by Alexey Stepin , Yaroslav Lyssenko
07/05/2010 | 02:42 PM
Our previous review of the two most advanced multi-GPU tandems, Nvidia GeForce GTX 480 SLI and Radeon HD 5870 CrossFire, made it clear that Nvidia’s solution is superior to its opponent in most applications. This is no wonder as the GF100 indeed has higher potential than the RV870, and GeForce GTX 480 cards are actually more expensive than Radeon HD 5870s. However, modern multi-GPU technologies are not limited to dual-processor configurations but can be used to combine three or even four graphics cores into a single graphics subsystem. AMD’s technology is more flexible at that. While Nvidia SLI requires two or more identical cards, CrossFire technology allows using different cards, e.g. a Radeon HD 5870 together with a Radeon HD 5850. So, what if we combine a Radeon HD 5870 with a Radeon HD 5970? Such a subsystem would have tremendous potential.
From a theoretical standpoint, Nvidia has nothing to pit against the combined power of 4800 ALUs grouped into 960 superscalar stream processors. In the worst case for AMD, when only one out of each five ALUs has any work to do due to poorly optimized code, AMD’s solution is equal or slightly inferior to Nvidia’s because the GF100 has a higher frequency of the computing subunits. In every other case, the 3-way CrossFire subsystem is far superior to the GeForce GTX 480 SLI tandem, especially in terms of texture-mapping performance. Yes, its practical efficiency is going to be lower than the theoretical expectations due to the synchronization of the Radeon HD 5970 and Radeon HD 5870 working at different clock rates, but even then the performance should be high enough to beat the opponent. That’s what we are going to check out in our tests today.
The foremost drawback of the proposed 3-way CrossFire subsystem is its high price. The Radeon HD 5970 still sells for no less than $650, so the whole configuration can hardly get under $1000 whereas the GeForce GTX 480 tandem easily does that. Anyway, we guess this competition is going to be interesting to fans of both GPU developers.
Besides, we will tell you about a unique version of Radeon HD 5870 offered by Gigabyte. The GV-R587SO-1GD (or Gigabyte Radeon HD 5870 Super Overclock) is going to be interesting for everyone who wants a fast and reliable Radeon HD 5870. This card will serve as the junior part of our 3-way CrossFire subsystem.
Gigabyte positions this Radeon HD 5870 as its fastest single-processor RV870-based solution. Its packaging is a matte black box designed in a very austere way.
The text on the box tells you that this is a Super Overclock product. You can read its name and see a Gigabyte logo in silvery letters. Here you can also learn that the card belongs to the Ultra Durable VGA+ series which features top-quality components and that its graphics core has passed a GPU Gauntlet test that guarantees stable operation at the increased clock rate. The various unique technologies employed in the Gigabyte Radeon HD 5870 SO are described on the back of the box. Most of them are useful indeed.
There is another nice-looking box, made from glossy cardboard and embellished with a Super Overclock logo, inside the external one. Inside it there is a foam-rubber tray with the graphics card. This packaging keeps the product safe from transportation dangers. The accessories to the card can be found nearby:
Not many things in there for such a highly positioned product but this Radeon HD 5870 is itself unique and exciting. Computer enthusiasts but seldom need such free bonuses as game discs or mouse pads. Everything you need to use the card is included into the box, particularly the Gigabyte OC Guru software on the driver disc. This utility is important for the Super Overclock series as it allows to flexibly control parameters of such graphics cards and keep track of temperatures, voltage regulator phases, power consumption, GPU load, etc.
For all its flexibility, Gigabyte OC Guru is not without downsides. We do not like its interface with the numerical indicators stylized as mechanical counters – these are just hard to read. MSI Afterburner with its RivaTuner-like graphs is much more user-friendly but cannot access the advanced features of the Gigabyte Radeon HD 5870 SO, particularly the GPU voltage adjustment option. The user manual provides a detailed description of the functionality and features of Gigabyte OC Guru, so even inexperienced users shouldn’t have any problems using it.
