by Alexey Stepin , Yaroslav Lyssenko
12/03/2008 | 01:48 PM
The GeForce 9800 GTX graphics card didn’t have a very happy fate. This follower of Nvidia’s GeForce 8800 GTX and Ultra was meant to rule in the $349 category but actually had to compete with the dangerous Radeon HD 4850 which was originally targeted into a cheaper market sector. As a result, many traits of this Goliath – such as its large dimensions, dual-slot cooler, expensive 12-layer PCB design and two power connectors – proved to be inappropriate, transforming into drawbacks. In contrast, the Radeon HD 4850, like the modest David, had a short 23cm PCB, a compact single-slot cooler and one power connector and delivered higher performance at a much lower price across a number of games.
Nvidia had to cut the price of the GeForce 9800 GTX even though the manufacturing cost of the reference card was quite high to start with, yet even this measure could not solve the performance problem. Even if the GeForce 9800 GTX cost the same money as the Radeon HD 4850, the latter looked preferable in a potential customer’s eyes. As Nvidia didn’t have anything else to pit against ATI’s RV770 processor, the only way was to squeeze some more performance out of the G92 chip. This attempt resulted in the GeForce 9800 GTX+ which differed from its predecessor with higher clock rates: 738/1836MHz against 675/1688MHz. This resembled what Nvidia had done with the GeForce 8800 Ultra earlier.
However, the manufacturing cost of the GeForce 9800 GTX+ would still have been too high if its reference design had been used. The reference PCB had been developed for quite a different price category and for the rather expensive 65nm version of the G92 chip. In contrast, the RV770 was originally manufactured on 55nm tech process and had a smaller size (260 against 330 sq. mm) and, accordingly, lower manufacturing cost than the G92 while being more complex. Trying to keep its positions in the mainstream sector, Nvidia hurried to transition its existing cores to 55mm tech process. Thus, the new card got a new heart – the G92b chip. The new revision doesn’t differ from the older G92 in anything save for the tech process, but the GeForce 9800 GTX+ has become more profitable to make.
Although Nvidia claims that the transition to the new tech process is complete in the mainstream sector (see this news story for details), you can still meet GeForce 9800 GTX+ with the older version of the core in shops. But in this review we’ll be dealing with the 55nm G92b processor. To perform this test we’ve got two GeForce 9800 GTX+ cards from Palit Microsystems, a well-known maker of inexpensive graphics cards who has also begun to produce prestige solutions under the Gainward brand. Besides gaming tests, we will also compare the power draw of the G92b as opposed to the G92.
Both versions of the Palit GeForce 9800 GTX+ come in medium-sized boxes.
The coloring of the boxes is black and green and silver and green, which agrees with Nvidia’s corporate colors, but the picture is somewhat odd. Instead of a girl, robot, mage or monster, the box shows an armor-clad frog - frobot. This image may be perceived as a hint at Froblin Demo, a program that demonstrates such advantages of the RV770 processor as a programmable tessellation unit and DirectX 10.1 support. These features are missing in Nvidia’s new G200 architecture, let alone G80/G92 chips.
The graphics card lies wrapped into an antistatic pack in a cardboard tray. Besides the card, the box contains the following:
The accessories are scanty just as you could expect considering the low price of the cards. On the other hand, there is everything you need to use all of the card’s features, including HDMI support. Software for playing HD video is the only missing thing here.
The 8-pin→6-pin power adapter may look odd to you because these connectors are mechanically compatible, and 8-pin graphics card connectors of modern PSUs often use an 8+2 design, but this adapter may indeed come in handy in some situations. The reference card from Nvidia is all hidden inside an exterior casing, and you can’t plug an 8-pin connector of your PSU into the card’s 6-pin connector because the 2 extra pins will press against the edge of the casing. That’s when you can use the adapter.
Overall, the accessories of the Palit GeForce 9800 GTX+ are up to the card’s market positioning. After all, G92-based solutions are now far from being flagship products as they were at the time the GeForce 9800 GTX was announced. The lack of a HD video player is a disappointment, but this lowers the price of the product, especially as Palit has never been an elite brand. The company is currently marketing its premium products under the Gainward brand. Thus, the accessories match the pricing of these cards.
