the core 100%, such as HD video playback, for instance. Therefore, they provided G200 with the opportunity to adjust not only the frequency, but also the voltage on its functional units and even disable some of these units if they weren’t needed at the moment.
G200 also fully supports HybridPower technology that we have already described in detail in one of our previous articles. It implies that a graphics card may get completely shut down in 2D mode transferring its functions to the integrated graphics core of the system chipset. For HybridPower to work you need a mainboard based on one of the following chipsets: Nvidia: nForce 780a, nForce 750, GeForce 8200 or GeForce 8300.
New Nvidia graphics core supports dual-precision floating-point operations and hence can be used for scientific, technical and financial calculations that require precision like that. The fact that G200 fully complies with IEEE 754 standard is definitely a plus, although ATI solutions boast this same feature since RV670. The method used in G200 to implement calculations in FP64 format may also pose some problems. As we have already pointed out above, each of the new chip’s shader processors received an additional computational unit, so that there appeared a total of 30 units like that. Their peak performance is only 90Gflops while by ATI RV770 this parameter is almost 2.5 times higher at lower levels of power consumption and heat dissipation. This certainly makes it hard to imagine that dual-processor cards on Nvidia G200 will ever make it to the market.
However, there are very few applications for GPGPU for the end-users: the only task is probably video transcoding that becomes more popular with the spreading of HD formats. This is where the computational capacity of contemporary graphics processors can really do a lot of good.
Making special processors responsible for physical model calculations has been on agenda for a long time and Ageia certainly was a pioneer here. They designed a special chip and released the world’s first add-on physics accelerator called PhysX. This new card didn’t because a serious success because it provided no qualitative breakthrough. Besides, the game developers didn’t hurry to support it. Later on Nvidia acquired Ageia and everyone forgot about hardware physics acceleration for a while. Especially, since in 2007 Intel acquired the developer of another popular physics engine aka Havok, because they cared a lot about promoting their CPUs and not that much about transferring physics calculations to GPU.
Nevertheless, Ageia’s efforts didn’t go wasted: with the launch of G200 Nvidia finally implemented fully-fledged physical model acceleration in the GPU. Moreover, it appeared not only in the new graphics cards, but also in the already existing GeForce 8/9 families. In fact, nothing can prevent hardware PhysX acceleration from working perfectly fine on ATI graphics solutions and all limitations are pure marketing, which a team of computer enthusiasts from NGOHQ.com have recently demonstrated to us.
We can only hope that Nvidia and AMD will not impose any strict bans, but on the contrary will support the use of graphics chips for physics calculations. Everything will benefit from that, that’s for sure.
Now time has come for us to take a closer look at the first graphics accelerators from the new family. We obtained Nvidia GeForce GTX 280 and GeForce GTX 260 from Gainward and Leadtek, so we have to check out their packaging and accessories bundles separately.
We have several times criticized Gainward for using the same design almost for all of their packaging. Luckily, they changed it completely for the accelerators on the new Nvidia chips.
Gainward designers did a great job: new package is very bright, attractive, but at the same time is laconically strict, which suits solutions from the highest price segment best of all. The package is pretty informative, although it is technically wrong to say “1GB DDR3”, because GDDR3 memory used in most graphics cards these days is a sub-type of DDR2 and has nothing to do with JEDEC DDR3 standard. However, this mistake is very widely spread among graphics card manufacturers and suppliers.
The package is very secure and will protect the graphics card inside the same way as before: the card is bubble-wrapped and placed in a small box. Cardboard corners hold this box inside the main box. There is enough room on the sides of the box for accessories. The bundle includes the following items:
If the package design and its protective functionality receive our most positive response, then the accessories bundle of Gainward GeForce GTX280 1024MB seems to be pretty modest. For example, there are no power adapters, such as 8-pin PCI Express, although there are a lot of power supply units out there even among the most powerful ones that still have only 6-pin graphics adapter connector. Nvidia GeForce GTX 280 will not work if you connect a 6-pin PSU connector to the 8-pin connector on the graphics card that is why the absence of the corresponding adapter is considered a serious drawback.
Another serious drawback is the absence of any software for HD video playback. It could be absolutely justified if we were talking about an inexpensive solution like ATI Radeon HD 3850, because bundling software like that could have a significant effect on the pricing and competitiveness of the product. However, for the graphics card from the highest price segment this sort of economical approach is unacceptable. The manufacturers who offer high-end graphics solutions usually try to ensure that there is some pleasant surprise awaiting the customer inside the box that is why we didn’t quite understand Gainward’s view on the accessories bundle. The full version of Tomb Raider: Anniversary game could be considered this “surprise”, but the game is not that new any more and doesn’t really reveal the potential of the new GeForce GTX 280.
