Overclocking nvidia geforce gtx 670 reference card. Architectural features, technical characteristics

• GPU: GeForce GTX 670 (GK104)
• Interface: PCI Express x16
• GPU operating frequency (ROPs): 915-1040 MHz (nominal - 915-980 MHz)
• Memory operating frequency (physical (effective)): 1500 (6000) MHz (nominal - 1500 (6000) MHz)
• Memory bus width: 256 bit
• Number of computational units in the GPU/block operating frequency: 7/915-1030 MHz (nominal - 7/915-1030 MHz)
• Number of operations (ALU) in block: 192
• Total number of operations (ALU): 1344
• Number of texturing units: 112 (BLF/TLF/ANIS)
• Number of rasterization units (ROP): 32
• Dimensions: 245×100×33 mm (the last value is the maximum thickness of the video card)
• PCB color: black
• Power Consumption (Peak 3D/2D/Sleep): 172/64/52 W
• Output Jacks: 2×DVI (Dual-Link/VGA), 1×HDMI 1.4a, 1×DisplayPort 1.2
• Multiprocessor support: SLI (Hardware)

The card has 2048 MB of GDDR5 SDRAM memory located in 8 chips (4 on each side of the PCB).

The card requires additional power, and two 6-pin connectors.

About the cooling system.

Nvidia Geforce GTX 670 2048 MB 256-bit GDDR5 PCI-E
The cooling system as a whole is traditional. The long shroud, which increases the card's size from 190mm to 245mm, has a standard cylindrical fan at the end. A radiator based on an evaporation chamber is installed on the core, and power transistors also have their own radiator; both are cooled by a single air flow. But the memory chips are left without cooling; let me remind you, they are located on both sides of the card. When the core heats up to maximum, the fan speeds up to about 60-66% (2300 rpm) of its maximum, which makes the noise a little noticeable.

We conducted a temperature study using the new version of the EVGA PrecisionX utility (author A. Nikolaychuk AKA Unwinder) and obtained the following results. I would like to remind you that the GTX 670 core operates at floating frequencies from 915 to 980 MHz (according to Nvidia’s declaration). At the same time, the actual core operating frequency reaches 1030 and even 1040 MHz. We don't yet know whether this is a feature that is generally true for all GTX 670s, or whether the reference card sample has slightly higher frequencies than mass-produced cards will have.

Let's return to monitoring.

Nvidia Geforce GTX 670 2048 MB 256-bit GDDR5 PCI-E

As we can see, after 6 hours of running the card under maximum gaming load, the maximum core temperature was 82 degrees, which, in principle, is normal for a top-series model.

The video card arrived to us without packaging or kit, so we are omitting the question of packaging.

Installation and drivers

Test bench configuration:

• Computer based on Intel Core i7-975 (Socket 1366)
  • Intel Core i7-975 processor (3340 MHz);
  • Asus P6T Deluxe motherboard on Intel chipset X58;
  • RAM 6 GB DDR3 SDRAM Corsair 1600 MHz;
  • hard drive WD Caviar SE WD1600JD 160 GB SATA;
  • Enermax Platimax 1200 W power supply.
• operating room Windows system 7 64-bit; DirectX 11;
• monitor Dell UltraSharp U3011 (30″);
• AMD Catalyst 12.4 drivers; Nvidia version 301.24

VSync is disabled.

Synthetic tests

The synthetic test packages we use can be downloaded here:

• D3D RightMark Beta 4 (1050) with a description on the website 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 a brief description: for Vista without SP1, for Vista with SP1.

As synthetic DirectX 11 tests, we used examples from the Microsoft and AMD SDKs, as well as the Nvidia demo program. First, there are HDRToneMappingCS11.exe and NBodyGravityCS11.exe from the DirectX SDK (February 2010).

