• GPU: 2 x Radeon HD 6970 (Cayman)
• Interface: PCI Express x16
• GPU frequency (ROPs) (MHz): 830 (o / s 880) (nominal - 830 (o / s 880))
• Memory frequency (physical (effective)) (MHz): 1250 (5000) (nominal - 1250 (5000))
• Memory bus width (bit): 2 x 256
• GPU Computing Units / Unit Frequency (MHz): 2 x 384/830 (o / c 2 x 384/880) (nominal - 2 x 384/830 (o / c 2 x 384/880))
• Number of operations (ALU) per block: 4
• Total number of operations (ALU): 2 x 1536
• Texture units: 2 x 96 (BLF / TLF / ANIS)
• ROP units: 2 x 32
• Dimensions (mm): 330 × 100 × 33 (the last figure is the maximum thickness of a video card)
• PCB color: black
• RAMDACs / TMDS: integrated into the GPU
• Output jacks: 1 × DVI (Dual-Link), 4 × mini-DisplayPort 1.2
• Multiprocessor support: CrossFire X (Hardware)

The card has 2 x 2048 MB GDDR5 SDRAM located in 16 chips (8 on each side of the PCB).

It makes sense to remind once again that the card requires additional power supply, with two 8-pin connectors! Hopefully AMD partners will invest in the appropriate power-splitter kit.

About the cooling system.

AMD Radeon HD 6990 2x2048 MB 2x256-bit GDDR5 PCI-E
Considering the particularly high energy consumption of the product, special requirements are imposed on the cooling system. Therefore, it is no longer possible to make the previously widely used method of positioning a cylindrical fan at the end of a cooler - for a simple reason: the processor units are spaced at the ends of the card. Therefore, the fan is located in the middle of the entire structure. Copper radiators with evaporation chambers and cooling fins are pressed against each core. All this is connected by a common platform, which is also pressed against the power elements in the middle of the PCB. A central cylindrical fan drives air in 2 opposite directions, cooling the radiators. At the same time, for a more highly efficient heat transfer, modern materials are used in the form of interlayers between microcircuits and radiators, which are very fragile and, in fact, disposable. That is why AMD does not recommend disassembling the video card, because after assembly it is no longer possible to obtain the same CO efficiency. As for the noise level, it is quite high under 3D load. The cylindrical fan accelerates quite strongly, and the noise of air passing through the radiators is noticeable. There is no noise in 2D mode or light 3D.

We conducted a study of the temperature regime using the EVGA utility (by A. Nikolaychuk AKA Unwinder) and received the following results:

As we can see, despite the simply colossal power consumption, the heating for the top-end accelerator is normal and does not exceed 90 degrees for the cores. That speaks about the high efficiency of the CO. At peak loads, consumption is about 360 W at the normal core frequency, and 418 W at the overclocking core frequency. So, stock up on heavy duty power supplies.

Equipment. Considering that reference samples never have a complete set, we will omit this question.

Installation and drivers

Test bench configuration:

• Intel Core i7-975 (Socket 1366) based computer
  • Intel Core i7-975 processor (3340 MHz);
  • Asus P6T Deluxe mainboard based on Intel X58 chipset;
  • RAM 6 GB DDR3 SDRAM Corsair 1600 MHz;
  • WD Caviar SE WD1600JD 160 GB SATA hard drive;
  • Tagan TG900-BZ 900W power supply unit.
• operating system Windows 7 64 bit; DirectX 11;
• Dell 3007WFP monitor (30 ″);
• ATI Catalyst 11.2 drivers; Nvidia version 266.58 / 266.66.

VSync is disabled.

Synthetic tests

The synthetic test packages we use can be downloaded from the following links:

• D3D RightMark Beta 4 (1050) with a description on the site http://3d.rightmark.org.
• D3D RightMark Pixel Shading 2 and D3D RightMark Pixel Shading 3- tests of pixel shaders versions 2.0 and 3.0 link.
• RightMark3D 2.0 with brief description: Vista without SP1, Vista with SP1.