Now let’s check out their design. One of these cards uses a non-reference PCB!
We won’t describe that Palit GeForce 9800 GTX+ which copies the reference card and cooler. You can just refer to our earlier reviews, for example to the review of the Gainward Bliss 9800 GTX 512MB. We’d better take a closer look at the other version which features a unique PCB design developed by Palit engineers.
First of all, this version lacks the main drawback of the reference GeForce 9800 GTX/GTX+. We mean its excessive length. The PCB from Palit is as long as the PCB of the Radeon HD 4850 or GeForce 9800 GT, i.e. 23 centimeters, making it easier to install the card into small system cases.
The power section still features an advanced four-phase GPU voltage regulator based on an NCP5388 controller from ON Semiconductor. Each phase has three power transistors, but there is a seat for a fourth transistor that may reinforce the power circuit if necessary. The PCB still has two power connectors, both of the PCIe 1.0 variety with a load capacity of 75W. These connectors face the user, like on the reference 9800 GTX/GTX+, but the PCB is now smaller and there is no real need for this orientation of the connectors. It would be handier to plug the cables in if the connectors were placed at the PCB’s butt-end. A dual-channel voltage regulator based on a Richtek RT9259A controller is responsible for the memory chips.
Palit has changed the positioning of the memory chips from semicircular around the GPU to the ordinary L-shaped layout to simplify the wiring (to make the PCB cheaper) and to save some space (to make the PCB shorter). The card is equipped with Samsung K4J52324QE-BJ08 memory.
These GDDR3 chips have a capacity of 512Mb (16Mb x 32), a voltage of 1.9V and a rated frequency of 1200 (2400) MHz, but their real frequency is 1100 (2200) MHz to comply with the official GeForce 9800 GTX+ specs. The Palit GeForce 9800 GTX+ that uses the reference design has the same memory frequency, too.
Palit’s cards both use the 55nm version of the G92 core also known as G92b.
The photo shows that the die has become much smaller. The chip is now revision B1 rather than the 65nm version’s A2 revision. This sample of the chip was manufactured on the 26th week of 2008, i.e. between the 22nd and 28th of June. The GPU of the other Palit card has the same marking but differs in its production date.
The main domain frequency is slightly overclocked (from 738MHz to 745MHz) on the reference-design card. The shader domain frequency is the same at 1836MHz, so this preoverclocking cannot make a big difference. The two cards have the same GPU configuration with 128 unified execution modules, 32 (64) texture processors and 16 render back-ends.
Each card is equipped with two dual-link DVI ports and a 7-pin mini-DIN connector. Additionally, each has a small 2-pin plug for connecting to the sound card’s onboard output to translate S/PDIF sound into HDMI. There are also two MIO connectors for building SLI configuration out of three such cards.
The unique PCB design called for a unique cooler. Well, the cooler is actually quite familiar to us.
In fact, it is a version of the reference cooler of the GeForce 7900 GTX card except that the heatsink consisting of thin aluminum plates is now smaller and doesn’t press against the card’s mounting bracket. Two heat pipes connect the heatsink to the copper base and ensure uniform distribution of heat. An axial fan is installed on the heatsink to blow downward as well as sideways. The developer must have wanted that some of the hot air were exhausted out of the system case through the slits in the mounting bracket. However, the air will largely remain inside because the heatsink’s ribs do not reach to those slits and the heatsink lacks an air-directing casing.
The fan has a 4-pin connection with PWM-based speed regulation. The whole arrangement is secured on the PCB with four spring-loaded screws. The cooler sits firmly and the GPU die is additionally protected with the metallic frame on its package.
Some of the airflow is driven downward to cool the PCB and its components such as memory chips which are additionally equipped with two heat-spreading plates, one plate for each four chips. There are no cooling elements on the power transistors of the power circuit.
This cooler doesn’t seem to have obvious advantages over the reference one that has inherited all the best features of the reference cooler of the GeForce 8800 GTX and is deservedly considered one of the best reference coolers ever. We are going to make our final verdict about Palit’s cooler after we check out its noise and efficiency in the next section of the review. We’ll also measure the power draw of the 55nm version of the G92.