We also took applications from both video chip manufacturers: Nvidia and AMD. The examples DetailTessellation11 and PNTriangles11 were taken from the ATI Radeon SDK (they are also in the DirectX SDK). Additionally, Nvidia's demo program, Realistic Water Terrain, also known as Island11, was used.

Synthetic tests were carried out on the following video cards:

• GeForce GTX 670 GTX 670)
• GeForce GTX 680 with standard parameters (hereinafter GTX 680)
• GeForce GTX 580 with standard parameters (hereinafter GTX 580)
• Radeon HD 7970 with standard parameters (hereinafter HD 7970)
• Radeon HD 7950 with standard parameters (hereinafter HD 7950)
• Radeon HD 7870 with standard parameters (hereinafter HD 7870)

To compare the results of the Geforce GTX 670 video card released today, these solutions were chosen for the following reasons. Geforce GTX 580 is the older single-chip model of the previous generation and is close to the GTX 570, which the new product replaces on the market, and the GTX 680 is the top single-chip model modern architecture Nvidia's Kepler.

We took the selected solutions from the competing company AMD for testing because the Radeon HD 7950 has a price close to the announced Geforce video card and is its direct competitor on the market. at the moment. The Radeon HD 7970 is taken as AMD's top model, a comparison with which can also be interesting, and the HD 7870 model is used in several tests as an auxiliary one, because it is also very close to the Radeon HD 7950 in speed.

Direct3D 9: Pixel Shaders tests

We will look at the texturing and fill rate tests from the 3DMark Vantage package below, and the first group of pixel shaders we use includes various versions of relatively low complexity pixel programs: 1.1, 1.4 and 2.0, found in older games, and it is very simple for modern video chips.

These tests are simple for modern GPUs, and the speed in them is often limited by texturing performance or fill rate. Therefore, these tests are not able to show all the capabilities of modern video chips, but they are interesting to us from the point of view of analogues of outdated gaming applications, of which there are still quite a lot.

Judging by our previous comparisons, the performance of new video cards in these tests is most often limited by fill rate, although the influence of the speed of texture modules can also be seen. But it is definitely implicit, since the GeForce GTX 680 did not become the winner as it should, based on texture performance. General results do not allow us to identify any exceptional characteristic that affects the overall rendering speed.

In these tests, AMD's top video card is clearly the leader, but the presented Geforce GTX 670 model should be compared not with the Radeon HD 7970, but with the HD 7950. Which is noticeably inferior in these tests. And here the situation is no longer in favor of the AMD board. Perhaps it was let down by software optimizations in the drivers (different versions were used), because there should not be such a difference. This can be seen in comparison with the HD 7870.

In these tests, the GeForce GTX 670 performs very well good level, inferior to the older GTX 680 by only about 10%, which is clearly less than the theoretical difference in the speed of texturing and mathematical calculations. And compared to the competing model Radeon HD 7950, everything is very good. Let's look at the results of more complex intermediate pixel programs:

The Cook-Torrance test is more computationally intensive, the difference in it approximately corresponds to the difference in the number of ALUs and their frequency, but it also depends on the speed of the TMU. Therefore, this test is historically better suited to AMD graphics solutions, but now GeForce based on the Kepler architecture is also very strong in it.

And the large difference in speed between the Radeon HD 7970 and HD 7950 again played in Nvidia’s favor - in a comparison of direct competitors, the GeForce GTX 670 again comes out as a clear winner. As for absolute leadership, in one of the tests (where fast mathematics is more important) Radeon is slightly faster, and in the other (where texture performance is more important) GeForce is ahead.

Direct3D 9: pixel shader tests Pixel Shaders 2.0

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

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

There are two variants of these shaders: those with a focus on mathematical calculations and those with a preference for sampling values ​​from textures. Let's consider mathematically intensive options that are more promising from the point of view of future applications:

These are universal tests, the performance in which depends on both the speed of ALU units and the texturing speed; the overall balance of the chip and the efficiency of execution of computer programs are also important in them. The test results show that AMD's architecture is significantly ahead of Nvidia's GPUs in these specific tasks.

