A new GPU with guarded secrets

A new chip

To allow the introduction of a lower priced GeForce 8800, Nvidia developed a new GPU based on GeForce 8 architecture while being conceived with lower production costs in mind. First of all, the NVIO, or chip in charge of managing the video outputs on the GeForce 8800 Ultra/GTX and GTS, has disappeared and is now integrated to the GPU. Second, the new GPU, the G92, was manufactured with the 65 nanometer production process which is a first for Nvidia as the G80 (previous GeForce 8800s) were engraved in 90nm and the G84 and G86 (GeForce 8600 and GeForce 8400) were in 80nm. The reduction in the width of engraving allowed the production of smaller and consequently less expensive chips.

Going from 90 to 65nm allows almost doubling the density of transistors and therefore to almost half the size of the chip. But the G92 isn’t simply a finer version of the G80 and there were other modifications. Unfortunately, Nvidia carefully avoided responding to any detailed questions concerning these differences and preferred to point out the main "features" of this product, especially careful not to speak of the chip itself. Ujesh Desai, General Manager for the GeForce, even went as far as saying, "G92 ? I don’t know it. What is that?" to a question concerning the G92. Those at the conference wanted to respond, "G92 is maybe what is written under the graphic card’s fan?". Nvidia actually seems to have decided to avoid speaking of chips and simply mentions the code name of the product line. In other words, goodbye G92 ("GPU capable of x with y number of units"), and hello D8P ("product at 200€"). This leaves more flexibility to the marketing department to adjust products and hide details from the competition as long as possible. The flip side is that tech journalists are left unsatisfied because it’s a bit annoying to not know the details of such a good GPU.


The diagram of the GeForce 8800 GT’s units when the actual chip theoretically contains more.
So be it. In terms of the GeForce 8800 GT’s innovations unveiled by Nvidia, there is now PCI Express 2.0 and VP2, the second generation video engine which handles H.264 acceleration at 100%. As for specifications, 112 scalar processors and 16 ROPs were announced for the GeForce 8800 GT compared to 128 scalar processors and 24 ROPs which make up the G80. It’s possible that the G92 contains at least 128 scalar processors but we believe the ROPs are indeed limited to 16 on the chip level just like the memory bus is limited to 256 bits. However, this is only a guess as Nvidia didn’t confirm any specific details.

In spite of everything, there were a few small technological details unveiled by Nvidia ; the scheduler has been improved in order to optimize load balancing as well as the structure of texturing units. While in the G80 there are blocks of 4 address processors each associated to 2 filtering units, in the G92 there are blocks of 8 address processors associated to one filtering unit as was already the case in the G84 and G86. Overall, this allows doubling the capacity of the GPU to access low precision textures when they are not filtered or limited to bilinear filtering. There is no difference in terms of the rate of complex filtering as well as for access to higher precision textures (FP16, FP32).

More than what Nvidia announces ?

While the G80 is composed of 681 million transistors, the G92 has 754 million. The G80 – 8 ROPs – 128 bit memory bus + NVIO + VP2 + address and texture processors should all add up to that figure ; however, it isn’t that simple and « 754 » seems a bit high. This is because with the G80’s 681 million transistors, there is most likely some percentage that aren’t used. A first GPU from any given generation always has bugged functions or ones that are judged useless in the end. Some units are not implemented in an optimal way and will be improved upon at some future time, etc. For this reason, this higher number of transistors leaves a number of unanswered questions as to the exact composition of the G92.


More than 16 ROPs and a 256 bit memory bus ? More than 128 scalar processors? A tessellation unit ? Support of double precision on floating numbers? Previously, Nvidia announced that they would release a GPU with this last capability this year (the G92 moreover was specifically mentioned) and it was specified that it would only be accessible through CUDA on Tesla and Quadro products. We have therefore been waiting for a development based on this information. When asked about this, Nvidia’s response was a simple "No, there is no double precision support." Is this just vague marketing meaning that the final "GeForce 8800 GT" doesn’t support this technology while the "G92" chip does? Or should we understand that Nvidia is backtracking due to a potentially troublesome implementation and prefers to wait for a future product where it will be better implemented ?

Specifications, the card, power consumption, the test

Spécifications

For specifications, the GeForce 8800 GT is close to the GeForce 8800 GTX in terms of calculation power and texture filtering : however, it is very far in memory bandwidth as its bus is limited to 256 bits.

