by Alexey Stepin , Yaroslav Lyssenko,
07/17/2008 | 10:18 PM
In our reviews we often overclock graphics cards in a “rational” way. In other words, we overclock without special tools or methods such as replacing the card’s native cooler with a liquid cooling system or modifying its power circuitry. One property of rational overclocking is that it can be continuously used in everyday operation. Although it often brings but small performance benefits, many gamers are interested in and practice this kind of overclocking.
Besides rational, there is extreme overclocking, which is quite a different thing. Such overclocking relies on extreme methods, up to cooling the overclocked card with liquid nitrogen or increasing the GPU and memory voltages in order to increase the overclockability of these components. As you know, nearly every GPU and memory chip can work at a clock rate far above the one set by the developer if you increase its voltage. Extreme overclocking leads to a sudden increase in the chip’s heat dissipation and a serious reduction of its service life. This explains why it is but seldom used by practical gamers. It is just a kind of sport in its own right in which record-breaking showings are more important than stable and problem-free operation. The excitement is high and so are the bets. Nearly every overclocker can tell you a story about how his graphics card died in the process of achieving a higher score in 3DMark. Some may even show you a whole collection of dead hardware.
It is flagship solutions that are usually overclocked in extreme ways because they give you the highest chance to set a new record you can boast about before other overclockers. We set ourselves a different goal in this review, however. We wanted to squeeze everything out of an ATI Radeon HD 3870 to check out the potential of the RV670 chip and see if it could at least theoretically be as fast as Nvidia’s G92-based solutions. It is an interesting question because there is currently a gap in the Radeon HD 3000 line-up between the relatively inexpensive single-chip Radeon HD 3870 and the dual-chip Radeon HD 3870 X2 whereas there is no such gap in Nvidia’s line-up.
True overclockers often limit themselves to fixing the record in 3DMark. We are not in for any records, yet we hope to achieve a GPU frequency of 1GHz with air cooling and test our Radeon HD 3870 in this mode in our traditional benchmarks. This will provide a complete picture of performance that may be interesting for ordinary gamers, not only overclockers.
Our experiments should not be understood as an invitation to overclocking or modifying any hardware. Everyone who wants to repeat them will do so at his own risk.
It has been a while since we last experimented with extreme overclocking. Many things have changed. Power circuitry of graphics cards has become far more complicated and now uses smart multi-channel programmable PWM-controllers. Among other things these controllers help implement power saving technologies by reducing the frequency and voltage of the core when the graphics card is idle.
A modern GPU has a few zones clocked at different frequencies. The GeForce 8/9 series is an obvious example but this also refers to modern Radeon HD chips which come with only one frequency declared but actually have a number of zones clocked at different frequencies. Moreover, the frequencies can change dynamically within the framework of PowerPlay technology.
You have to account for these peculiarities when you want to overclock a modern graphics card. For example, the original BIOS of the Radeon HD 3870 doesn’t contain divisors for frequencies above 862MHz. Therefore, your attempt to set the core frequency above this value will hang the system irrespective of the GPU voltage. Fortunately, this limitation is not fundamental and can be easily bypassed by means of a modified BIOS you can download here. You can use the ATIFlash tool (we used version 3.59) to reflash the BIOS. Then you can try to volt-mod your Radeon HD 3870.
The GPU power circuit of this card is based on the uPI uP6201 controller. It is difficult to find any documentation on this chip and we couldn’t find it, either. However, the volt-mod is most simple: you should connect a trimming resistor between the uP6201’s Pin 13 and the ground. We used a multiturn 100kOhm resistor. The points of connection are marked with blue in the photo below:
You can use one of the Vgpu control points marked in red to keep track of the GPU voltage. By default, the GPU voltage should be within 1.3V in 3D mode.
It is just as simple to volt-mod the memory circuit: another 100kOhm trimming resistor is connected to Pins 3 and 7 of the uPI uP6101BSA chip. These points are marked with green in the photo:
You can monitor the memory voltage in the point marked as Vgddr control. The memory voltage is 1.89V by default.