Gainward GeForce GTX280 1024MB makes a very favorable overall impression due to new stylish packaging. However, a $499+ graphics card definitely doesn’t deserve a poor accessories bundle like that. We could have closed our eyes to that, if it hadn’t been for the missing power adapters. We hope that Gainward fixes this omission ASAP, which will certainly make this solution a much ore attractive buy for the end-user.
The youngest solution in the new Nvidia family will be represented by a new graphics card from Leadtek, which solutions have received great response from Xbit Labs in the past. Leadtek WinFast GTX 260 has every chance to repeat its predecessors’ success. The card retails in a bright-colored white-and-orange box decorated with a war robot:
This aesthetically modest design is, however, attractive enough. Besides, it really stands out and is quite eye catching. The front of the box is not very informative and the only useful message there is the PCI Express bus version and amount and type of onboard graphics memory. By the way, unlike Gainward, Leadtek states the memory type absolutely correctly. Moreover, here you can also see a mention of the full version of a popular RPG Neverwinter Nights 2 game bundled with the card.
The card and accessories are all very thoroughly packed and are protected with soft inserts. Besides the graphics accelerator, you also get the following items:
In fact, the accessories bundle of Leadtek WinFast GTX 260 suffers from exact same drawbacks as that of Gainward GeForce GTX 280 1024MB: it is pretty scarce and missing some important components. This time, the power adapters are there, although they are not that necessary since the card features common 6-pin connectors that are available on all contemporary PSUs. However, if your PSU only has one 6-pin connector, one of the bundled adapters may be very useful.
What was actually missing in this box, were the DVI-I → HDMI adapter and S/PDIF cable. Of course, not all of you will be connecting the gaming PC to a plasma or large LCD screen, however, the absence of a corresponding adapter in the era of HDMI interface is a pretty serious drawback.
Just like Gainward GeForce GTX280 1024MB, Leadtek WinFast GTX 260 ships without any HD video player software. Since the corresponding adapters are also missing, you will have to spend some extra cash if you want to use this card in a home entertainment center. However, new Nvidia cards are not the best choice for a HTPC anyway: they are too bulky and consume too much power, so they will not feel at home inside most multimedia cases even if they are equipped with a sufficiently powerful PSU.
Leadtek’s Neverwinter Nights 2 game seems to be a better choice than Gainward’s Tomb Raiser: Anniversary, but tastes differ, so I am sure a lot of gamers will prefer the latter. They are almost identical from the technical standpoint and cannot actually reveal the potential of the new graphics cards.
Our conclusion about the package and accessories of Leadtek WinFast GTX 260 will be moderately positive – while the former definitely scores big, the latter is satisfactory at best. Despite lower price of Leadtek WinFast GTX 260 compared with the above discussed Gainward’s card, the company should consider improving their accessories bundle. At least they should include DVI-I → HDMI adapter and S/PDIF cable, and ideally, they could also throw in special software for HD video playback. Anyway, Leadtek’s solution does make a very well-balanced and positive impression.
Now we should move on to discussing the technical peculiarities of the new Nvidia graphics accelerators. There will be only one chapter devoted to this topic, because both models use the same PCB layout and differ just slightly from one another, at least on the outside.
As we have already mentioned several times, Nvidia is attracted to all gigantic. While using a long PCB and a strong power supply system wasn’t absolutely necessary for Nvidia GeForce 9800 GTX, the new graphics cards family on Nvidia G200 tells a completely different story.
Even at first glance both new cards look just the opposite from ATI Radeon HD 4800 illustrating the different approaches to GPU and graphics cards design chosen by two main players of the graphics market. Huge PCB and extremely complex layout are an inevitable necessity resulting from high power consumption of the graphics core. They used not the most advance manufacturing process for the core consisting of 1.4 billion transistors, which is an absolute record. The 512bit memory bus also contributed to the layout complexity, while ATI’s new solutions still use a simpler 256-bit bus. Both cards, GTX 280 and GTX 260, use unified design that makes it rally hard to joggle the prices. Also, graphics cards with proprietary layouts are also very unlikely to appear, because of extremely complex Nvidia GeForce GTX 200 PCB design.
Nvidia GeForce GTX 280/260 hidden inside the cooling system are not very interesting from the technical prospective, because all the details are right beneath the solid massive casing, which is very hard to remove. The front and he back of the casing are held together with crews and special clips, which we have already seen before. For example, the casing of Nvidia GeForce 8800 GT cooler used the same clips.
It is pretty hard to take the casing apart, but we did it nevertheless, and revealed all the most exciting parts of the card.