The card

The GeForce 8800 GT has a single slot cooling system with a rather nice design. On the other hand, its efficiency is lower than what we find on the 8800 GTS/GTX and Ultra systems but this is only logical as they are much more imposing. For this reason, the fan accelerates more often with the GT but this is nothing too dramatic. In idle, the card is quiet, slightly less than an 8800 Ultra and it’s still in the "silent" class. In action, the fan can be heard when it accelerates but produces much less noise than, for example, a Radeon HD 2900 XT.

There is 512 MB of 1 ns GDDR3 memory by Qimonda. The card "only" has a PCI Express 6 pin power connector. On the other hand, there is only one single SLI connector which makes the GeForce 8800 GT incompatible with the soon to be available 3-way SLI.

Note that this is a reference version. The first cards available in stores will be similar and in the beginning Nvidia will only deliver completed models to its partners. This allows it to reveal nothing or very little of the finished product while at the same time making it available for the launch. After this, it’s up to the partners to come up with their own design.

In terms of overclocking, we went from 600/1500/900 MHz (GPU, shader core, memory) to 690/1700/1060 MHz, which represents an overclocking of around 15%. We expected slightly better, but this result is average for graphic cards. Riva Tuner 2.06 beta was used in order to be able to independently overclock the 3 frequencies of the GeForce 8800 GT.

Power consumption

We evaluated the power consumption of the different cards. Measurements were taken at the wall socket, in order to have the total power consumption of the power supply, in this case, an Enermax Galaxy 850W.


A large advantage of 65 nanometers is that power use is greatly reduced. While consumption is still relatively high and requires a PCI Express connection we are very far from the extremes of the GeForce 8800 Ultra and Radeon HD 2900 XT.

On a positive note, power consumption in standby is strongly reduced and the GeForce 8800 Ultra uses 50 watts more than this 8800 GT !

The test

For this test, we used 10 games, four of which support DirectX 10. All tests were carried out in 1920x1200 and 1280x1024 with anisotrope filtering activated when an option in the game. HDR was activated every time when available.

Finally, transparency/adaptive antialiasing were activated in multisampling mode as Nvidia has just implemented a new and more efficient version in its drivers. For AMD, it is forced in many games and its activation is now possible via the control panel.

Of course, all updates to Windows Vista relative to performance were installed.

Configuration

Intel Core 2 Extreme QX6850
nForce 680i SLI EVGA
2 GB DDR2
Windows Vista
Forceware 169.04 (169.05 for Crysis)
Catalyst 7.10

HD Video: performances and quality

HD video playing

While the first cards of any given generation are generally very poorly equipped for HD video, luckily AMD and Nvidia take the time to correct this point on the following versions. In this way, while the GeForce 8800 GTS/GTX/Ultra are not very efficient in terms of the acceleration of HD video decoding because they have the VP1 video engine of the GeForce 7, the GeForce 8800 GT has the VP2 like the GeForce 8400 and 8600.

We measured the performances of the different solutions in reading two HD DVDs, one in VC1 (King Kong) and the other in H.264 (Babel). The test configuration here changed and was now based on a Core 2 Duo E6400. For software, we remained in Windows Vista and used Power DVD version 3104.


The GeForce 8800 GT holds to its promises and with the 8600 GTS is far ahead of the other high end solutions.

We also took a look at video quality and based our evaluations on the analysis tests of the HD HQV suite (go herefor the details of these tests).


The GeForce 8800 GT has the same behavior as the GeForce 8600 GTS. The rest of the GeForce line has problems with inverse telecine and doesn’t manage to reproduce these videos with satisfactory quality.

Only the Radeon HD 2900 XT obtained a score of 100, the difference being in terms of digital noise reduction. For Nvidia, this is activated with drivers in the control panel with a setting between 0 and 100 and set too high it can produce visual defects. We settled for 51%. Note that this test is subjective and that in some videos it’s best to be able to deactivate this effect when noise can benefit display of the movie. AMD does not allow deactivation or the adjustment of this parameter.

Enemy Territory : Quake Wars

Enemy Territory : Quake Wars

While Quake Wars is based on the Doom 3 engine, it has undergone some evolution such as megatexturing which facilitates the work of artists; however, there is the additional cost in terms of decoding and access to megatextures. In the end, Quake Wars is a little more resource heavy than Doom 3 or Quake 4.

We saved a demo in a sequence versus 4 bots. Given that artificial intelligence was not calculated in the timedemo, results were less affected by the CPU than in actual gameplay or at least in this case versus our bot adversaries.

All parameters were set to a maximum in the game including 16x anisotropic filtering. The patch 1.2 was used.