Synthetic tests were carried out on the following video cards:

• Radeon HD 6990 OC with GPU frequency increased to 880 MHz - Dual-BIOS factory overclocking mode (hereinafter HD 6990 OC)
• Radeon HD 6990 HD 6990)
• Radeon HD 6970 with standard parameters (hereinafter HD 6970)
• Radeon HD 5970 with standard parameters (hereinafter HD 5970)
• Geforce GTX 570 SLI two GTX 570 video cards in SLI mode with standard parameters (hereinafter GTX 570 SLI)
• Geforce GTX 580 with standard parameters (hereinafter GTX 580)

To compare the results of the new dual-GPU video card Radeon HD 6990, these solutions were chosen because: Radeon HD 5970 is the company's previous dual-GPU solution for the upper price range, the strongest one before the release of the model presented today; Radeon HD 6970 is the fastest single-chip solution from AMD from the same HD 6900 series, based on a similar Cayman video chip.

With solutions from Nvidia, things are a little more complicated. The Geforce GTX 580 is basically the company's fastest graphics card based on the latest GPU. It is not a competitor for the presented AMD video card in terms of price, but its results are interesting precisely as the maximum for single-board solutions from Nvidia at the moment.

But a pair of GTX 570s in SLI mode is taken as a kind of indicator under the conditional name "GTX 590", roughly showing the probable lower the level of performance expected in the spring of the company's two-chip solution. Perhaps such a real Nvidia video card will be based on two full-fledged GTX 580 chips, but just in case we insured ourselves by taking non-top GPU versions as an example.

Direct3D 9: Pixel Filling Benchmarks

This test determines the peak texel rate in FFP mode for a different number of textures applied to one pixel:

Even single-chip solutions in our test for filtering 32-bit (8 bits per color) textures show numbers that are far from theoretically possible. And we got strange results from dual-GPU ones. Nvidia SLI is clearly not working correctly in this test (the numbers are extremely low), and both AMD cards based on two GPUs do not work properly in CrossFire and / or run into the memory bus bandwidth.

Accordingly, there is no point in examining the results of this test from RightMark, we’ll better analyze the texturing speed figures according to the corresponding test from the 3DMark Vantage package below. The only thing that is clear is that all AMD solutions are stronger than all Nvidia graphics cards. And just for our own comfort, consider the same results in the fill rate test:

The numbers show a clear limitation of multi-chip technologies and in them we see everything the same as in the previous diagram. Nominally, the best result remains with the new top-end solution of the Radeon HD 6900 family, although the single-chip analogue is not too far behind. But all this is not very important, because the test results are incorrect.

Direct3D 9: Pixel Shaders benchmarks

The first group of pixel shaders that we are considering is very simple for modern video chips; it includes various versions of pixel programs of relatively low complexity: 1.1, 1.4, and 2.0, found in old games.

The tests are very simple for modern GPUs, even in a single copy, and the rendering speed in them is often limited by the texturing performance. Therefore, they do not show all the capabilities of modern video chips. Nevertheless, dual-GPU solutions get a good performance boost in them. The speed difference between the HD 6970 and the HD 6990 OC ranged from 23% to 91% in these tests. And the more difficult the test, the greater the gain from the second GPU.

The new HD 6990 performed well, outperforming the HD 5970 in all modes, although the difference between them was not very great. But in comparison with Nvidia's solutions, everything is generally fine - the latter are far behind. The reason for this is the poor SLI efficiency for two GTX 570s, which are not far behind one GTX 580. Also, the lack of texturing speed is clearly to blame for the poor performance of Nvidia solutions.

Let's look at the results of more complex pixel programs of intermediate versions:

And this time it turned out about the same as in the most complex tests of the last section. The SLI system of two GTX 570s again competes with a maximum of a single-GPU HD 6970, and not with real dual-GPU models from AMD. In both tests, the gain from CrossFire was almost 100%, but SLI is only effective in the Water subtest. The Cook-Torrance benchmark, which is computationally intensive, does not perform well on an SLI configuration for some reason.

It is clear that these tests are better suited for the AMD architecture, whose chips have a larger number of math and texture units. And the Radeon HD 6990 outperformed its dual-GPU predecessor with a margin, although the difference in Cook-Torrance was also not very large. But competitors from Nvidia are resting again.

Direct3D 9: Pixel Shaders 2.0 benchmarks

These tests of DirectX 9 pixel shaders are more complicated than the previous ones, they are close to what we see in multi-platform games, and fall into two categories. Let's start with the simpler shaders version 2.0:

• Parallax Mapping- a texture mapping method familiar from most modern games, described in detail in the article.
• Frozen Glass- complex procedural texture of frozen glass with controllable parameters.