The performance of new AMD video cards in the “Frozen Glass” test is significantly higher than that of the new product, but this applies more to the Radeon HD 7970, while the younger model HD 7950, although it remained ahead, is not so much. Most likely, Nvidia chips simply do not perform this task efficiently, or the drivers are not well optimized for it. It may also be that the speed is limited by throughput. Indeed, in comparison with the previous GTX 580 model, the new product showed an extremely small increase in speed.

In the second “Parallax Mapping” test, the new Nvidia video card showed performance comparable to what we got from the Radeon HD 7950, although the latter is again clearly hampered by something in the software part, since the older AMD model remains the sole leader. Nvidia solutions in this test cannot catch up with the top AMD board for some reason, such as insufficient bandwidth. Let's consider these same tests in a modification with a preference for samples from textures over mathematical calculations:

For boards with Nvidia GPUs, the situation has become better, and the GeForce GTX 670 is now always faster than the Radeon HD 7950 and slightly slower than the HD 7870 only in the first test. But still, the speed of GeForce comes up against something, and modern AMD chips work more efficiently in these tasks. The new GeForce GTX 600 series video card is quite strong in the Parallax Mapping test, and lags only behind the two older models of both companies, and in the Frozen Glass test the lag behind them is greater. The difference between GTX 680 and GTX 670 is 9-12%, which is less than theoretical.

These were legacy tasks, with an emphasis on texturing and 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 Direct3D 9. They are most indicative from the point of view of modern PC games, including many multi-platform ones. The tests differ in that they heavily load both the ALU and texture modules; both shader programs are complex and lengthy, and include a large number of branches:

• Steep Parallax Mapping- a much more “heavy” type of parallax mapping technique, also described in the article Modern terminology of 3D graphics.
• Fur— a procedural shader that renders fur.

These tests are no longer limited by the performance of texture samples only, but most of all depend on the efficiency of shader code execution. In the most difficult DX9 tests from the first version of the RightMark package, Nvidia video cards were previously stronger, but in the latest architecture, AMD has changed the situation and now it is the top solution on the GCN architecture chip that shows the best result in the PS 3.0 comparison.

But our today's hero also showed good results, losing to its direct competitor Radeon HD 7950 in only one of the tests, winning in the second. “Fur” generally shows an excellent result, better than that of the HD 7950 and HD 7870. At the same time, the new product lags behind the older model Geforce GTX 680 by 14-17%, which is much closer to the theoretical superiority of the top-end video card over the one presented today. And it means that there is almost no emphasis on throughput in this test.

Direct3D 10: PS 4.0 pixel shader tests (texturing, loops)

The second version of RightMark3D included 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 added the ability to enable self-shadowing and shader supersampling, which further increases the load on video chips.

These tests measure the performance of pixel shaders running in cycles with a large number of texture samples (in the heaviest mode, up to several hundred samples per pixel) and a relatively small ALU load. In other words, they measure the speed of texture samples and the efficiency of branches in the pixel shader.

The first test of pixel shaders will be Fur. At the lowest settings, it uses 15 to 30 texture samples from the height map and two samples from the main texture. The Effect detail mode - “High” increases the number of samples to 40-80, the inclusion of “shader” supersampling - up to 60-120 samples, and the “High” mode together with SSAA is characterized by maximum “heaviness” - 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 results in the “Low” and “High” modes should be approximately the same.

In this test, performance depends largely on the number and efficiency of TMUs, but the efficiency of executing complex programs is also strongly influenced. And in the version without supersampling, the effective fill rate and memory bandwidth also have an additional impact on performance. The results at the “High” level of detail are up to one and a half times lower than at the “Low” level.