The voltage regulator circuitry takes up the entire right part of the PCB, about 1/3. As we have expected, it turned out pretty powerful. There is a seven-phase regulator circuitry with a Volterra VT1165MF PWM controller that is responsible for supplying GPU with power. Voltage regulators like that, using almost no traditional electrolytic capacitors, are often referred to as digital. However, it is actually completely wrong: the only difference from traditional regulators is in the higher operational frequency, which allows trading electrolytic capacitors for smaller ceramic ones. It makes the entire system much more reliable. Moreover, it would be very hard to fit a lot of electrolytic capacitors even on a large PCB like that, so the biggest advantages of the new high-frequency voltage regulator are its reliability and small size. Nvidia GeForce GTX 260 has 5 phases instead of 7 and fewer electronic components on the PCB.
The GPU consisting of 1.4 billion transistors and manufactured with 65nm process should boast horrific power consumption that is why it makes total sense to have two additional power supply connectors. Nvidia GeForce GTX 280 has one 8-pin connector that can bear up to 150W. Note that you can’t connect a 6-pin power cable to it – the card will not power on and a red light on the bracket will indicate power problems. The “lighter” GeForce GTX 260 uses the same PCB, but with a pair of regular 6-pin PCI Express 1.0 connectors with 75W capacity. It has no LED indicator either.
Memory voltage regulator is based on Richtek RT9259A chip that we have already come across before in our Gainward and ECS solutions reviews. Since its components are located around the power part of the PCI Express x16 slot, we can conclude that memory received its power from this PCB segment. So, all additional power connectors service GPU only, which is not surprising at all considering its serious appetite for power.
Nvidia used a completely different approach to increasing the memory bandwidth than ATI that bet on increasing the frequency due to faster GDDR5 memory. Nvidia GeForce GTX 200 family had its memory bandwidth increased in a more traditional way: by using a wider 512bit bus for the top model and 448bit bus for the slower model. The latter uses two memory chips less, which determines the corresponding capacity and width. They doubled the number of spots for the memory chips: new cards can accommodate up to 16 chips instead of 8, with half the spots located on the reverse side of the PCB.
Nvidia GeForce GTX 280 memory
Nvidia GeForce GTX 260 memory
Wider memory bus makes the PCB layout much more complex and expensive to produce, however, it is partially compensated by more widely spread and cheaper memory GDDR3 chips. Both graphics cards discussed today use the same Hynix H5RS5223CFR 512Mbit (16Mx32) chips. Nvidia GeForce GTX 280, however, uses chips modification with N2C suffix that operates at 1200 (2400) MHz, while Nvidia GeForce GTX 260 is equipped with slower N0C chips working at 1000 (2000) MHz nominal frequency. In the first case the memory is clocked at a slightly lower frequency of 1100 (2200) MHz, while in the second case the frequency equals the nominal 1000 (2000) MHz. The top model features 1024MB of memory, while the slower model is equipped with 896MHz.
Since Nvidia GeForce GTX 280 and Nvidia GeForce GTX 260 have different memory bus width, their memory subsystem bandwidths make 140.8GB/s and 112GB/s respectively. So, from this prospective, the top model is far ahead of ATI Radeon HD 4870, while the other one yields to it a little bit. However, as we have already pointed out several times with ATI R600 as an example, it is not enough to have significant memory subsystem bandwidth, you should be able to take advantage of it in a smart way. ATI Radeon HD 4870 proved capable of that, however only tests will show what we can expect from Nvidia GeForce GTX 200 here.
Nvidia GeForce GTX 280 GPU
Nvidia GeForce GTX 260 GPU
The size of the graphics processor is truly impressive: it is a real giant, just like G80 was. Unfortunately, we can’t really see because we couldn’t remove the chip heat-spreader easily and decided not to apply any force to it fearing to damage the die. However, the number mentioned earlier will give you a good idea of the size: 576sq.mm. It makes ATI RV770 and Nvidia G92 look like midgets with their 260 and 330sq.mm sizes. This is the price you have to pay for the record-breaking number of transistors and not the most advanced production process. Unlike Nvidia GeForce 8800 GTX, the new G200 based solutions have no metal frame around the GPU edges, although there is a spot for it.
Nvidia’s chip marking is more informative than ATI’s, because it also indicates clearly that the new chip is codenamed G200 and not GT200, D10U, NV55 or NV60. Both chips have A2 revision and are manufactured on week 18 2008, between April 27 and May 3. The interesting thing is that the chip on board of Nvidia GeForce GTX 280 is marked as G200-300-A2, while the chip on Nvidia GeForce GTX 260 – as G200-100-A2. Looks like the middle number has to do with frequency and/or number of operational functional units. With a die that huge G200 yields could be extremely low, so Nvidia may be using not only the chips with lower frequency potential, but also the ones with several defective units.