There are two variants of these shaders: with a focus on mathematical calculations, and with a preference for fetching values ​​from textures. Consider mathematically intensive options that are more promising from the point of view of future applications:

These are universal tests that depend both on the speed of ALUs and on the speed of texturing; the overall balance of the chip is important in them. The performance of the new AMD graphics card model in the Frozen Glass test was only 50% higher than that of the single-chip Radeon HD 6970, although even this figure was sufficient to surpass both the GTX 580 and two GTX 570s with a margin. Alas, but the test again, clearly not for Nvidia, whose solutions are inferior due to weak texturing.

In the second test "Parallax Mapping" Nvidia's solutions feel a little better, but even there they lose, especially given the relatively low SLI efficiency - look, the result of the GTX 570 SLI is close to the GTX 580. But here it is curiously different - the speed of the HD 6990 turned out to be more than twice as high as HD 6970! But the box opened simply - the lack of optimization of the first drivers for the new architecture was to blame, and the updated drivers gave such a powerful acceleration.

It is logical that for Nvidia's solutions everything has once again become noticeably sadder, since with the texturing speed of the latest AMD chips, things are much better. Therefore, the latter are only increasing their advantage. Even a system with two GTX 570s loses to a single-GPU HD 6970 in both tests in the modification with emphasis on texturing. Well, the new two-chip top from the HD 6900 family turns out to be the fastest solution. It decently outperformed its predecessor HD 5970, which is understandable theoretically, because the performance of the TMUs is most important here.

All these were outdated tasks, mainly with an emphasis on texturing, and less often on fillrate. Next, we will look at the results of two more pixel shader tests - but this time version 3.0, the most complex of our pixel shader tests for the Direct3D 9 API. They are most indicative in terms of modern PC games, many of which are multiplatform. The tests differ in that they heavily load ALUs and texture units, both shader programs are complex and lengthy, and include a large number of branches:

• Steep Parallax Mapping- a much more "heavy" version of the parallax mapping technique, also described in the article.
• Fur- a procedural shader that renders fur.

In our toughest DX9 benchmarks, Nvidia's graphics cards have always outperformed AMD's solutions, in contrast to previous benchmarks in the subsection. This is due to the fact that the last two DX9 tests are not limited by the performance of texture sampling, but rather depend on the efficiency of the pixel shader code execution.

But Nvidia again let down a bit the comparatively low efficiency of dual-GPU rendering, since the GTX 570 SLI was not that much stronger than one GTX 580. But the speed of the HD 6990 OC is exactly twice as fast as the performance of the single-GPU counterpart HD 6970, which indicates the maximum efficiency of CrossFire. As a result, in tests of complex pixel shaders version 3.0, the new top-end video card from AMD was able not only to catch up with competitors, but also to outrun a system of two GTX 570s in one of the tests.

The speed in both PS 3.0 tests is weakly dependent on memory bandwidth and texturing, but the code is complex, which both new architectures from Nvidia and AMD do well. In addition, in this test, the difference between the efficiency of the previous and newest architectures of the company (Cypress and Cayman) was found to be maximum.

Direct3D 10: PS 4.0 Pixel Shader Tests (Texturing, Loops)

The second version of RightMark3D includes two familiar PS 3.0 tests for Direct3D 9, which were rewritten for DirectX 10, as well as two more new tests. The first pair adds the ability to enable self-shadowing and shader supersampling, which additionally increases the load on video chips.

These tests measure the performance of executing pixel shaders with loops, with a large number of texture samples (in the heaviest mode, up to several hundred samples per pixel) and a relatively low ALU load. In other words, they measure the texture sampling rate and branching efficiency in a pixel shader.

The first pixel shader test will be Fur. At the lowest settings, it uses 15 to 30 texture samples from the heightmap and two samples from the main texture. Effect detail - “High” mode increases the number of samples up to 40-80, enabling “shader” supersampling - up to 60-120 samples, and “High” mode together with SSAA differs in maximum “severity” - from 160 to 320 samples from the height map.

Let's first check the modes without supersampling enabled, they are relatively simple, and the ratio of the results in the “Low” and “High” modes should be approximately the same.