After the BIOS update to solve the 862MHz problem, the described hardware modification enabled us to adjust the GPU and memory voltages in a wide range. The cooler’s fan was set at its maximum speed by means of RivaTuner to keep the GPU temperature within reasonable limits. Our attempt to overclock the GPU produced the following results:
Obviously, we’ve got not the best sample of the ATI Radeon HD 3870. We only achieved a frequency of 1GHz at a voltage of 1.71V although other cards were reported to be stable at that frequency with 1.55-1.6V voltage. We could not lift the frequency to 1.1GHz at 1.75-1.77V and we didn’t increase the voltage further because 1.7V was already dangerous for the 55nm chip with a default voltage of 1.33V. The temperature showings were indicative of that: the fan working at the max speed, the GPU was as hot as 69-72°C when idle and over 90°C hot under load. We wouldn’t recommend you to repeat our experiment because there is a risk of damaging the card even if you provide enough cooling. Well, ordinary users do not use this kind of overclocking whereas true enthusiasts can’t be stopped by any warnings.
We managed to overclock the memory chips to 1242 (2484) MHz at a voltage of 2.05V. This increased the memory bandwidth from 72 to 79.5GBps. That’s not much. Anyway, that’s not the memory bandwidth that is going to be the main factor in games.
The card successfully passed a couple of test cycles in 3DMark06 and 3DMark Vantage at 999/2484MHz but would hang up in games after a while. We achieved perfect stability at 972MHz GPU and 1224 (2448) MHz memory frequencies. The card passed all of our tests successfully at these clock rates.
It is important to know the power consumption of the card you overclock, so we checked out how much power a Radeon HD 3870 overclocked to a GPU frequency of 1GHz needed. We performed our measurements on a special testbed configured like follows:
The 3D load was created by 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 the 2D Transparent Windows test from PCMark05. Here are the results:
Click to enlarge
PowerPlay technology is disabled at extreme overclocking and the core frequency is not varied depending on operation mode. This explains the high numbers in the 2D and Peak 2D modes. The most interesting number is the one referring to the 3D mode: the overclocked card consumed 10W more than the ATI Radeon HD 3870 X2 and the load on both 12V power lines was considerably higher than the allowable limit of 75W. This is the tradeoff of extreme overclocking with GPU volt-modding.
And we doubt the overclocked Radeon HD 3870 is going to be comparable to the ATI Radeon HD 3870 X2 in terms of performance. Increasing the clock rate is a dead-end, actually. The future belongs to multi-core solutions, homogeneous or heterogeneous.
For our performance tests of overclocked ATI Radeon HD 3870 we put together 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 - Adaptive Anti-Aliasing/Multi-sampling for ATI Catalyst and Antialiasing – Transparency: Multisampling for Nvidia ForceWare. As a result, our ATI and Nvidia driver settings looked as follows:
ATI Catalyst:
Nvidia ForceWare:
For our tests we used the following games and synthetic benchmarks:
First-Person 3D Shooters
Third-Person 3D Shooters
RPG
Strategies
Synthetic Benchmarks
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. 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 overclocked ATI Radeon HD 3870 we have also included the following graphics accelerators to participate in our test session:
The tests were performed in the following resolutions: 1280x1024/960, 1600x1200 and 1920x1200. If the game didn’t support 16:10 display format, we set the last resolution to 1920x1440. We used “eye candy” mode everywhere, where it was possible without disabling the HDR/Shader Model 3.0/Shader Model 4.0. Namely, we ran the tests with enabled anisotropic filtering 16x as well as MSAA 4x antialiasing. 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.
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 2.9.1. We measured not only the average speed, but also the minimum speed of the cards where possible.
This game doesn’t support display resolutions of 16:10 format, so we use a resolution of 1920x1440 pixels (4:3 format) instead of 1920x1200 for it.
The extremely overclocked version of Radeon HD 3870 is far ahead of itself working at the default frequencies. For example, the average performance growth amounts to 20% at 1600x1200 whereas the minimum speed grows up by 35%.
There is a smaller performance growth at 1920x1200 but it may indeed be helpful for the gamer in some situations. On the other hand, Nvidia’s solutions are even faster.
BioShock doesn’t support FSAA when running in Windows Vista’s DirectX 10 environment. We benchmark graphics cards without FSAA in this game.
ATI’s cards feel at ease in BioShock without any overclocking. Running at a GPU frequency of 972MHz the Radeon HD 3870 nearly wins the test. It is only inferior to the dual-processor Radeon HD 3870 X2 and to the Nvidia GeForce 9800 GTX. Note that the minimum speed of the Radeon HD 3870 X2 is considerably higher at resolutions above 1600x1200.