Nvidia GeForce GTX 280 has all core units up and running: 240 unified shader processors, 80 texturing processors and 32 raster units. Its operational frequency equals 602MHz for the main domain and 1296MHz for the computational one. Nvidia GeForce GTX 260 has more modest features: only 192 active shader processors, 64 texturing units and 28 RBE. Its core frequencies are 576/1242MHz respectively.
Just like by Nvidia G80 based graphics cards, the display controllers are singled out in a separate NVIO chip located to the left of the GPU next to the DVI connectors. This prevented the record-breaking number of transistors in the main die from getting even higher than that.
The marking reads NVIO2-A2, the second version of NVIO chip. However, its functionality remained the same and it still doesn’t support DisplayPort interface. There is a special spot for the translator chip on the card PCB right beneath the MIO connectors. There is also a contact slot laid out for the corresponding connector right above the top DVII port. However, we doubt they will ever roll out Nvidia GeForce GTX 200 with this interface onboard. As for MIO, both Nvidia GeForce GTX 200 models have two of these, so that you could use them in triple-SLI configurations. Their power consumption and heat dissipation will certainly be on the corresponding level, which will barely please even the most risky enthusiasts. SLI connectors are covered with protective rubber caps.
There is a plastic panel above the seven-pin mini-DIN connector functioning as an analogue video out, so you will be able to see the glowing of the LED indicating the power status of your Nvidia GeForce GTX 280 through the slits in the bracket. If everything is OK, it glows green; but if one of the connectors is not plugged in properly or if there is a 6-pin plus in the 8-pin connector socket, it will glow red and the card won’t start. As for the additional functions, we have to mention a two-pin connector next to the power supply connectors that you use to connect to a sound card to translate the sound from S/PDIF to HDMI, because G200 has no proprietary sound core. The only interesting thing about the reverse side of the PCB is the insulation pads beneath the DVI and mini-DIN ports. Since the card is locked within a solid metal casing of the cooling system, these pads prevent electric locking.
You may go farther and fare worse. Nvidia developer team that worked on the cooling system for the new graphics card family understands that well enough. In fact, Nvidia GeForce GTX 280/2600 cooler is yet another variation of ideas that have been first implemented in Nvidia GeForce 8800 GTX cooling solution. It is one of the best reference systems combining excellent cooling efficiency with comfortable acoustics.
Of course, when they adapted the cooling system for the new graphics card family they took into account the peculiarities of the new graphics cards, such as high heat dissipation. As a result, the heatsink became much bigger and the metal external casing also became a part of the heat dissipating system, because as you may notice it warms up pretty significantly during work. Don’t try to grab Nvidia GeForce GTX 280 after it has been running under peak load for a while, because you will hardly enjoy the burning sensation in your fingers.
The top of the graphics processor contacts with a copper base connected with the heatsink via heatpipes. They use conventional thick gray thermal compound as thermal interface. The base has corresponding lugs where it contacts the memory chips, NVIO chip and voltage regulator components. These lugs are topped with fibrous pads soaked in white thermal compound – a traditional Nvidia solution. Since there are a few memory chips on the reverse side of the PCB, too, they also require cooling. The bottom part of the casing works as a cooler. If you have no experience with coolers like that, it is pretty hard to take the cooling system apart, as it is secured with 10 screws and a number of clips. In fact, Nvidia GeForce GTX 280/260 looks like a massive monolithic metal block with only a bracket and a PCI Express x16 slot connector on the outside. The developers scored big for the cooling system design: it is extremely reliable, protects the card against all sorts of physical damage, unless you apply special effort to breaking it, which we seriously doubt.
The system is equipped with a standard radial fan that we have already seen in previous Nvidia products. The airflow it creates cools down the heatsink and ousts warm air outside the system case through the slits in the bracket. The fan uses a four-pin power connector, like the fans of all other contemporary graphics cards. It features a rotation speed sensor and supports PWM control that ensures flexible adjustment of the fan rotation speed depending on the thermal conditions. The fan rotor bears an imprinted Nvidia logo, which may make life harder for the company’s partners who would usually put their own sticker there. In this case the sticker will not hold well and since the fan is rotating it will eventually come off and block the rotor. This may inevitably cause the card to overheat, exactly the same way it once happened to one of our Nvidia GeForce 7950 GX2 solutions, which lower fan failed for exactly the same reason.
Overall, the cooler design is very successful and we have no reason to believe that it may not cope with cooling the new Nvidia GeForce GTX 280, not to mention Nvidia GeForce GTX 260. However, G200 is expected to have extremely high heat dissipation, so before we make any conclusions we have to check out the real power consumption numbers.