The Palit GeForce 9800 GTX+ cards both have a 55nm G92b core, so we can see the benefits of the thinner manufacturing process. We measured the power consumption of the graphics cards on the following testbed:
The 3D load was created by means of the first SM3.0/HDR test from 3DMark06 running in a loop at 1600x1200 with 4x FSAA and 16x AF. The Peak 2D mode was emulated by means of the 2D Transparent Windows test from PCMark05. This test is important as it simulates the user’s working with application windows whereas Windows Vista’s Aero interface uses 3D features. Some enthusiasts criticize this interface for loading the graphics card even in 2D mode but there is in fact no difference when DWM is turned off.
Click to enlarge
So, we can see a great improvement over the 65nm version of GeForce 9800 GTX+. The difference is almost 34 watts. The new tech process proves to be advantageous for the end-user. Note that the card with Palit’s own design consumes less power than the card with Nvidia’s reference design. The difference isn’t big, though.
The card with the nonstandard cooler proved to be better than the card with the reference cooler in terms of temperatures. Its GPU temperature was 51°C in idle mode and 58°C under load. The reference cooler kept the GPU temperature at 55°C and 70°C, respectively. The GPU is beyond any danger anyway as it can work even at 90°C without any harm, yet the nonstandard cooler is superior, obviously due to its more advanced heatsink. But what cooler is the quieter of the two?
We measured the cards’ noise with a digital sound-level meter Velleman DVM1326 using A-curve weighing. The level of noise at a distance of 1 meter from the working testbed with a passively cooled graphics card inside was 43dBA. We’ve got the following results:
So, both cards are quite comfortable in terms of noisiness. They are not exactly silent, yet you can hardly hear them because there are a lot of other noise sources in a gaming system such as hard drives, PSU, CPU cooler, and system fans. It is only in a completely silent computer that you will be able to hear the hiss of the fan of the Palit card. This fan is working at its minimum speed almost always because the 55nm G92 generates much less heat than the 65nm version.
We also tried to overclock the cards. The reference-design card managed to speed up from its default 738/1836MHz GPU and 1100 (2200) MHz memory frequencies to 792/1944MHz and 1250 (2500) MHz, respectively. The card with Palit’s own PCB design was stable at 770/1890MHz GPU and 1200 (2400) MHz memory frequencies. Thus, the simplified PCB design and nonstandard power circuit worsened the card’s overclockability despite the more efficient cooling. We benchmarked the reference-design card at the overclocked frequencies. As you can guess, the overclocked card with the unique design would be slower than it, but faster than at the default frequencies.
For our performance tests of Palit GeForce 9800GTX+ graphics cards we used the following testbed:
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 by both: AMD/ATI and Nvidia. Also, to ensure maximum image quality, we enabled transparent texture filtering. As a result, our ATI and Nvidia driver settings looked as follows:
For our tests we used the following games and synthetic benchmarks:
First-Person 3D Shooters
Third-Person 3D Shooters
We selected the highest possible level of detail in each game using standard tools provided by the game itself from the gaming menu. The games configuration files weren’t modified in any way, because the user doesn’t have to know how to do it. The only exception was Enemy Territory: Quake Wars game where we disabled the built-in fps rate limitation locked at 30fps. Games supporting DirectX 10 were tested in this particular mode.
Besides both Palit GeForce 9800GTX+ cards we have also included the following graphics accelerators to participate in our test session:
We used the most widely spread resolutions: 1280x1024, 1680x1050 and 1920x1200. Everywhere, where it was possible without disabling the HDR/Shader Model 3.0/Shader Model 4.0, we ran the tests with enabled MSAA 4x antialiasing and anisotropic filtering 16x. These were our default settings in all tests except 3DMark. We enabled them from the game’s menu. If this was not possible, we forced them using the appropriate driver settings of ATI Catalyst and Nvidia ForceWare drivers. As we have already said, we didn’t modify the games configurations files.
Performance was measured with the games’ own tools and the original demos were recorded if possible. Otherwise, the performance was measured manually with Fraps utility version 2.9.6. In the latter case we ran the test several time and took the average number for the performance charts. We measured not only the average speed, but also the minimum speed of the cards where possible.