Extreme overclocking isn’t very beneficial from a practical standpoint. It only improves the minimum speed at 1920x1200 which was below 30fps before the overclocking.
The overclocked Radeon HD 3870 is as fast as the GeForce 9800 GTX at resolutions above 1280x1024 (it has a lower minimum of speed, though). This is not enough to have really high results in the tech demo from Call of Juarez DX10, however. The Radeon HD 3870 X2 remains unrivalled.
The overclocking leads to a 15% increase in speed at 1280x1024. Unfortunately, the OpenGL driver from ATI’s Catalyst suite is still imperfect. The Radeon HD 3870 is slower than its opponents even with its GPU overclocked by 30%.
This game is tested at the High level of detail, excepting the Shaders option which is set at Very High. This way we try to achieve a compromise between image quality and speed.
The overclocking increases the performance of the Radeon HD 3870 by about 20% but the game is still far from playable even at the relaxed graphics quality settings. The overclocked Radeon almost beats the GeForce 9600 GT and gets close to the GeForce 8800 GT even though the absolute performance growth is very low.
It looks like you have to wait for the next generation of graphics cards from ATI and Nvidia to enjoy Crysis fully. We have some apprehensions that the game will run fast enough even then, though.
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.
Working at its limit, the Radeon HD 3870 takes third place at 1280x1024 and second at the higher resolutions where it is only slower than the GeForce 9800 GTX.
The speed is far above the required limit but it is really risky to use your graphics card in such a harsh operating mode. We can see that the overclocked Radeon HD 3870 is not far slower than the more expensive competing solutions, though.
The effect from the increased frequency is quite obvious but not strong enough for the Radeon HD 3870 to catch up with the GeForce 8800 GT 512MB or GeForce 9600 GT. The gap is small at 1920x1200 and the speed is good enough for comfortable play, but the noise from the fan and the high temperature of the core won’t let you relax for real. It’s no fun playing when the graphics card is roaring like hell and can fail any moment because of the core voltage at 1.7V.
The game doesn’t support FSAA when you enable the dynamic lighting model, but loses much of its visual appeal with the static model. This is the reason why we benchmarked the cards in S.T.A.L.K.E.R. using anisotropic filtering only.
The overclocking is no good for S.T.A.L.K.E.R. because this game is optimized for GeForce cards. The average speed gain of the overclocked Radeon is only 14-15%. ATI’s solutions can only do well in this game when they have more GPUs: the Radeon HD 3870 X2 is quite competitive against top-end single-chip cards from Nvidia.
Our overclocking helped the Radeon HD 3870 reach the level of the GeForce 8800 GT 512MB in terms of average performance. The minimum speed didn’t grow up, though. The overall performance level is still extremely low, too.
Unfortunately, the overclocked Radeon HD 3870 is even unable to overtake the GeForce 9600 GT. The gap is 5-6% at 1920x1200. The game is quite playable at the highest resolution if you’ve got a Radeon HD 3870 that can work at such extremely high frequencies without overheating, failing, and producing unbearable noise.
Overclocked to a GPU frequency of 972MHz, the Radeon HD 3870 wins two out of the three tested resolutions (among the previous-gen single-chip solutions), which would be an excellent result if it were not for the low minimum speed that must be due to the faulty design of the Radeon HD’s texture processors.
We think TES IV looks best with enabled FP HDR and test it in this mode.
The modified Radeon HD 3870 is about as fast as the Radeon HD 3870 X2 in terms of average performance but shows a higher minimum of speed.
The overclocked card is only second to the GeForce 9800 GTX (the gap is a mere 2-3%) at the first two resolutions in open scenes. At 1920x1200 it is no worse than the Radeon HD 3870 X2 and GeForce 8800 GT 512MB.
The new add-on to Company of Heroes is tested in DirectX 10 mode only since it provides the highest quality of the visuals.
The average frame rate of the overclocked Radeon HD 3870 is but slightly higher than that of the non-overclocked sample. More considerable is the increase in minimum speed at 1280x1024 where the gameplay becomes much smoother. The card can’t make it to the results of the Radeon HD 3870 X2 but is only 5% slower than the GeForce 8800 GT 512MB at 1600x1200 and higher resolutions. That’s a good result, but it comes at a price.