Thus, we don’t have any serious complaints about the packaging and accessories of the Gigabyte Radeon HD 5870 SO. This product will be conspicuous on a shop shelf. The card itself is protected against any hazards that may occur during transportation and storage and comes with everything you need to use all of its features. Now, let’s take it out of the box and have a closer look at its design.
This Radeon HD 5870 does not have anything in common with the reference design. It features an original PCB with thicker metallization layers (the so-called “2oz Copper PCB” technology) and colored differently than the reference PCB.
Gigabyte did not use a cooler that would exhaust the hot air out of the system case. Instead, there is a cooling system with two large fans similar to what we saw in our Palit GeForce GTX 470 review. Such coolers are generally rather effective but not if there are other add-on cards nearby which block one of the fans. As usual, the most interesting things are hidden beneath the heatsink:
The PCB space is utilized more efficiently than on the reference card: there is almost no free room at the back of the Gigabyte Radeon HD 5870 SO. The GPU power circuit is considerably reinforced using six rather than four phases.
Overclockers will appreciate the opportunity to monitor the card’s voltages at the control points located in the bottom right corner of the PCB. We can also note that the card uses high-quality chokes with ferrite cores.
An ADP4100 chip from On Semiconductor is responsible for managing this complex power system. The same chip is used on GeForce GTX 275 cards. Software-based GPU management is implemented via the Gigabyte OC Guru utility. You also see NEC/Tokin 0E907 film capacitors with a capacitance of 900 µF that look like memory chips, and six SMD LEDs indicating the operation of the GPU voltage regulator. When the latter is working normally, the indicators are green.
The above-noted film capacitors are also employed in the memory voltage regulator. Some of its components are in the back part of the PCB while others are near the CrossFire connectors. Oddly enough, both power connectors the graphics card has are of the 6-pin variety whereas the original design from AMD provides for one 8-pin PCIe 2.0 connector with double load capacity. Well, as we found out, single-processor RV870-based cards have no real need for such 8-pin connectors.
There are eight GDDR5 memory chips from Hynix on the PCB, marked as H5GQ1H24AFR-T2C. The same chips with a capacity of 1 Gb (32 Mb x 32) each are installed on Radeon HD 5970 cards. The T2C suffix means their ability to be clocked at 1250 (5000) MHz and this is indeed the memory frequency of the Gigabyte Radeon HD 5870 SO, delivering a peak memory bandwidth of 160 GBps. This is higher than the reference card’s 153.6 GBps but lower than the GeForce GTX 480’s 177.4 GBps. The difference isn’t huge, however, and even negligible considering the weaker texture-mapping subsystem of the GF100 processor.
The graphics card automatically lowers its memory frequency to 960 (3840) MHz in power-saving mode. Compared with the reference card from AMD, which lowers its memory frequency to 300 (1200) MHz, this is not very economical, but graphics memory does not contribute much to a graphics card’s overall power draw.
Like any other Radeon HD 5870, this version has the following GPU configuration: 1600 ALUs grouped into 20 SIMD cores, 80 texture-mapping processors and 32 raster back-ends. The RV870 chip of our sample of the card was made on the 7th week of 2010. It is marked just like any other such core, but Gigabyte culls the best chips for its Super Overclock series, capable of working at increased frequencies. The testing procedure is rather gloomily referred to as GPU Gauntlet. Here, the core clock rate is increased from 850 to 950 MHz and its voltage is raised from 1.15 to 1.18 volts. You can use the Gigabyte OC Guru tool to raise the voltage even higher, but this didn’t help us in our experiments. The card is less economical than the reference one in power-saving mode, lowering its GPU frequency to 680 MHz instead of 157 MHz. Thus, the Gigabyte Radeon HD 5870 SO is going to be less economical in desktop applications and during HD video playback.