The performance in this test is most dependent on the number and efficiency of TMUs and ROPs. The results with the detailing of the “High” level are approximately one and a half times lower than with the “Low”, as it should be according to theory. In Direct3D 10 tests of procedural rendering of fur with big amount Nvidia's solutions used to be stronger for texture fetching, but the Radeon HD 6900 family noticeably pulled up to them.

In the variant without supersampling, the effective fill rate and memory bandwidth have a greater impact on performance. Therefore, the single-chip solution from Nvidia turned out to be ahead of the single-chip HD 6970, but due to the high efficiency of CrossFire, we again see that the HD 6990 has become the leader and outperforms the SLI configuration of two GTX 570 cards in the factory overclocked mode. It seems that Nvidia would like to make a dual-GPU card not based on a GTX 570, but on two GTX 580s ...

Let's look at the result of the same test, but with “shader” supersampling enabled, which quadruples the work, perhaps in such a situation something will change, and the memory bandwidth with fill rate will have less effect:

As always, the inclusion of supersampling increases the theoretical load by four times and the comparative results of Nvidia solutions drop noticeably. Now HD 6970 performs at the level of GTX 580, and two GTX 570s can only compete with the outdated HD 5970. The new dual-GPU solution HD 6990 outperforms all other video cards in both frequency modes. Again, we see a very good result from a fresh solution from AMD.

The second DX10 shader test measures the performance of complex pixel shaders with loops with a large number of texture samples and is called Steep Parallax Mapping. At low settings, it uses 10 to 50 texture samples from the heightmap and three samples from the main textures. When you turn on the heavy mode with self-shadowing, the number of samples doubles, and supersampling increases this number four times. The most complex test mode with supersampling and self-shadowing selects from 80 to 400 texture values, that is, eight times more than the simple mode. We first check the simple options without supersampling:

The second pixel-shader test Direct3D 10 is somewhat more interesting from a practical point of view, since varieties of parallax mapping are widely used in games, and heavy options like our steep parallax mapping are used in many projects, for example, in Crysis and Lost Planet games. Besides, in our test, in addition to supersampling, you can enable self-shadowing, which doubles the load on the video chip, this mode is called “High”.

The diagram is very similar to the previous one without supersampling anti-aliasing, but the positions of both Nvidia solutions have weakened. In the updated D3D10 version of the test without SSAA, the HD 6970 gets closer to the GTX 580, although it lags a little behind. On the other hand, the new Radeon HD 6990 now outperforms the GTX 570 SLI in all modes and frequency configurations. In this test, it has a clear advantage over the previous dual-GPU solution HD 5970. Let's see what will change the inclusion of supersampling, because it usually causes an additional drop in speed on motherboards with Nvidia GPUs.

When supersampling and self-shadowing are enabled, the task is even more difficult; enabling both of these options together increases the load on the cards by almost eight times, causing a large drop in performance. The difference between the speed indicators of the tested video cards has changed, the inclusion of supersampling affects the same as in the previous case - AMD cards slightly improved their indicators relative to Nvidia solutions.

And again, with SSAA enabled, the single-chip HD 6970 shows results at the level of the GTX 580, and the new product based on two GPUs outperforms the SLI system from a pair of GTX 570s. The comparative difference between Cayman and Antilles is close to twofold, which again speaks of the excellent performance of CrossFire. SLI also performs well, but the relative weakness of the GTX 570 prevents a pair of these cards from catching up to the HD 6990.

In general, based on a couple of recent tests, we can conclude that the dual-chip board of the HD 6900 family released today outperforms its potential competitor from Nvidia if it is based on two GTX 570s or has similar characteristics: the number of execution units and the clock speed. ...

Direct3D 10: PS 4.0 Pixel Shader Benchmarks (Compute)

The next couple of pixel shader tests contain the minimum number of texture fetches to reduce the impact of TMU performance. They use a large number of arithmetic operations, and they measure exactly the mathematical performance of video chips, the speed of execution of arithmetic instructions in a pixel shader.

The first math test is Mineral. This is a complex procedural texturing test that uses only two texture data samples and 65 instructions like sin and cos.

The results of the extreme mathematical tests usually correspond to the difference in frequencies and the number of execution units, with little influence on their effectiveness. In such cases, the modern AMD architecture has a huge advantage over competing video cards from Nvidia, and this perfectly explains the results obtained.