As in similar DX9 tests, in tasks of procedural fur visualization with a large number of texture samples, over a couple of generations of graphic architectures, AMD not only reduced the difference with Nvidia boards, but with the release of GCN even took the lead. And now we very often see the Radeon HD 7970 among the leaders in such comparisons, which indicates high efficiency execution of complex pixel programs.

The Geforce GTX 670 model reviewed today showed results at the level of the GTX 580, that is, almost worse than all of them, which may indicate either a decreased efficiency in the execution of complex shaders in Kepler or a lack of memory bandwidth or effective fillrate. There is no way to compete with the competitor Radeon HD 7950, because it is much stronger.

Let's look at the result of the same test, but with shader supersampling enabled, which increases the work by four times: perhaps in this situation something will change, and memory bandwidth with fill rate will have less effect:

Yes, both video cards of the new GeForce GTX 600 line clearly improved the results relative to the old GeForce GTX 580, but when supersampling was enabled, which quadrupled the theoretical load, the results of Nvidia solutions as a whole still deteriorated significantly compared to the performance of AMD video cards. The difference in rendering speed in this task was already high, but now it has become simply huge.

The tested new product from Nvidia is ahead only of the GTX 580, and loses up to 50% to its competitor from AMD - the video card of the Radeon HD 7950 model. And the top board from this HD 7000 series shows the highest performance in the test, which indicates GCN’s “love” for complex calculations. The advantage in these tests is clearly with AMD chips, which prefer pixel-by-pixel calculations.

The next DX10 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 height map and three samples from the main textures. Enabling heavy mode with self-shadowing doubles the number of samples, and supersampling quadruples this number. 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. Let's first check simple options without supersampling:

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

The diagram is very similar to the previous one without the additional inclusion of SSAA, and Nvidia's solutions in this test did not improve their position one gram. The new GeForce GTX 670 in the updated D3D10 version of the test without supersampling still lags behind its direct rival Radeon HD 7950 and outperforms only the GTX 580. Let's see what the change will be when enabling supersampling, because it usually causes a strong drop in speed on Nvidia boards.

And here everything is about the same as in “Fur”. When supersampling and self-shadowing are enabled, the task becomes even more difficult; enabling both options together increases the load on the cards by almost eight times, causing a serious drop in performance. The difference between the speed indicators of the tested video cards has changed, the inclusion of supersampling has an effect, as in the previous case - AMD video cards have clearly improved their relative performance compared to boards based on Nvidia chips.

Although this time the difference between the GeForce GTX 670 and the Radeon HD 7950 has decreased somewhat, and the lead over the GTX 580 is no longer so great. It is clear that the Radeon HD 7970 is again far ahead, but even the younger AMD model is very good and outperforms even the GeForce GTX 680, not to mention the younger modification. Kepler's efficiency in these tasks is clearly lacking...

Direct3D 10: PS 4.0 Pixel Shader Tests (Compute)

The next couple of pixel shader tests contain a minimum number of texture fetches to reduce the performance impact of the TMU units. They use a large number of arithmetic operations, and they measure precisely 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 samples of texture data and 65 sin and cos instructions.

The results of limiting mathematical tests most often more or less correspond to the difference in frequencies and the number of computational units, but with the influence of different efficiency of their use. AMD architectures of the last few years have had an overwhelming advantage over competing Nvidia video cards in such cases, but it was in Kepler that the number of stream processors and peak mathematical performance increased significantly.

That’s right - the results of video cards are located on the diagram approximately in accordance with the theory. And the GeForce GTX 670 in this test was completely ahead of its direct competitor, the Radeon HD 7950! For the first time in many, many years, the Californian company's GPU outperformed a similarly priced competitor's graphics card offered at a similar price. However, in this test the Radeon should have had a higher speed, but both GeForce GTX 600s performed clearly better this time.