Nvidia continues to recommend using high-capacity power supply units for their graphics accelerators. However, as we have already proven multiple times any 400W-450W PSU will be sufficient to ensure their normal operation. They claim that the new Nvidia GeForce GTX 280 solution has peak power consumption of 236W, however, in order to find out how this number correlates with the actual power consumption, we took some standard measurements using a special testbed:
The 3D load was created by running the first SM3.0/HDR test from 3DMark06 in a loop at 1600x1200 with forced 4x FSAA and 16x AF. This test is known to load the GPU beautifully, even though it may be somewhat obsolete for the contemporary technical standards. It suits better than 3DMark vantage tests. The Peak 2D load is emulated by means of the 2D Transparent Windows test from PCMark05. This benchmark is pretty acute because it emulates active work with windows and Windows Vista Aero uses 3D, as we all know. We may replace it with a more popular HD video playback test later on.
This test session produced the following results:
Click to enlarge
As usual, the developer’s numbers turned out way higher. And impeccable operation of the new Nvidia flagship solution on our testbed equipped with a modest 410W PSU proves it.
Of course, Nvidia GeForce GTX 280 boasts pretty serious peak power consumption. However, it not only failed to reach 200W, but never even caught up with Nvidia GeForce 9800 GX2. Nvidia GeForce GTX 260 also demonstrated pretty ordinary power consumption readings: the maximum was only 136W, which is only 6W more than the power consumption of its closest competitor – ATI Radeon HD 4870.
We have to point out that the use of an 8-pin power connector for the top model is absolutely justified: it does receive much more than 75W, which is the maximum allowed workload for a 6-pin connector. Although in 2D mode the new cards drop their GPU frequency to 100MHz and lower the memory frequency significantly, too, they are not very economical in this mode: 65nm process and extremely complex design of the new G200 series chips do their job here. However, ATI Radeon HD 4870 wouldn’t be economical in a similar situation either, although it supports advanced ATI PowerPlay technology.
It is interesting that the Nvidia GeForce GTX 280 power consumption measurements in fact prove that ATI picked the right strategy by betting on multi-processor solutions for the high-end segment. As you know, Nvidia GeForce 9800 GX2 has two G92 GPUs onboard that gives it 256 shader processors, 128 texturing processors, and 32 RBE. In other words, it boasts potentially higher computational and texturing performance than Nvidia GeForce GTX 280 at the same level of power consumption. Anyway, we will b able to capitally impeach or acquit the idea of creating powerful monolithic GPUs for the high-end segment only after we complete our tests in real gaming applications.
The thermal conditions of the new family turned out surprisingly mild: in idle mode the GPUs of both graphics cards heated up only to 45-46ºC, which they certainly owe to the highly efficient cooling system. This system worked perfectly well in 3D mode keeping the GPU temperature around 68-70ºC for Nvidia GeForce GTX 260 and no higher than 72ºC for Nvidia GeForce GTX 280. It is a truly excellent result, however, the price you have to pay for it is the higher level of generated noise. The readings from our Velleman DVM1326 noise-meter were as follows:
The cooling system runs pretty quietly in 2D mode, almost as quietly as the best VGA coolers out there including the ones used on previous generation Nvidia solutions. However once the workload increases, the situation changes dramatically. Nvidia GeForce GTX 280 turns out as noisy as GeForce 9800 GX2, which is not surprising considering its high power consumption. And most noise comes not from the fan motor, but from the airflow hissing through the heatsink. The less powerful Nvidia GeForce GTX 260 is quieter under high workload, however, you can still distinguish its noise against the background of other system components. These are the consequences of using 65nm manufacturing technology.
To be fair we have to say that despite poor acoustic performance, the cooling on both cards is superb. You can certainly sacrifice that for the sake of quieter operation by lowering the rotation speed of the cooler fan using corresponding software. Just make sure the GPU temperature never exceeds 105ºC, because at this point the automatic overheating protection kicks in and reduces the frequencies and hence slows down the cards.
The new cards no longer suffer from any compatibility issues that used to be common for G80 and G92: both cards from Nvidia GeForce GTX 200 family booted and worked perfectly fine on our test boards without PCI Express 2.0, as well as on an Intel X38 based board supporting this protocol.