The theoretical difference in AMD's favor is so great that even a system with two GTX 570s performs only slightly better than a single-GPU Radeon HD 6970! It is clear that even the HD 5970 outperforms two GPUs from Nvidia, although the theoretical difference between them is even greater.

But the theory changes little in the event of a confrontation between HD 6990 and GTX 570 SLI - the new top-end video card from AMD is 80% faster in simple mathematical calculations, compared to a hypothetical dual-GPU Nvidia video card based on two GF110s with the characteristics of GTX 570. In this case, and CrossFire and SLI effectively coped with the parallelization of work on two chips, and the difference between the HD 6970 and HD 6990 OC is close to twofold.

Consider the second shader computation test called Fire. It is heavier for ALU, and there is only one texture fetch in it, and the number of instructions like sin and cos is doubled, to 130. Let's see what has changed with increasing load:

And this time all the solutions remained in approximately the same positions. The only difference is that the relative performance of the two Cypress chips has improved slightly. Otherwise, there are practically no changes. Since the rendering speed in the test is limited exclusively by the performance of the shader units, the HD 6990 is again the clear leader.

The new model is followed by a dual-GPU HD 5970 from the same AMD, and both systems based on Geforce are inferior to them. The GTX 570 SLI is again not far from the single-chip HD 6970. Nvidia in the next generation clearly needs to change something, increasing the mathematical power.

Direct3D 10: geometry shader benchmarks

The RightMark3D 2.0 package contains two tests of the speed of geometry shaders, the first version is called "Galaxy", the technique is similar to the "point sprites" from previous versions of Direct3D. It animates a particle system on the GPU, a geometric shader from each point creates four vertices that form a particle. Similar algorithms should be widely used in future DirectX 10 games.

Changing the balancing in the geometry shader tests does not affect the final rendering result, the final image is always exactly the same, only the scene processing methods change. The "GS load" parameter determines in which of the shaders the calculations are performed - in vertex or geometric. The number of calculations is always the same.

Let's consider the first variant of the Galaxy test, with calculations in a vertex shader, for three levels of geometric complexity:

The ratio of speeds at different geometric complexity of scenes is approximately the same for all solutions, the performance corresponds to the number of points, with each step the FPS drop is almost twofold. The task for modern video cards is not particularly difficult, overall performance is limited not only by the geometry processing speed, but also by the memory bandwidth or fill rate to a certain extent (only within the GPU of one manufacturer).

Alas, although the new AMD chips are distinguished by increased geometric performance, they have nothing to brag about in this test. The factory-overclocked HD 6990 OC is only slightly ahead of the single-GPU GTX 580. It's no wonder that two GTX 570s in SLI became the sole benchmark leader, 50% faster than the HD 6990. And the new AMD card is not far behind the previous generation Radeon HD 5970. Perhaps this is the fault of the memory bandwidth limitation or the ROP units do this work ineffectively, as we noted earlier.

Let's see how the situation will change when transferring part of the calculations to the geometry shader:

When the load changed in this test, the numbers for solutions from both Nvidia and AMD remained almost unchanged. Although video cards of the HD 6900 family in this test still react slightly to changes in the GS load parameter, which is responsible for transferring part of the calculations to the geometric shader, showing results slightly higher than in the previous diagram. But this does not change anything, in fact, a serious lag behind the GPU architecture of Nvidia remains. Let's see what will change in the next test, which assumes a heavy load on geometry shaders.

Hyperlight is the second geometry shader test that demonstrates the use of several techniques at once: instancing, stream output, buffer load. It uses the dynamic creation of geometry using rendering in two buffers, as well as a new Direct3D 10 feature - stream output. The first shader generates the direction of the rays, the speed and direction of their growth, this data is placed in a buffer that is used by the second shader for rendering. For each point of the ray, 14 vertices are built in a circle, up to a million output points in total.

A new type of shader program is used to generate "rays", and with the "GS load" parameter set to "Heavy" - also for their rendering. That is, in the "Balanced" mode, the geometry shaders are used only for creating and "growing" rays, the output is carried out using "instancing", and in the "Heavy" mode, the geometry shader is also involved in the output. Let's look at light mode first:

It is immediately clear that the test for comparing multi-chip systems is simply not suitable. For some reason, both CrossFire and SLI in this test are not only inoperable, but also show abnormally low results, much worse than similar single-chip systems. This is somewhat similar to the fill rate test performed by the GTX 570 SLI. Well, we compared single-chip video cards back in December - Nvidia's solutions are clearly faster here.