The game having a frame rate limiter, you should consider the minimum speed of the cards in the first place.
This game is counted among our benchmarks yet it cannot say anything about the overclocked RV670: the ATI Radeon HD 3870 easily hits the speed limit in every display mode even at its default frequencies.
The extreme overclocking helps the Radeon HD 3870 catch up with the GeForce 8800 GT 512MB and GeForce 9600 GT. The overclocked memory coupled with the more efficient ring-bus memory controller says its word at the higher resolutions: the gap from the GeForce 9800 GTX is just 1fps at 1600x1200. At a resolution of 1920x1200 the overclocked Radeon HD 3870 even takes first place, beating every opponent and showing the best minimum speed.
3DMark is an overclocker’s favorite tool and boasting with a new record-breaking score is something what every overclocker loves. We are not into any records and our results are modest. However, the modified Radeon HD 3870 takes third place among the tested cards. It might be even second if it were not for the dual-processor Radeon HD 3870 X2 that is obviously unrivalled. Let’s check out the individual tests now.
High frequencies are good but the faulty design of the RV670’s texture processors is a negative factor. As a result, the Radeon HD 3870 X2 enjoys but a small lead while the overclocked Radeon HD 3870 can’t match the GeForce 8800 GT 512MB.
It’s different in the SM3.0/HDR tests that require all the computing resources the GPU has. The overclocked Radeon scores only 232 points less than the GeForce 9800 GTX while the Radeon HD 3870 X2 leaves Nvidia’s solutions far behind, winning by nearly 2000 points.
As we said above, the RV670-based solutions can’t show their best in the SM2.0 tests due to the faulty design of their texture-mapping units. The overclocked Radeon HD 3870 is barely faster than the reference card and is inferior even to the GeForce 9600 GT. It overtakes the latter in the second test, but that’s all it can do.
The results of the SM3.0/HDR tests are much brighter for ATI. In both tests the modified and overclocked Radeon is as fast as the GeForce 8800 GT 512MB and but slightly slower than the GeForce 9800 GTX. This would be good if the card could run in such a harsh mode continuously. Alas, these results are of purely theoretical interest.
To minimize the influence of the CPU on the 3DMark Vantage tests we select the Extreme profile that uses 1920x1200 resolution, 4x FSAA, and anisotropic filtering.
Our attempt to overclock the ATI Radeon HD 3870 to an extremely high clock rate proved that this GPU should not be expected to increase its performance dramatically if its frequency is increased by 30%.
Click to enlarge
Click to enlarge
The diagrams show that the performance benefits are small. The overclocked Radeon HD 3870, working at its limit, often fails to reach the level of the GeForce 8800 GT 512MB. For example, the speed of Crysis in DirectX 10 mode and at the highest quality settings increases by 20% but that only means that the frame rate grows from 15 to 18fps. The more advanced GPUs delivered higher performance anyway. The same is true for other DirectX 10 applications such as Call of Juarez, Lost Planet and World in Conflict. On the other hand, the extreme overclocking helped achieve acceptable performance at least at 1280x1024 in such games as Call of Duty 4 and S.T.A.L.K.E.R. and improved the average frame rate in some other games.
You should be aware that these performance benefits are achieved by means of heavy overclocking that can’t be harmless whereas installing a more advanced cooling solution would require an additional investment.
The dual-processor Radeon HD 3870 X2, consuming about the same amount of power, beats the GeForce 9800 GTX where AFR support is implemented properly. The more affordable GeForce 8800 GTS beats the overclocked Radeon HD 3870.
The lack of linear performance growth at overclocking is due to many factors such as architectural features of the GPU (the lack of texture filter or rasterization units), driver flaws, insufficient application-specific optimizations, etc. This cannot justify the result, though.
Although TSMC’s 55nm tech process developed in collaboration with ATI/AMD has a good frequency potential and RV670-based cards can work at a frequency of 900MHz quite easily, there is no sense in releasing Radeon HD 3870 with increased frequencies. There can be a higher performance growth if each of the chip’s 26 frequencies is increased, perhaps not proportionally, but it is not required: the new ATI RV770 chip is already here.
Summing it up, extreme overclocking is only really interesting for overclocking communities and for setting records. It is a kind of sport, actually. This sport may be exciting and spectacular, especially if practiced with exotic tools like liquid nitrogen, but has little practical value.