As we could expect, the 7MHz higher frequency of the main domain of the original Palit card provides but a negligible performance gain that varies from 3% at 1280x1024 to less than 1% at 1920x1200. That’s normal considering that the shader domain and memory frequencies are the same. Overclocking increases the average frame rate by 8 to 14%. This is enough for the Palit card to overtake the Radeon HD 4850 but the latter is fast enough to allow playing at the same resolutions with the same comfort.
The overclocked card from Palit delivers the same average speed as the Radeon HD 4850 and a higher bottom speed. That’s good for a product with an old architecture, especially considering the demands of the Call of Juarez engine. Comparing the two Palit cards between each other, the unique version is no more than 1 fps ahead of the reference version.
Overclocking is not very rewarding here, improving performance by 7% at best. This is not enough to compete with the Radeon HD 4850, let alone Radeon 4870 with its super-fast GDDR5 memory. On the other hand, you can play at any popular resolution, including 1920x1200, at the overclocked as well as default frequencies.
The overclocking gain is small in Crysis, too. However, the GeForce 9800 GTX+ cards feel better than the Radeon HD 4850, at least at low resolutions. At the high resolutions the overall level of performance is too low to play the game on these cards at the highest graphics quality settings.
The frame rate is fixed at 30fps in this game as this is the rate at which the physical model is being updated at the server. Thus, this 30fps speed is the required minimum for playing the game.
Our overclocking adds 3-9% to the average speed, making the Palit card competitive to the Radeon HD 4870. This has no practical value, however, because of the frame rate limit. Both cards ensure a double reserve of speed even at 1920x1200.
The system’s CPU is obviously the limiting factor at 1280x1024. At 1680x1050 the tested variants of GeForce 9800 GTX+ differ by less than 2%. It is only at a resolution of 1920x1200 pixels that the overclocking gain amounts to 4% and more. The GeForce 9800 GTX+ looks better than the Radeon HD 4850 here but you can barely feel this during practical play.
The Radeon HD 4870 enjoys a serious advantage, indicating that this game appreciates high graphics memory bandwidth. The other cards are inferior in this respect and keep close to each other, the Radeon HD 4850 being somewhat slower at low resolutions and the GeForce 9800 GTX+, at high resolutions. The maximum benefit from our air-based overclocking amounts to 10%, but it can’t make the game playable with the advanced special effects turned on.
Dead Space has modest system requirements: the overclocked Palit and the Radeon HD 4870 stand out among the others, showing similar speeds. The overclocking gain is a mere 1-4% here.
Our overclocking is more rewarding in Tomb Raider, amounting to 4-13% of average speed. This addition is likely to go unnoticed by the gamer, however, because all the cards, except the GeForce 9800 GT at 1920x1200, are very fast here.
ATI’s solutions are superior in this test. The Palit card overclocked to 792/1944MHz core and 2500MHz memory can’t do anything about that, especially as this overclocking increases the average frame rate by only 6-10%. However, the bottom speed of all the versions of GeForce 9800 GTX+ is high enough to play the game comfortably at 1680x1050.
Based on Unreal Engine 3, this game does not support FSAA directly but you can force this feature from the driver’s Control Panel. And you’ll indeed get rid of those jaggies. However, this results in a performance hit at high resolutions although the GeForce 9800 GTX+ look preferable to the Radeon HD 4850 at low display modes.
The G92 architecture can’t do anything in this game to oppose the Radeon HD 4850. The gap is 19-20% irrespective of the resolution whereas the maximum growth in average speed at the overclocked frequencies is no higher than 6%. Moreover, there is a gap in terms of bottom speed which may affect your playing comfort in heaviest scenes of the game.
The overclocking effect can only be observed at 1920x1200 in open scenes. Otherwise, the Radeon HD 4850 and the three versions of GeForce 9800 GTX+ are equal to each other. The game is unable to benchmark modern graphics cards adequately, so we are going to replace it with Fallout 3 which runs on a similar but improved engine.