In this first test, the GeForce 8800 GT finds itself halfway between the GTS and GTX. The Radeon HD 2900 XT is behind and we can see that the GeForce 8600 GTS is clearly not in the same league.


Once antialiasing was activated, the performances of the GeForce 8800 GT fall, theoretically, because of its reduced memory bandwidth. However, this could also be a lack of memory as is the case with the GeForce 8800 GTS 320.

Half Life 2 - Episode 2

Half Life 2 Episode 2

Still based on the Source Engine, Half Life 2 Episode 2 doesn’t really have anything new on the technological level. It simply optimizes and more heavily relies on the engine’s capabilities, making the game more resource heavy than its previous versions. We carry out a demo with all game options set to a maximum including anisotropic filtering which is in 16x.


The GeForce 8800 GT is on the heels of its bigger sibling and clearly puts some distance between it and the 8800 GTS. The 8600 GTS is at a third of its performance level.


It’s the exact same situation once 4x antialiasing is activated while the Radeon HD 2900 XT’s performances seem slightly more affected.

S.T.A.L.K.E.R.

S.T.A.L.K.E.R.

We carry out an identical movement and measure the framerate with fraps. The test was done in high quality, complete dynamic lighting, and high detail followed by another session in medium quality, complete dynamic lighting, maximum details (anisotropic filtering 16x) and grass' shadows. S.T.A.L.K.E.R. uses an engine based on differed rendering, which is fundamentally incompatible with MSAA and makes the use of antialiasing impossible. A type of filtering of edges carried out with a shader can be activated but results are mixed. The 1.00004 patch is used.


The GeForce 8800 GT is situated between a GTS and GTX. The GeForce 8800 GTS 320 is largely behind given that 320 MB is insufficient for this game.

Rainbow Six : Vegas

Rainbow Six : Vegas

The first PC game based on the Unreal Engine 3.0, Rainbow Six : Vegas is still a very resource heavy game. We measure performances in the introductory scene. The HDR mode is activated as it is more or less obligatory, because without it banding is very noticeable. Shadows are set to “low” because a higher quality in this domain lowers performance too much in certain areas.


Initially designed for the Xbox 360, this game seems to have a natural affinity to the Radeon HD 2900 XT which has a similar architecture to the game console’s graphic chip. It’s no surprise then that the Radeon finishes first this one time. The GeForce 8800 GT is very close to the 8800 GTX as calculation power is of high importance here.


The game does not support antialiasing but Nvidia has implemented it to drivers like it does with Oblivion. This is contrary to AMD who unfortunately didn’t make this effort. This time the GeForce 8800 GT is situated halfway between the GTS and GTX. The GTS 320 does not have enough memory to launch the game in 1920x1200.

Oblivion

Oblivion

We saved a specific movement in order for it to be always identical and the test reproducible. Of course, HDR was activated.


Given that Oblivion is very dependent on calculation power in HDR mode, the GeForce 8800 GT places close to the GTX.


The standing is similar with FSAA 4x.

Colin McRae DIRT

Colin McRae DIRT

To test Colin McRae‘s latest opus which is very resource heavy we carried out a well defined sequence in high quality mode. Note that the activation of antialiasing is highly recommended given the way the menus are rendered and because post processing effects amplify aliasing. The patch 1.2 was used.


Once again, the GeForce 8800 GT places between the GTS and GTX. The Radeon is equivalent to a GeForce 8800 GTS while the GeForce 8600 GTS is largely behind.


No big change with antialiasing except that the 320 MB of the GeForce 8800 GTS 320 aren’t enough.

Bioshock

Bioshock

The first game based on the Unreal Engine 3.0 to support DirectX 10, Bioshock has great graphics even in DirectX 9 mode while it is less resource heavy than Rainbow Six : Vegas. We carry out a well defined sequence of movement with all options pushed to a maximum.


In DirectX 9 mode, the Radeon HD 2900 XT is rather fast and very close to the GeForce 8800 Ultra. The other performances aren’t too surprising.


Like with Rainbow Six : Vegas, Nvidia allows the activation of antialiasing for this game which doesn’t normally support it. AMD doesn’t offer this option. For some unknown reason, it was impossible to activate antialiasing on the GeForce 8800 and 8600 GTS.


In DirectX 10 mode, GeForce 8 performances are similar to what we obtained in DirectX 9 while performances mysteriously plummet for the Radeon.

Note that it is not yet possible to activate antialiasing in DirectX 10 mode for Nvidia and AMD.