However, the numbers may change in the next diagram, in a test with more active use of geometry shaders.

But no, we see exactly the same thing here - the Hyperlight test is not suitable for multi-chip graphics systems. Both Nvidia SLI and AMD CrossFire perform equally inadequately. In any case, according to previous comparison materials and the results of the GTX 580 and HD 6970, we can see that Nvidia graphics cards cope with geometry shaders much better than AMD solutions.

And it is unlikely that the Radeon HD 6990 will change anything here, well, except that two GPUs from AMD will still catch up with one chip produced by Nvidia. Probably, AMD engineers have yet to solve the problem of parallelizing the operation of the geometry setup blocks, which these tests can run into.

Direct3D 10: speed of fetching textures from vertex shaders

The Vertex Texture Fetch tests measure the speed a large number texture samples from the vertex shader. The tests are similar in essence and the ratio between the results of the cards in the tests "Earth" and "Waves" should be approximately the same. In both tests, it is used based on texture samples, the only significant difference is that the Waves test uses conditional transitions, while the Earth does not.

Consider the first test "Earth", first in the "Effect detail Low" mode:

Previous studies have shown that the results of this test are influenced by both texturing speed and memory bandwidth, to some extent. And multichips, in contrast to the last test of geometric shaders, perform very well here.

The Radeon HD 6990, based on two new GPUs, performs very well, even just above the two GTX 570s in SLI. Which are far behind only in the simplest mode (for Nvidia video cards in this test, the performance is clearly limited by something).

The efficiency of dual-GPU configurations is quite good, the difference between HD 6990 and HD 6970 is almost two times. That is why the new AMD solution becomes the leader of the test. Let's look at the performance in the same test with an increased number of texture fetches:

In the second VTF test, the ratio of the results changed noticeably, especially in the hard mode. Although Nvidia graphics cards continue to perform poorly in the lightest conditions. Probably, with a small number of polygons, the rendering speed is limited by something (it doesn't look like memory bandwidth), and in this case the new AMD motherboard even outperforms a potential dual-GPU solution from Nvidia.

But in heavy modes, the difference remains in favor of Nvidia. The previous two-GPU top Radeon HD 5970 shows speed at the level of a single-GPU GTX 580, and the HD 6990, even when overclocked, lags behind a system of two GTX 570s, although not too much.

Let's consider the results of the second test of texture samples from vertex shaders. The Waves test has a smaller number of samples, but it uses conditional jumps. The number of bilinear texture samples in this case is up to 14 ("Effect detail Low") or up to 24 ("Effect detail High") for each vertex. The complexity of the geometry changes in the same way as in the previous test.

The results in the second test of vertex texturing - "Waves", are traditionally very different from what we saw in the previous diagrams. In this test, single-chip video cards show similar results, which can be attributed to the limited video memory bandwidth, but the efficiency of two-GPU rendering is very high - almost 100%.

And therefore, given that the Cayman is a full-frequency and full-fledged top-end GPU, and there are two of them installed on the HD 6990, this solution comes out on top in comparison, and the SLI system of two Geforce GTX 570 lags behind the new AMD board quite strongly. Let's consider the second variant of the same test:

The difference compared to the previous diagram is extremely small. Although the speed of Nvidia video cards "sagged" much less than that of AMD solutions. Therefore, in the mode with a small number of polygons, the Radeon HD 6990 presented today remains the best, but on average and heavy the system on two GTX 570s is already in the lead. And a single GTX 580 competes with the old, but two-chip HD 5970.

3DMark Vantage: Feature tests

The synthetic benchmarks from the 3DMark Vantage suite may show us something that we missed earlier. Feature tests of this test suite have DirectX 10 support and are interesting because they differ from ours. When analyzing the results of new video cards in this package, we can make some new and useful conclusions which eluded us in the tests of the RightMark family. Unfortunately, the company's even newer test suite, 3DMark11, does not contain specialized synthetic tests and is not at all interesting to us in this case.

The first test is the texture sampling speed test. Used to fill a rectangle with values ​​read from a small texture using multiple texture coordinates that change every frame.

The results of video cards in the texture test by Futuremark are always close to the theoretically possible level of texture fetching speed and the efficiency of video cards from AMD and Nvidia in it is slightly higher than in ours. In this test, we always get a different ratio of results compared to RightMark, and even more so this time.