Let's look at the second shader calculation test, which is called Fire. It is heavier for an ALU, and there is only one texture fetch, and the number of sin and cos instructions has been doubled, to 130. Let's see what has changed with increasing load:

And in the second mathematical test, the relative result of the new Nvidia product turned out to be lower, and it lost to the Radeon HD 7950 by about as much as it should according to theory. Therefore, the younger of the two Radeons took the lead, although not by much. The rest of the positions have not changed, the GTX 580 from the previous generation of Nvidia boards is far behind - that's what the new architecture means, suitable for applications of this kind.

The rendering speed in this test is limited solely by the performance of the shader units and their efficiency, so both Radeon boards show strong results, and the top one became the leader in the comparison. But the GeForce GTX 680 and GTX 670 are not inferior as much as they were in previous generations, when the difference was almost several times greater. Therefore, we confirm the conclusion from the Kepler review - in extreme computing tasks, with the release of the new architecture, the difference between AMD and Nvidia has become not at all as large as it was before. Moreover, in the comparison between GeForce GTX 670 and Radeon HD 7950 there is no clear leader at all.

Direct3D 10: geometry shader tests

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

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

Let's look at the first version of the Galaxy test, with calculations in the vertex shader, for three levels of geometric complexity:

The ratio of speeds for different geometric complexity of scenes is approximately the same for all solutions, performance corresponds to the number of points, with each step the FPS drops almost twofold. This task is not very difficult for modern video cards, and performance is limited either by the speed of geometry processing or by memory bandwidth.

The most noticeable difference is between the new generation of Nvidia boards and AMD solutions. If in previous tests with pixel shaders AMD boards were clearly overall more efficient and faster, then the first geometry test shows that Nvidia boards are still ahead in such tasks. Our new GeForce GTX 670 is not too far behind the GTX 680 and is ahead of the top single-chip board from the previous generation by a margin.

As for comparing the new product with its main competitor, the result of the comparison is logical - the difference between the GeForce GTX 670 and Radeon HD 7970 is almost one and a half times. In this generation, Radeon boards were only able to catch up with the best of Fermi, and that’s good. Let's see how the situation changes when we transfer part of the calculations to the geometry shader:

As the load changed in this test, the numbers remained the same for the legacy Nvidia board and improved slightly for the newer AMD and Nvidia boards. All video cards in this test react weakly to changing the GS load parameter, which is responsible for transferring part of the calculations to the geometry shader, so all the conclusions remain the same. The GeForce GTX 670 model presented today is more than one and a half times faster than the Radeon HD 7950. Let's see what will change in the next test, which assumes a greater load on geometry shaders.

“Hyperlight” is the second test of geometry shaders, demonstrating the use of several techniques at once: instancing, stream output, buffer load. It uses dynamic geometry creation by drawing to two buffers, and new opportunity Direct3D 10 - 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, which is used by the second shader for drawing. 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 programs is used to generate “rays”, and with the “GS load” parameter set to “Heavy” - also to draw them. That is, in the “Balanced” mode, geometry shaders are used only to create and “grow” rays, the output is carried out using “instancing”, and in the “Heavy” mode, the geometry shader is also involved in output.

Relative results in different modes also roughly correspond to the change in load: in all cases, performance scales well and is close to theoretical parameters, according to which each subsequent level of “Polygon count” should be slightly less than twice as slow. So far there are few changes, except that AMD boards have moved higher.

In this test, rendering speed is limited primarily by geometric performance, but with some impact from video memory bandwidth. This time, the GeForce GTX 670 broke away from its rival Radeon HD 7950 by only 30-40%. The numbers should change a lot in the next chart, in a test with more active use of geometry shaders. It will also be interesting to compare the results obtained in the “Balanced” and “Heavy” modes with each other.

This time, the diagram with the transfer of calculations to the geometry shader has changed quite seriously and all Nvidia video cards turned out to be clearly faster than AMD cards, even if we take the GeForce GTX 580 of the previous generation. Well, the new Kepler-based boards cope with the task even better, since this test rests specifically on the performance of geometric blocks, with the power of which Nvidia solutions are more than good.