Company of Heroes

Company of Heroes

Given that Company of Heroes received a DirectX 10 patch that adds a real plus on the graphics level, we decided to add it to our test protocol. All graphic settings were pushed to a maximum except for terrain details which remained on High (Ultra mode is reserved for DirectX 10). Textures were also limited to High, because the game indicates a lack of system memory for the GeForce in DirectX 9 mode with Ultra textures.

We run the integrated test on version 1.71.


The GeForce 8800 GT is just barely ahead of the Radeon and largely surpasses the 8800 GTS.


Once 4x antialiasing is activated the performances of the Radeon HD 2900 XT plummet while those of the GeForce 8 line show consistent results.


In DirectX 10, performances are clearly inferior as the engine displays more advanced graphics. The Radeon is equivalent to an 8800 GTS 320 and behind the 8800 GT.


The standings are identical with antialiasing and a comparatively larger reduction in performance for the Radeon HD 2900 XT. Note that the GeForce 8600 GTS shows its limits and is very far behind.

World in Conflict

World in Conflict

Very resource heavy and with nice graphics, it’s only natural World in Conflict joins our test suite. We carry out the internal test with the patch 1.0002. All game options are pushed to a maximum.


The Radeon HD 2900 XT doesn’t shine in this game and is behind the GeForce 8800 GTS 640 MB. The 320 MB version is out of the race due to insufficient memory.


With antialiasing 4x, the GeForce 8800 GT seems to be only slightly limited by its memory bandwidth and is not that far from the 8800 GTS.


While GeForce 8 performances are slightly reduced in DirectX 10 mode, those of the Radeon HD 2900 XT increase placing it equivalent to a GeForce 8800 GT.


This is the one and only time that the GeForce 8800 GTS 640 MB surpasses the 8800 GT.

Crysis - demo SP

Crysis - demo SP

Although we don’t usually use demos in our test suite, it’s hard to pass up Crysis due to the fact that this game is so much awaited and redefines modern graphics quality. We carry out the internal test in DirectX 9 High mode (forced via the addition of –dx9 behind the executable)and in DirectX 10 Very High mode.

Note that there was some polemic regarding a bug in Nvidia’s 169.04 driver related to deformed reflections on the water surfaces which weren’t updated enough. Due to the fact that it was calculated less often, performances improved or at least in certain conditions like in 1920X1200 with AA4x in DirectX 10 mode (+30%). In the beginning, we thought this was a bug in the game itself ; however, it showed that by renaming the executable in order that the Nvidia driver could no longer detect it, the problem was corrected. For this reason, this bug seemed a bit suspicious.

We contacted Nvidia about this and they confirmed the presence of the bug and the related gain in performances it could add. Nvidia provided us with a driver that no longer has this problem (169.05), which of course we used here. They also gave precise information on the bug in question in order that there be no further confusion on the subject.

Crysis’ engine tries to avoid calculating water reflections as much as possible, which is only logical. To do this, three pieces of information are used: the time since the last update, the movement of the camera, and the number of "water" pixels visible on the screen. Of course, it would be ridiculous to recalculate an entire reflection for a handful of pixels. Moreover, it is this last value that is the source of the problem.

To determine the number of visible "water" pixels, the engine carries out an occlusion test in the GPU, returns the result to the system which uses it to determine if the reflection should be updated or not. To avoid wasting time (the CPU waits for the result from the GPU) for any given image, the engine relies on the occlusion test result from the previous image, which can, for example, be one image behind. In SLI, this can be a problem as a one image delay can mean that GPU2 will then have to wait for GPU1 to carry out the occlusion test. There is therefore no performances gain with SLI. The demo (and most likely the game) will be updated to take into account multi-GPU systems by synchronizing the utilization of occlusion tests in order to avoid blocking part of the system.

In the meantime and so that we can see the gains in SLI (potentially causing other bugs), Nvidia in a way falsifies the occlusion test by always giving the same information to the engine: the water is visible but only on a very small number of pixels. In the case where the test actually indicates that water is not visible, the GeForce 8 needlessly processes the water (obviously without updating the reflection). In the case when the test detects that there is a large amount of water that is visible and a reflection to be updated, the GeForce 8 indicates that there is still very little visible and the bug appears. In the end, water can be a significant part of the screen in the demo’s test and the GeForce 8 benefits from this.

However, how is an SLI optimization related to the problem in question? There is an indirect effect because here a bug in the driver applies an SLI profile even when there is a single GPU. And this is what Nvidia corrected with its 169.05 driver. So this is still a problem in SLI and will remain one until this game has been updated for this to function properly.

Having pointed this all out, we can now move on to test results.