Palit’s cards are in the lead here. The Radeon HD 4870 can only get close to them at 1920x1200. However, this Radeon is more expensive while the overclocked Palit is still a little ahead of it.
Like some other projects from EA, Spore has an integrated frame rate limiter, so you should compare the bottom speeds of the cards in the first place.
The Radeon HD 4800 series cards win this test as they reach the frame rate limit at all resolutions. The GeForce 9800 GTX+ have lower bottom speeds even at 1680x1050. At a resolution of 1920x1200 the GeForce 9800 GTX+ may slow down to below comfortable level. Overclocking helps improve the average speed somewhat but the bottom speed is still too low. Thus, overclocking makes no practical difference in this game.
The speeds are overall low in this test, so the resolution of 1280x1024 is the most interesting one. Here, the GeForce 9800 GTX+ doesn’t accelerate much at the overclocked frequencies but delivers a good bottom speed, allowing you to play comfortably. The Radeon HD 4800 cards have lower speeds and you can feel the difference in practice.
Even though the good old 3DMark06 isn’t quite adequate today, it is quite correct saying that the G92-based solutions are inferior to the RV770-based ones. The Palit cards are superior in the SM2.0 tests but it is only at the overclocked frequencies that they can get close to the Radeon HD 4850.
The results of the individual SM2.0 tests agree with what we’ve said above but it is only in the first test, which loads the texture processor subsystem, that the Palit cards are much faster.
The SM3.0/HDR tests produce predictable results, too. The overclocked Palit GeForce 9800 GTX+ is close to the Radeon HD 4850 in the second test only. No wonder the ATI solution is better in overall scores even though 3DMark06 can make use of but a small portion of the potential of the RV770 architecture.
We minimize the CPU’s influence by using the Extreme profile (1920x1200, 4x FSAA and anisotropic filtering).
Both versions of Palit GeForce 9800 GTX+ are ahead of the Radeon HD 4850, but far inferior to the Radeon HD 4870. The gap cannot be bridged even through overclocking.
Although the first test is easier for ATI’s architecture, the Radeon HD 4850 is slower than its opponents in it. The second test puts a load on the texture processors, but the Radeon HD 4850 is as fast as the non-overclocked Palit cards there.
The overclocked Palit is in between the Radeon HD 4850 and HD 4870, which is a good result for an obsolete architecture in a DirectX 10-oriented test.
Our tests of two Palit GeForce 9800 GTX+ graphics cards show that the G92 architecture will not leave the battlefield without a fight. However, it cannot achieve an overwhelming advantage over the Radeon HD 4850 as our summary diagrams show.
At a resolution of 1280x1024 pixels the G92 enjoys a serious advantage, over 15%, in six out of the 17 tests. In two tests the Palit cards were far slower than their opponent. The overclocking gain varied from 1 to 10%, the average gain being 4.3%. Thus, overclocking such cards is an unrewarding business unless you want to set a new record in 3DMark.
At a resolution of 1680x1050 pixels, Nvidia is superior in five tests but the average overclocking gain is now about 6%.
It is the 1920x1200 results that are indicative of the decline of the G92 era. The Palit GeForce 9800 GTX+ are more than 15% faster than the Radeon HD 4850 in two tests only, namely in Half-Life 2: Episode Two and Tomb Raider: Legend, but lose in four tests (Call of Duty 4, The Witcher, Spore and World in Conflict). The average overclocking gain is 8.5%, which doesn’t give any edge to the G92-based solutions.
The two versions of Palit GeForce 9800 GTX+ are good products if you don’t mind their outdated architecture. The version with a nonstandard PCB design is preferable. It is smaller and has better cooling and requires somewhat less power. Moreover, this version may prove to be cheaper due to its simpler design. Its lower overclockability is its only drawback, but you can’t expect substantial benefits from overclocking such cards anyway unless you use some extreme overclocking methods. Such overclocking just adds a few frames per second in Fraps and other benchmarks but cannot be felt during actual play.
So, if you want to buy a GeForce 9800 GTX+, the Palit GeForce 9800 GTX+ with nonstandard PCB can be a good choice. The reference version of GeForce 9800 GTX+ should only be bought for serious overclocking.