To study the theoretical potential of the new Nvidia GeForce GTX 200 cards and to compare it against previous generation Nvidia graphics accelerators and ATI Radeon HD 4800 we assembled the following testbeds:
According to our testing methodology, the drivers were set up to provide the highest possible quality of texture filtering and to minimize the effect of software optimizations used by default. Also we enabled transparent texture filtering. As a result, our ATI and Nvidia driver settings looked as follows:
ATI Catalyst:
Nvidia GeForce:
To study the theoretical performance of our today’s testing participants we used the following applications:
ATI Radeon HD 4800 have no chances in “pure” fillrate tests, because they have much fewer texturing processors and RBE/ROP units than Nvidia GeForce GTX 200. New solutions on Nvidia G200 are far ahead and as always, prove their best in operations with Z-buffer and stencil buffer. Their performance increases almost linearly compared with Nvidia GeForce 9800 GTX, because all the differences are of purely quantitative nature.
However, don’t get too excited and predict an indisputable victory in games for the new Nvidia solutions basing only on the results of this test. As you know, Nvidia uses texture processors with two texturing filtration units per each two addressing units, that makes “free” anisotropic filtering impossible. According to our practical experiments, you should divide the number of Nvidia’s texture processors by 2 every time anisotropic filtering is used. As a result, 80 texturing units of the new Nvidia GeForce GTX 280 will in fact work as 40.
Yes, new Nvidia solutions have a serious advantage over ATI Radeon HD 4800 family: they have twice as many RBE/ROP units. However, this parameter doesn’t really affect the gaming performance any more. It may have certain influence on the performance only if anti-aliasing is enabled in resolutions over 1920x1200. In fact, you shouldn’t really base any conclusions on the results of this test, because they have no connection with reality and only indicate that G200 is today’s fastest monolithic GPU when it comes to fillrate. The gaming performance is affected by much more factors, namely, GPU’s computational capacity. Although the only thing we know about it at this time is that it is definitely superior to G92.
Marko Dolenc’s Fillrate Tester doesn’t support even Shader Model 3.0, but suits very well for analyzing graphics architecture performance with older shader code versions. The results obtained in this benchmark may be pretty valuable for those who still play old games. Besides, they give a more or less good idea of the GPU potential.
In fact, all above mentioned shaders test the RBE performance, which is no issue for G200. However, the GPU performance drops dramatically on more complex shaders and gives in completely during per-pixel lighting emulation. Moreover, Nvidia GeForce GTX 280 solution loses even to ATI Radeon HD 4850. The card priced at $499+ yields to a budget solution going for only $199 – what a disaster!
Shader Particles test from 3DMark06 suite is not a fully-fledged graphics test as it emulates a physical model of massive particle systems’ behavior. Collision calculations are performed using pixel shaders and the result is displayed on the screen using vertex shader texture samples. Nevertheless, this test can measure the mathematical GPU performance just fine.
Here the new Nvidia graphics card family seems to be a little better off. However, only lower GPU frequency by ATI Radeon HD 4850 saves the new Nvidia solution from fiasco. However, ATI Radeon HD 4870 seems to be much harder to beat, although Nvidia GeForce GTX 280 almost manages to do it in 1920x1200. This is more than modest performance for a $499+ graphics card with 180W power consumption.
Looks like similar test from 3DMark Vantage suite employs only 1/5 of ATI Radeon 4800’s computational capacity. As a result, Nvidia G200 manages to take the lead. Although this victory is not the success of G200. To be more exact, its only advantage is the fact that scalar architecture from Nvidia doesn’t depend as much on software optimizations as super-scalar architecture from ATI.
Perlin Noise test from 3DMark06 suite estimates the GPU performance in the most complex Shader Model 3.0. It generates a texture using 48 texture samples and 447 mathematics instructions:
As we have expected, TI Radeon HD 4800 is beyond competition here, while Nvidia GeForce GTX 200 performs very modestly. The top model from the new family again lost to the younger fellow from the rivalry camp, although the gap between them is not that big at all.
Increasing the number of texturing units or RBE units will not help to make up for the lower computational capacity. Unfortunately, it is too late and they no longer can fix the power misbalance in the new G200. We can only hope that the new solution will not be so hopeless in real games.
3DMark Vantage POM test displaying the complex landscape with parallax occlusion mapping method again gives the lead to Nvidia GeForce GTX 200.
Although ATI Radeon HD 4850/4870 outperform GeForce 9800 GTX, Only the top model of the two can actually catch up with GeForce GTX 260. We don’t know why, however these could be either software optimizations, or hardware issues. In either case it is very sad outcome for ATI Radeon HD 4000.
Shader Math test is a slightly more complex version of Perlin Noise from 3DMark06.
Here Nvidia GeForce GTX 200 solutions manage to win a formal victory, which indicates that both newcomers boast pretty good mathematical performance. However, the top model, Nvidia GeForce GTX 280, is barely ahead of ATI Radeon HD 4870, while the younger Nvidia GeForce GTX 260 model yields about the same little bit to ATI Radeon HD 4850. So, despite greater complexity of the new Nvidia GeForce GTX 200 cards, they are not always far ahead of the simpler designed ATI Radeon HD 4800.