Like the reference Radeon HD 5870, this card has two DVI-I connectors, one HDMI and one DisplayPort. And like the reference card too, it allows connecting three display devices simultaneously, but one of them must use the DisplayPort. Besides, it has a couple of CrossFire connectors for building a graphics subsystem with up to four graphics cards. For example, we ran our Gigabyte Radeon HD 5870 SO together with a Radeon HD 5970 without any problems.
The cooling system employed by Gigabyte does not look like the reference one but is similar to the cooler we saw on the Palit GeForce GTX 470. A couple of rather aggressive-looking 80mm Everflow T128010SH fans are blowing at the big flat aluminum heatsink connected with four heat pipes to the copper base.
Most of the hot air remains inside the system case. The heat-exchanger that contacts with the GPU die is ordinary enough and does not use the direct-touch technology. Standard dark-gray thermal grease serves as a thermal interface. The rest of the components lack any heatsinks and are cooled by the air flow. Despite the larger dimensions, the whole arrangement is secured on the PCB with only four spring-loaded screws. So, we wouldn’t recommend you to take this card by the cooler’s casing lest you should damage the GPU die.
The design of the cooler is quite familiar to us already and we know that it can cool the GPU effectively. However, it requires that the system case itself be properly ventilated. And it doesn’t work that effectively if there is a neighboring add-on card blocking one of the fans. If you install a Gigabyte Radeon HD 5870 SO into your system, you must make sure the expansion slot beneath it is empty or occupied by a small card. Otherwise, the left fan won’t be of much help, dropping the efficiency of the whole cooler. We’ll test that efficiency in the next section.
Since Gigabyte Radeon HD 5870 SO uses a nonstandard PCB and works at pre-overclocked frequencies, we measured its power consumption on our standard testbed:
The new testbed for measuring electric characteristics of graphics cards uses a card designed by one of our engineers, Oleg Artamonov, and described in his article called PC Power Consumption: How Many Watts Do We Need?. This device facilitates and automates the measurement process. As usual, we used the following benchmarks to load the graphics accelerators:
Except for the maximum load simulation with OCCT, we measured power consumption in each mode for 60 seconds. We limit the run time of OCCT: GPU to 10 seconds to avoid overloading the graphics card's power circuitry. Here is what we managed to obtain using this testing methodology:
The Gigabyte card is not as economical as the reference Radeon HD 5870 in the desktop and video playback mode due to its factory overclocking. The GPU clock rate is not lower than 680 MHz even in 2D applications, which is a peculiarity of this version of Radeon HD 5870. Besides, the card refused to run OCCT:GPU for more than a few seconds. It would hang up and produce various visual artifacts, although was perfectly stable in games. This may be some kind of protection against severe synthetic tests such as OCCT or FurMark.
The cards from AMD and Gigabyte differ somewhat in terms of load distribution among the different power lines as they have different power subsystems. The Gigabyte Radeon HD 5870 SO loads the +3.3V line less but has higher load on the first +12V connector.
It must be noted that the voltage regulator design implemented by Gigabyte engineers is more efficient than the reference one in 3D applications. Notwithstanding the factory overclocking, the card’s peak power draw is but slightly higher than that of the reference sample. The results of the multi-GPU subsystems are approximations calculated by adding up the power consumption of their constituents in 3D mode. It is clear that the three AMD RV870 processors are going to be more economical than the pair of Nvidia GF100 chips. This naturally comes out of the difference in the number of transistors in these chips.
The cooling system of the Gigabyte Radeon HD 5870 SO performs excellently, keeping the GPU as cool as 60°C under load. No reference cooler from AMD or Nvidia can match that result even at a lower GPU frequency. Most importantly, the described product is rather compact as opposed, for example, to the Zotac GeForce GTX 480 AMP! which has a dual-slot form-factor. You must only make sure that no expansion card blocks the cooler’s left fan and prevents it from delivering its full performance.