The GeForce 8800 GT is close to the GTX and just ahead of the Radeon HD 2900 XT. Note that only the GeForce 8800 GT/GTX/Ultra allows playing in High mode and 1280x1024 and even then just barely. This game is very demanding on all levels including for memory because the 320 MB of the GeForce 8800 GTS 320 is already insufficient.


With antialiasing, a number of cards start to fall off. While the GeForce 8800 GTX and Ultra manage to do alright, the GT plummets in 1920x1200. As for the HD 2900 XT, its performances are cut in half!


With all details pushed to their limits and in DirectX 10, the GeForce 8800 GT has trouble putting distance between it and the 8800 GTS 640 MB in 1920x1200. However, it is still ahead of the Radeon HD 2900 XT.


Playing with all settings set to a maximum in 1280x1024 with the current biggest high end card? Impossible and it’s a humbling experience for all cards. We can’t wait for the GeForce 9800 in SLI!

Recap

Recap

Although individual game results are interesting, we calculated a performance index based on all tests with the same weight for each game. A 100 was given to the GeForce 8800 GTX in 1280x1024.


On average the GeForce 8800 GT is 10% behind the GeForce 8800 GTX, which is quite good considering the card sells for half the price. It places closer to the GTX than to the GTS and is very far ahead of Nvidia’s previous upper mid-level card, the GeForce 8600 GTS. The GeForce 8800 GT also finishes ahead of the Radeon HD 2900 XT.


With antialiasing 4x filtering activated, the GeForce 8800 GT places a little closer to the GTS or farther from the GTX, depending on the game. Either way, performances are still very good. The GeForce 8800 GTS 320 MB is clearly limited by its memory in this area. As for the Radeon HD 2900 XT, its bugged ROPs in terms of antialiasing forced AMD to use an alternative solution which is very flexible but it stifles performances.

Note that the average scores of some cards was reduced because they aren’t capable of antialiasing in Rainbow Six : Vegas and Bioshock. However, in our opinion, excluding these games would not change the standings. You can consult this possibility here.


In DirectX 10, the GeForce 8800 GT isn’t disappointing and is still between the GTS and GTX and ahead of the Radeon HD 2900 XT.


In DirectX 10 with FSAA 4x, the GeForce 8800 GT is more limited most likely by its memory bandwidth ; however, it is still equivalent to a Radeon HD 2900 XT and GeForce 8800 GTS 640 MB which all finished with more or less the same scores. In these extreme conditions, the GeForce 8800 GTX and Ultra hold an advantage.

Conclusion

Conclusion

After reading this article, it’s obvious that Nvidia hits hard with this GeForce 8800 GT. While high end products from Nvidia’s first DirectX 10 architecture convinced us, this was not the case for mid-range versions. Indeed, the GeForce 8600 GT and 8600 GTS left something to be desired with their slightly deceiving performances. This meant there was no real DirectX 10 alternative to a rather expensive high end because Nvidia was unbothered by the competition.


In terms of pricing, this GeForce 8800 GT places itself between the high end and mid-range while offering high end performances. For 230 to 250€, you will get between 75 and 90% of what the GeForce 8800 GTX is capable of which is double the price. This is a good deal for gamers, all the more so that we are currently seeing the arrival of many amazing games.

However, improvements weren’t only on the gaming level. Nvidia addressed power consumption, which was significantly reduced thanks to a more modern production process, as well as integrating a more evolved video engine which enables the GeForce 8800 GT to be the first high performance graphic card adapted to HD video.


Is this the perfect card? This is an impossible question a few days before AMD’s release of the Radeon HD 3800 with its future support of DirectX 10.1. We therefore will delay answering this question until our upcoming article which will compare the two solutions. Nevertheless, it’s evident that it will be difficult for AMD to rival the GeForce 8800 GT in terms of performances at this price level.

We finish with one last small detail : the prices announced by manufacturers, especially Nvidia. Given that this is a very important factor when analyzing a product, for some time now GPU manufacturers announce very low prices that don’t correspond to the current market. Thus Nvidia announced a launch price of $199 (or 179€ taxes included) and assured us that cards would be available for this amount. After having checked with its partners, we realized (unsurprisingly) that they were totally incompatible with a retail price of $199 even if the store or dealer didn’t add any costs! Nvidia then said that they were talking about a future price and then said this was for upcoming card equipped with 256 MB. We point out that Nvidia Europe, which knows that the tech press does not like to be mislead in this manner, announced a more realistic 219 to 249€. Either way, for this price the GeForce 8800 GT is an excellent product and we can only wonder about the motives of Nvidia US. Maybe to scare AMD ?