X-bit Mark test illustrates remarkably well both: mathematical GPU performance and their architectural success. It includes shaders made of mathematical instructions as well as shaders with a lot of texture samples.
This is a complete fiasco for GeForce GTX 200. Nvidia GeForce GTX 280 lost to ATI Radeon HD 4870 in all tests. Moreover, in some of the tests such as Plaid Fabric, 27-Pass Fur and Dynamic Branching – it was a really bad beating. Nvidia GeForce GTX 260 also lost to ATI Radeon HD 4850, though not as badly and even got ahead of the competitor in NPR test (hatch, 8 textures).
So, we managed to reveal the major weakness of the new Nvidia G200: its computational unit working at 1GHz+ frequency doesn’t always bring in the victory. Remember that computational processors are the ones determining the performance in contemporary games that use a lot of shader effects with numerous complex calculations. However, it is indeed very hard to compete with new ATI solutions featuring 800 ALU.
The results are not very optimistic so far: new Nvidia cards didn’t impress us with their mathematical talents, which will inevitably affect the gaming performance in far not the best way. These results are actually pretty strange for solutions aiming at indisputable leadership.
In this test Nvidia GeForce GTX 200 family doesn’t act logically at all: it yielded not only to ATI Radeon HD 4800, but even to its predecessor - Nvidia GeForce 9800 GTX! The only explanation we could come up with – lower clock frequencies resulting from extreme complexity of the new G200 chip, which do not get compensated by increased computational capacity. It could be a software issue at best, but in the worst case these results may indicate the presence of some architectural issues. No good news for the newcomers.
Nvidia GeForce GTX 200 behaves exactly the same way in the vertex shaders processing test. Both models turn out in the last places. However, these are theoretical tests, and in real games things may be different.
Even if we disregard the fact that Nvidia GeForce GTX 200 won in 3DMark Vantage thanks to idling of 4/5 of ATI Radeon HD 4800’s computational capacity, we will still notice that the improvement is minimal even compared to Nvidia GeForce 9800 GTX. The results demonstrated by Nvidia GeForce GTX 260 show it best of all.
X-bit Mark geometry benchmark doesn’t hold back ATI Radeon HD 4800 and the cards on RV770 GPU get far ahead. The gap between the leaders and Nvidia GeForce GTX 200 reduces in scenes with 8 light sources, although ATI cards remain ahead. Moreover, new Nvidia graphics cards perform less efficiently than their predecessor, Nvidia GeForce 9800 GTX, and even bigger number of shader processors doesn’t help.
All this once again proves that new Nvidia solutions have certain problems with geometry processing. As we have already said, these problems may be caused by lower clock frequencies of the new Nvidia GeForce GTX 280/260 solutions. However, frequencies only couldn’t possibly cause such a dramatic failure.
Overall, Nvidia GeForce GTX 200 family cannot boast much in geometry benchmarks. In most cases G200 demonstrated considerably lower potential than ATI RV770 GPU, even despite gigantic chip size and unprecedented amount of transistors used. New Nvidia GPU is evidently not so well-balanced: with the die size approaching 600sq.mm G200 features extremely large texturing and raster blocks at the expense of computational capacity. Only the results obtained in real gaming applications will be able to rehabilitate the new Nvidia GeForce GTX 280 and Nvidia GeForce GTX 260 solutions.
As we have revealed in our ATI Radeon HD 4850 review, ATI and Nvidia solutions provide very high anti-aliasing quality these days. Since GeForce GTX 200 inherited its FSAA algorithms from the predecessors, it doesn’t make sense to discuss any game screenshots one more time. We suggest focusing on another matter: graphics card performance with enabled FSAA.
Theoretically, Nvidia G200 should easily cope even with the most extreme FSAA modes thanks to 32 RBE units and 512-bit memory access bus. Nevertheless, we have to check it out in tests, before drawing any conclusions.
It makes sense to discuss only 1920x1200 resolution, because in lower resolutions enabling FSAA has minimal influence on the performance except a few extreme modes, such as ATI edge detect CFAA.
Nvidia CSAA modes differing from the classical MSAA only by bigger number of samples on the grid, are pretty liberal when it comes to the video memory capacity and bandwidth. And our results prove it very well. When Nvidia GeForce GTX 280 switches from classical MSAA 4x to CSAA 16xQ, combining MSAA 8x and 16 grid samples, the performance drops only 10%, which is truly impressive. At the same time, ATI Radeon HD 4850 lose just a little more,12%, when switching from MSAA 4x to MSAA 8x. However, it features half the RBE units and has no 512-bit memory bus.