And what about noise? Well, the fans of the Gigabyte Radeon HD 5870 SO rotate at about 2800 RPM in 3D applications and are perfectly audible. Fortunately, the spectrum of this noise is rather agreeable to the ear and does not irritate. It is a soft hiss of the air. The problem is that the fans do not slow down in 2D applications. This is the tradeoff for the GPU temperature of 39°C in 2D mode. Well, we could hardly expect an enthusiast-targeted graphics card to be completely silent.
For all its positioning, the graphics card didn’t do well in our overclockability tests. At its default GPU voltage it would hang the system up even at a GPU frequency of 960 MHz. Yes, the Gigabyte OC Guru tool offers flexible options for controlling the GPU and memory voltages but we didn’t risk using them. Volt-modding, both software and hardware based, often leads to an untimely death of the graphics card. The Gigabyte Radeon HD 5870 SO is superior in frequencies to the reference card anyway and we already investigated the overclockability of the RV870 core in our earlier reviews. So, let’s focus on the subject of this one as we are going to benchmark a 3-way CrossFire configuration. Our Gigabyte Radeon HD 5870 SO works with a Radeon HD 5970 without any problems. Let’s see what performance they deliver.
We are going to investigate the gaming performance of our three-way CrossFire configuration using the following universal testbed:
We used the following versions of ATI Catalyst and Nvidia GeForce drivers:
The ATI Catalyst and Nvidia GeForce graphics card drivers were configured in the following way:
Below is the list of games and test applications we used during this test session:
First-Person 3D Shooters
Third-Person 3D Shooters
Semi-synthetic and 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, because the ordinary user doesn’t have to know how to do it. We updated our test modes and ran our tests in the following resolutions: 1600x900, 1920x1080 and 2560x1600. Unless stated otherwise, everywhere, where it was possible we added MSAA 4x antialiasing to the standard anisotropic filtering 16x. We enabled antialiasing from the game’s menu. If this was not possible, we forced them using the appropriate driver settings of ATI Catalyst and Nvidia GeForce drivers.
The triple-way CrossFire graphics configuration from AMD will be competing against the today’s most powerful Nvidia SLI tandems built of GeForce GTX 480 and GeForce GTX 470. For the purposes of comparison, we will also include the results for the less powerful CrossFire configurations and single-GPU flagship products from AMD and Nvidia.
Performance was measured with the games’ own tools and the original demos were recorded if possible. We measured not only the average speed, but also the minimum speed of the cards where possible. Otherwise, the performance was measured manually with Fraps utility version 3.1.2. In the latter case we ran the test three times and took the average of the three for the performance charts.
This game is benchmarked in DirectX 11 mode with the highest graphics quality settings.
The 3-way RV870-based configuration expectedly beats the GeForce GTX 480 SLI. The difference is small and due to the architectural specifics of the graphics cores from AMD and Nvidia. That subsystem is going to cost more than $1000 and won’t be quiet, especially in 3D applications, yet it proves that AMD is just as advanced technologically as Nvidia.
The ordinary Radeon HD 5870 CrossFire tandem feels good in this game, being only inferior to the GeForce GTX 480 SLI at 2560x1600. With the addition of a third GPU, AMD’s graphics platform goes ahead and enjoys a considerable lead at the highest resolution. Take note that the third graphics core does not produce serious performance benefits at 1920x1080 and 1600x900.
The 3-way CrossFire is not very successful in this test, only matching the GeForce GTX 480 SLI at 2560x1600. We cannot yet benchmark graphics cards at nonstandard resolutions with two or more monitors whereas the maximum single-monitor resolution of 2560x1600 is playable even on the single-GPU flagship cards from both AMD and Nvidia.