Therefore, Geforce has an advantage over AMD chips with a traditional graphics pipeline. The new GeForce GTX 670, even in heavy mode, shows results like the Radeon HD 7950 on average. That is, there is about 70-80% difference in speed between them, and this is a lot. So, although the results of competitor boards based on Tahiti have noticeably improved, latest solutions based on the GK104 chip are significantly ahead of them in this category of tests.

Direct3D 10: texture fetching speed from vertex shaders

The Vertex Texture Fetch tests measure the speed of a large number of texture fetches from the vertex shader. The tests are essentially similar, so the ratio between the cards' results in the Earth and Waves tests should be approximately the same. Both tests use displacement mapping based on texture sample data, the only significant difference is that the Waves test uses conditional branches, while the Earth test does not.

Let's look at the first "Earth" test, first in the "Effect detail Low" mode:

Our previous research has shown that the results of this test can be affected by both texturing speed and memory bandwidth, especially in easy mode. And the results of Nvidia video cards are completely limited by something else incomprehensible - look, the numbers for all GeForces in all conditions are simply identical. And in general, between boards of similar class, the difference in this test is sometimes very small.

In this test, the Radeon HD 7970 came out far ahead, and the same HD 7950, which is the main rival for the GTX 670, showed slightly better results than the new product under consideration. But the difference is not that big, especially in easy and medium modes. The new board of the GTX 600 family was able to compete with the HD 7950 only in light and medium modes, and lagged behind by almost a third in heavy modes. Due to the slight difference with the GTX 680, we can assume that this happened due to low fill rate and/or bandwidth. Let's look at the performance in the same test with an increased number of texture samples:

The relative position of the cards on the diagram has changed mainly due to the fact that Nvidia boards provided the same rendering speed in all modes, unlike AMD solutions, which lost ground a little. That is, our version about the speed of Californian video cards hitting some kind of barrier is confirmed. And now the results of the GeForce GTX 670 are not just close to the speed of the Radeon HD 7950, but it turned out to be faster than its opponent in all modes, not just in easy.

Let's look at the results of the second test of texture fetches 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 the previous test.

The results in the second "Waves" vertex texturing test are nothing like what we saw in the previous charts. But we note the same oddity in them - all the presented Nvidia video cards lined up almost in the same line. But for Radeon boards everything is different; we can separately highlight the excellent speed of the top-end Radeon HD 7970, which was the best in comparison.

And the Radeon HD 7950 was clearly stronger than the Geforce GTX 670 board presented today. All tested Nvidia solutions again ran into something unclear, showing almost identical results. Let's consider the second version of the same test:

This time there were also changes similar to those we saw earlier - AMD video cards slightly worsened their results, and Nvidia cards also suffered in light modes. This allowed the Californian company's boards to get a little closer to the results of the Radeon HD 7970 and HD 7950, but still, the new Kepler architecture presented today lost to its direct competitor, the Radeon HD 7950, in two out of three modes. Vertex texturing tests once again measured something unclear.

3DMark Vantage: Feature tests

Synthetic tests from the 3DMark Vantage package will show us what we previously missed. Feature tests from this test package support DirectX 10 and are interesting in that they differ from ours and are still relevant. When analyzing the results of the new Nvidia video card in this package, we will draw some new and useful conclusions that eluded us in tests from the RightMark family of packages.

The first test is a texture fetch speed test. This involves filling a rectangle with values ​​read from a small texture using multiple texture coordinates that change every frame.

Although the Futuremark test does not show the theoretically possible level of texture fetch performance, the efficiency of AMD and Nvidia video cards in it is quite high and the comparative figures of the models are always quite close to the corresponding theoretical parameters. But not always - after all, according to theory, the best video card in comparison should be the Geforce GTX 680 model, but it showed less efficiency in the test compared to the Radeon HD 7970 and lost the lead to it.