So, the diagram clearly shows that AMD video cards are much faster in texturing compared to their competitors. Even the single-chip model Radeon HD 6970 shows results close to two GTX 570s, not to mention dual-GPU solutions. The dual-GPU HD 6990 is the clear leader in the test, because the texture performance of the Cayman has noticeably increased compared to the Cypress, and there are also two GPUs installed, not one. An excellent result for a novelty!

This is a fill rate test. A very simple pixel shader is used with no performance limitation. The interpolated color value is written to the offscreen buffer (render target) using alpha blending. A 16-bit FP16 off-screen buffer is used, which is most often used in games that use HDR rendering, so this test is quite timely.

Here we also see a completely different situation compared to our fillrate test. The Futuremark subtest measures ROP performance, with almost no impact on video memory bandwidth. Here Nvidia cards compete quite well with AMD solutions and do it quite well.

The single-GPU Geforce GTX 580 is clearly faster than the analogous Radeon HD 6970, and the two GTX 570s are only slightly lacking in the HD 6990. And yet, in order for the future dual-GPU card from Nvidia to compete strongly for the HD 6990, it will need to be based on two chips with characteristics similar to the GTX 580, not the GTX 570.

The influence of the memory bandwidth on the test results can be seen from the HD 6990 figures in two modes. At first it may seem strange that there is almost no difference between them, but the explanation for this is very simple - in the mode of increased frequencies, only the GPU clock frequency changes, and the video memory frequency remains nominal, limiting the memory bandwidth. Therefore, the results are shown to be similar.

Feature Test 3: Parallax Occlusion Mapping

One of the most interesting feature tests, as a similar technique is already used in games. It draws one quadrangle (more precisely, two triangles), using a special technique Parallax Occlusion Mapping, which simulates complex geometry. Quite resource-intensive ray tracing operations and high resolution depth map are used. Also this surface is shaded using the heavy Strauss algorithm. This is a test of a very complex and GPU-heavy pixel shader containing numerous texture selections for ray tracing, dynamic branches and complex lighting calculations using Strauss.

This test differs from other similar ones in that the results in it depend not exclusively on the speed of mathematical calculations or the efficiency of branch execution or the speed of texture fetching, but on a little bit of everything. And to achieve high speed, the right balance of performance of different GPU units is important. Affects the speed and efficiency of branching in shaders.

Comparative results for AMD video cards are very similar to what we saw in the texture performance test a little higher. Well, the solutions of the Nvidia board received a slight increase in speed, which means that not only texture performance affects the test results. As a result, even taking into account the pulling up of Nvidia solutions, the new dual-GPU model from AMD remains the leader. A pair of GTX 570s successfully compete only with the HD 5970 on the previous generation GPU.

The test is interesting in that it calculates physical interactions (imitation of tissue) using a video chip. Vertex simulation is used, using combined work of vertex and geometry shaders, with several passes. Use stream out to transfer vertices from one simulation pass to another. Thus, the performance of vertex and geometry shaders execution and stream out speed are tested.

Rendering speed in this test depends on many parameters, the main of which are geometry processing performance and geometry shader execution efficiency. As in our tests, Nvidia's graphics cards perform clearly better in this application, well ahead of the competition.

The test clearly shows the advantage of the new generation of AMD architecture over the old one. After all, the Radeon HD 6970, based on one Cayman, copes with the dual-GPU HD 5970 (perhaps its CrossFire did not work in this test). The performance gain from the second Cayman was almost 100%, but this still did not allow even getting close to a pair of GTX 570s. Moreover, even one GTX 580 shows the speed only 20% lower than that of the two-GPU novelty. Most likely, this is due to the difference in the organization of the graphics pipeline for processing geometry.

Physical simulation test of effects based on particle systems calculated using a video chip. Vertex simulation is also used, each vertex represents a single particle. Stream out is used for the same purpose as in the previous test. Several hundred thousand particles are calculated, all are animated separately, and their collisions with the height map are also calculated.

Similar to one of the tests in our RightMark3D 2.0, particles are rendered using a geometry shader, which creates four vertices from each point, forming a particle. But the test most of all loads shader units with vertex calculations, stream out is also tested.