Even the 3-way subsystem from AMD cannot do anything with the resolution of 2560x1600 in this game. The highest performance is still delivered by the GeForce GTX 480 SLI but its bottom speed is not high enough for comfortable play using the highest graphics quality settings. It is the 3-way CrossFire that wins at the lower resolutions, yet we wouldn’t recommend it for this game. It is noisy and consumes a lot of power whereas you can play the game just as comfortably on the single Radeon HD 5970.
The 3-way Radeon HD 5800 configuration is in the lead at resolutions up to 2560x1600. But at 2560x1600 it is only as fast as the GeForce GTX 480 SLI, delivering a higher bottom speed. These graphics solutions are overkill for this game and may only be useful for panoramic mode with multiple monitors. Otherwise, a single Radeon HD 5870 is quite enough for comfortable play.
This game is tested without multisampling antialiasing as it worsens the textures and provokes a performance hit.
AMD wins this test but it can hardly be worth the difference in price and the increased noise. The 3-way CrossFire and the GeForce GTX 480 SLI are almost the same in terms of bottom speed and offer the same gaming comfort.
We use the game’s DirectX 10.1 and DirectX 11 modes for graphics cards that support them.
Nvidia’s solutions have higher bottom speed throughout this test whereas the AMD platform doesn’t improve its bottom speed much even with the addition of a third GPU. The 3-way CrossFire cannot overtake the GeForce GTX 480 SLI tandem in terms of bottom speed, either.
This game is tested in multiplayer mode that uses the OpenGL API. The integrated benchmark does not report the bottom frame rate.
Such advanced graphics configurations as Radeon HD 5800 3-way CrossFire are redundant for this game. You may only need them for playing at panoramic resolutions. Anyway, we can note that the AMD platform is ahead at 2560x1600, leaving the GeForce GTX 480 SLI tandem far behind.
This game’s integrated benchmark does not report the bottom frame rate.
Combining a Radeon HD 5970 with a Radeon HD 5870 into a single graphics subsystem makes sense here. This is the only graphics subsystem to make the game playable at 2560x1600, its bottom speed being always higher than 25 fps. The single Radeon HD 5970 is close to that result, yet the gameplay is smoother on the Radeon HD 5800 3-way CrossFire.
Once again we can see no practical benefits from using a Radeon HD 5970 together with a Radeon HD 5870 in CrossFire mode. This configuration has the same bottom speed as the single Radeon HD 5970. The game having modest system requirements, there is no real need for a top-performance multi-GPU solution.
The integrated benchmark does not report the bottom frame rate.
Top-end multi-GPU configurations are not called for in this game, either. The frame rate is over 200 fps in each mode. The Radeon HD 5800 3-way CrossFire and the GeForce GTX 480 SLI seem to hit the ceiling set by the game engine and the capabilities of the testbed’s CPU.
We enforced full-screen antialiasing using the method described in our special Mass Effect 2 review.
The 3-way CrossFire platform can challenge the GeForce GTX 480 SLI at 2560x1600, losing at the lower resolutions. Take note that its scalability is 50% at 1600x900, which is higher than with the ordinary CrossFire tandems.
We enable the DirectX 11 mode for graphics cards that support it.
The 3-way CrossFire configuration enjoys an advantage at 1600x900 but looks no better than the GeForce GTX 480 SLI at higher resolutions. We don’t see a reason for investing into these expensive and non-economical solutions as you can achieve the same performance with a much cheaper graphics card.
The game’s integrated benchmark cannot report the bottom frame rate. We use DirectX 10 and 10.1 modes here.
Nvidia GPUs remain unrivalled in this game. The most significant achievement of the 3-way CrossFire configuration is that it delivers the same average speed as the GeForce GTX 470 SLI at 2560x1600. The GeForce GTX 480 SLI isn’t much faster at that resolution, though.
We use DirectX 11 mode for graphics cards that support it.
Multi-GPU solutions of both developers seem to have problems with bottom speed in this game. The Nvidia SLI tandems are more or less acceptable with their current drivers at low resolutions whereas AMD’s configurations are no better than 30 fps.