In other words, we see once again that even with smaller resources available one can often achieve very decent results if using these resources in a smart way. Almost as good as in case one simply increased the capacity without introducing any architectural optimizations. The results demonstrated by Nvidia GeForce GTX 260 prove it, too: it lost 18% of speed when switching from MSAA 4x to CSAA 16xQ, although it features a 448-bit memory access bus with 112GB/s bandwidth, which is much higher than by ATI Radeon HD 4850.
All in all, we see that Nvidia’s CSAA algorithms provide an almost ideal compromise between performance and anti-aliasing quality. They barely affect the performance of the new GeForce GTX 200 family, so the owners of new graphics cards shouldn’t ever question using them, while the GeForce 9800 GTX users may want to reconsider even enabling the common MSAA 4x.
At first glance, the new Nvidia GPU that tends to be the world’s fastest monolithic graphics processor these days made very ambiguous impression. Unlike ATI RV770, it cannot boast that many innovations compared with the previous generation graphics processors, so we can’t call it a revolutionary. G200 is more likely to be considered further extensive development of ideas first brought up in G80. Moreover, we sometimes get the impression that they were in a real hurry when designing this chip. In fact, there was no need for rush, because ATI chose a completely different strategy and gave up the high-performance monolithic GPU concept altogether.
Actually, G200 could be regarded as “G92 on steroids”. Just look at the increased number of all functional units: ALU, TMU, RBE, and wider 512bit memory bus. The only significant architectural change is the addition of the third shader processor into computational clusters that use to have only two of those.
The results of preliminary theoretical benchmarks turned out not very optimistic. New Nvidia’s solution lost to a simpler and cheaper ATI RV770 in a lot of synthetic benchmarks except the fillrate test and texture sampling pure performance. Theoretically, G200 based solutions should feel at home in those games that have a lot of high-resolution complex textures and shaders working mostly with textures, and should lose to ATI only in those games that require high mathematical performance. Moreover, 512-bit memory bus and 32 RBE may be extremely useful in high resolutions with enabled anti-aliasing, which will definitely attract a number of hardcore gamers.
We don’t quite understand why Nvidia bet on texturing performance. This parameter becomes less important because most games are developed for multiple platforms right from the start. Since gaming consoles have limited hardware resources, the main focus here is on special effects that use a lot of complex mathematical calculations on few texture samples. Even the graphics is initially designed for lower resolutions than those typical of contemporary PC platforms.
However, what made the new Nvidia G200 solution truly arguable, is not choosing texturing performance over arithmetic, or monolithic GPU design, but a combination of both. As a result, the new GPU die grew enormously big and complex: 576sq.mm and 1.4 billion transistors. However, it still appeared less powerful than a pair of G92s in an SLI configuration. Gigantic size and unprecedented complexity made this GPU extremely expensive to make and lowered the production yields. Besides, G200 requires complex PCB layout because of 512-bit memory bus. As a result, the cards based on this new GPU, Nvidia GeForce GTX 280 and Nvidia GeForce GTX 260, turned out pretty expensive and Nvidia had to drop the prices right after the launch of ATI Radeon HD 4800. It was certainly painful for them, if we take into account a lower margin they ended up with.
Nvidia G200 may become the last premium graphics card based on a single mega-chip, because the company has already faced pretty serious difficulties: look at the lower frequency of the new GPU that caused it to lose in several benchmarks. Further increase of the GPU potential may lead to even greater problems.
As for the specific graphics cards we have discussed today, Nvidia GeForce GTX 280 and Nvidia GeForce GTX 260, we will be able to share our final verdict only after releasing the gaming benchmark results. According to our preliminary estimates, the top model has every chance to become a performance leader at least until ATI Radeon HD 4870 X2 comes out. However, the performance advantage over ATI Radeon HD 4870 will most likely be not as dramatic as Nvidia had hoped. As for Nvidia GeForce GTX 260, it looks pretty competitive priced at $299, although its pretty large size may prevent it from becoming a gaming bestseller (it may not fit in a number of system cases).
Gainward GeForce GTX280 1024MB belongs to the top price range, but its relatively scarce accessories bundle slightly spoils the overall impression:
Highs:
Lows:
This solution may become pretty attractive, if it retails for no more than ATI Radeon HD 4870. Just like in the previous case, we were very pleased with stylish quality packaging, but upset about scanty accessories with a few important items missing.
Highs:
Lows:
When you consider the highs and lows for Gainward and Leadtek graphics cards, please, keep in mind that GeForce GTX 200 is targeted for gamers in the first place. Therefore, DVI-I → HDMI adapter, integrated sound core and software player for Blu-ray/HD DVD movies may not be a necessity at all.