The results of another test from the 3DMark Vantage package are similar to those we saw in the previous diagram, and here the geometry processing speed is also most important. Therefore, the last generation in the form of a Radeon HD 5970 card has become an outsider, lagging behind both systems from Nvidia, which are the clear leaders, as well as the new models of the HD 6900 family based on Cayman chips.

The power of the chips of the new AMD architecture is only enough to compete with the single-GPU Geforce GTX 580, and a pair of GTX 570s ran far ahead. It turns out that in synthetic tests of imitation of tissues and particles from the 3DMark Vantage test suite, in which geometric shaders are actively used, the solutions of the HD 6900 line continue to lag behind competing video cards with excellent geometry processing speed.

The last feature test of the Vantage package is a mathematically intensive test of the video chip, it calculates several octaves of the Perlin noise algorithm in a pixel shader. Each color channel uses its own noise function for more load on the video chip. Perlin noise is a standard algorithm often used in procedural texturing and uses a lot of math.

But in a purely mathematical test from the Futuremark package, showing the peak performance of video chips in extreme tasks, we saw approximately the same as in similar tests from our RightMark 2.0 package - a complete defeat of Nvidia solutions.

But here's the strange thing - the result of HD 6990 in normal mode is abnormally low. With a theoretical difference of 6%, it turned out to be more than a quarter of the lag behind the overclocked mode. Why? We think that the too clever power management system PowerPlay is to blame for this, which reduced the clock frequency on the overclocked version, and with it the performance of the solution when the set power consumption threshold was reached. And in "OC" mode, the system simply spread its wings.

Otherwise, there were no revelations. Nvidia's boards are far behind AMD solutions even from other price ranges in peak mathematics, so the comparison is far from in their favor. The two GTX 570s miss even the Radeon HD 5970, let alone the HD 6990 in the factory overclocked Dual-BIOS. Nvidia and three GF110s on one board won't help here ...

Conclusions on synthetic tests

Based on the results of synthetic tests of a new video card from the Radeon HD 6900 family, based on two Cayman GPUs, as well as the results of other models of video cards produced by both manufacturers of discrete video chips, we can conclude that this is the fastest video card on the market today, and An excellent replacement for the Radeon HD 5970.

Thanks to the installation of two powerful GPUs based on the new AMD architecture, the Radeon HD 6990 video card released today is even further ahead of the competitor in the face of the best (but single-chip) Nvidia solution - the Geforce GTX 580. And before the release of a dual-chip video card from this company, if any will take place , it is the Radeon HD 6990 that will be the fastest. In most synthetic tests, it outperformed other models, except for those rare applications in which performance is limited by the speed of geometry processing, tessellation, and geometry shader execution.

Even the single-chip Radeon HD 6970 showed decent speed in synthetics, and the high efficiency of the CrossFireX mode allowed the HD 6990 to become almost twice as fast. And the advantage of the video card presented today is especially noticeable in texturing tests and peak mathematical performance, which has become traditional for AMD solutions.

Of the potential drawbacks, in addition to lagging in geometric performance, we can note not very outstanding results in mathematical tests, compared to the HD 5970. And also the very real possibility that the PowerTune power management system can reduce clock frequencies when reaching maximum power consumption in some the most demanding synthetic tests, not allowing them to show higher performance.

But these little fly in the ointment make little difference. As we already wrote above, the new solution will certainly turn out to be the most productive video card in gaming tests as well. And the results of the Radeon HD 6990 in synthetic tests simply have to be confirmed by the corresponding figures in the "game" part of our material. Accordingly, the price of the solution turned out to be high, but it is fully justified by the fact that the Radeon HD 6990, in principle, has no competitors.

Speaking of competitors. According to numerous rumors, Nvidia is ready to release its version of a dual-chip board on two GF110 chips soon. And the outcome of the two-GPU battle will depend on what characteristics these two GPUs endow, whether they are closer to the GTX 570 or the GTX 580. According to our tests, it is obvious that it is highly desirable for Nvidia to get as close as possible to the frequencies and the number of execution units available in the GTX 580 to outperform the announced solution, especially in the factory overclocked mode.

But it will not be so easy to do this, because the main limitation will be the rather high power consumption of the GPU from Nvidia. In the meantime, the characteristics of their future video card are unknown, and we can only guess about them. And to make the fortune-telling, let's look at the results of the Radeon HD 6990 gaming tests.