The 3-way CrossFire configuration cannot beat the GeForce GTX 480 SLI tandem at any resolution. Both solutions are comparable in terms of performance, but the AMD one costs more.
We minimize the CPU’s influence by using the Extreme profile (1920x1200, 4x FSAA and anisotropic filtering). We also publish the results of the individual tests across all resolutions.
Overclockers and benchmarkers should appreciate the performance of the 3-way CrossFire platform in 3DMark as it scores over 19,000 points. It is much faster than the Radeon HD 5870 CrossFire tandem.
The main opponents differ more in the second test than in the first one, the Radeon HD 5800 3-way CrossFire being in the lead at every resolution.
GF100-based solutions are always strong in this test, and the AMD platform cannot win even in the 3-way configuration, although it outperforms the GeForce GTX 470 SLI tandem in terms of average performance. The GeForce GTX 400 delivers a higher bottom speed, though.
We can’t say definitely that AMD’s 3-way graphics solution is superior to its opponent. It was indeed faster than the GeForce GTX 480 SLI tandem in some tests but the gap was only large in a very few cases. And there were games where the third GPU did not help it outperform Nvidia’s solution. Besides, the 3-way CrossFire platform did not show high scalability as the summary diagrams indicate.
At the resolution of 1600x900, the 3-way CrossFire configuration wins 10 out of 19 tests and equals the GeForce GTX 480 SLI tandem in one more test. It loses 8 tests, too. The maximum advantage of 51% was observed in Resident Evil 5 and the maximum loss of 35% in H.A.W.X. We must note that this resolution is not really meant for such top-performance solutions as Radeon HD 5800 3-way CrossFire and GeForce GTX 480 SLI.
When it comes to 1920x1080, the score is 11 to 8, but the average advantage of the AMD solution over the GeForce GTX 480 SLI is only 4% now. The difference in the individual tests varies from +49% in Resident Evil 5 to -36% in Tom Clancy’s H.A.W.X. Compared with the dual-processor Radeon HD 5870 CrossFire tandem, the 3-way CrossFire configuration is an average 20% faster, the maximum advantage being 55%. This is hardly worth the $1000 investment into building and assembling such a configuration.
At the resolution of 2560x1600 the 3-way CrossFire wins 12 and loses 7 tests, the maximum loss to the GeForce GTX 480 SLI being no larger than 25%. This is a good result considering that such top-end solutions are meant for playing at the highest resolution, yet the average 6.5% advantage over the GeForce GTX 480 SLI is far from impressive. The scalability is good, the third GPU adding up to 85% to the performance of the AMD platform. The average performance boost is about 33%.
So, just as expected, the Radeon HD 5800 3-way CrossFire is only good for playing at high resolutions: 2560x1600, 2560x1440 or panoramic ATI Eyefinity modes. There is no reason to run such a fast, noisy and power-hungry subsystem in less resource-consuming modes. It must also be noted that the Radeon HD 5970 + 5870 tandem consumes far less power than the GeForce GTX 480 SLI tandem. As for the price factor, the Radeon HD 5970 costs $650-700 and the Radeon HD 5870 about $400. The total cost will be higher than that of two GeForce GTX 480 cards. The small advantage of the AMD solution in terms of performance can hardly justify this difference in price.
The Radeon HD 5870 model from Gigabyte we used for our tests leaves a very positive impression. It refused to overclock at its default GPU voltage, but it already comes with pre-overclocked frequencies: 950 MHz GPU and 1250 (5000) MHz memory. This card is built with high-quality components and offers handy management and monitoring options. Its cooler easily keeps the GPU temperature below 60°C and doesn’t produce much noise. We recommend the Gigabyte Radeon HD 5870 SO (GV-R587SO-1GD) to everyone who wants a fast and reliable version of Radeon HD 5870. It won’t disappoint you.