But in the case of comparing a pair of Geforce GTX 670 and Radeon HD 7950, everything turned out differently. The new product has overtaken its direct competitor, and the GeForce GTX 580 from the previous generation has remained far behind. Even in conditions of low bandwidth, the result of the GTX 670 turned out to be quite good, which once again shows that some of the shortcomings of the Fermi architecture have been successfully resolved in the new generation of GPUs.

This is a fill rate test. A very simple pixel shader is used that does not limit performance. The interpolated color value is written to an off-screen buffer (render target) using alpha blending. The 16-bit off-screen buffer of the FP16 format is used, which is most often used in games that use HDR rendering, so this test is quite timely.

The situation in the ROP block performance test is very different. We determined earlier that the numbers for this subtest from 3DMark Vantage, although they show the performance of ROP units, are greatly influenced by the amount of video memory bandwidth (the so-called “effective fill rate”). This test often measures memory bandwidth rather than ROP performance.

This is what happened this time too, because the new GeForce GTX 670 model has exactly the same memory bandwidth performance as the GTX 680. Therefore, their results are almost identical. Moreover, the Geforce GTX 670 showed more results at the level of the Radeon HD 7970 than its direct competitor HD 7950. Although the difference between all the new generation GPUs is small, they are all far ahead of the Geforce GTX 580, which indicates improvements in the efficiency of the TMU units.

One of the most interesting feature tests, since a similar technique is already used in games. It draws one quadrilateral (more precisely, two triangles) using a special Parallax Occlusion Mapping technique that simulates complex geometry. Quite resource-intensive ray tracing operations and a high-resolution depth map are used. This surface is also shaded using a heavy Strauss algorithm. This is a test of a very complex and heavy pixel shader for a video chip, containing numerous texture samples during ray tracing, dynamic branching and complex lighting calculations according to Strauss.

This test differs from the ones we conducted above in that the results depend not solely on the speed of mathematical calculations, the efficiency of branch execution, or the speed of texture fetches, but on everything at once. And to achieve high speed, the correct balance of the GPU is important here, as well as the efficiency of executing complex shaders.

In synthetics from 3DMark Vantage, Geforce and Radeon boards show approximately the same relative results as in similar tests from our test package. The new GeForce GTX 670 board, although it is ahead of the previous GTX 580, is 16% behind the top-end GTX 680. This test is also affected by simplifications in Kepler computational units; the efficiency of the new architecture in such tasks is lower than that of Fermi and GCN.

Therefore, in comparison with the competing AMD board from the Radeon HD 7900 family based on the latest GCN architecture, Nvidia's new product did not shine in the test. The new GeForce GTX 670 lost to the HD 7950 by more than 25%. So, in such computing tasks, AMD video cards still cope with the work more efficiently.

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

The rendering speed in this test depends on several parameters at once, but the main influencing factors are the performance of geometry processing, the efficiency of execution of geometry shaders, and the performance of ROP blocks. Due to the influence of geometric blocks, it is logical that Nvidia video cards perform well in this application, outperforming the corresponding Radeons.

The GeForce GTX 670 model presented today easily outperforms its competitor in the form of the Radeon HD 7950 model. Since this is one of the tests in which the advantage of Nvidia solutions with several geometric blocks is visible, even the top AMD solution is left behind. It is also interesting that the new product based on the GK104 chip presented today is quite a bit ahead of the outdated model based on the GF110 chip - the difference between them is only slightly more than 1%. Most likely, bandwidth limitations or low fill rates are to blame.

Test of physical simulation of effects based on particle systems calculated using a video chip. Vertex simulation is also used, each vertex representing 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 our RightMark3D 2.0 tests, particles are rendered using a geometry shader that creates four vertices from each point to form a particle. But the test most of all loads shader units with vertex calculations; stream out is also tested.

The results of this test from the 3DMark Vantage package would be similar to those we saw in the previous diagram, if not for the difference in speed between the GeForce GTX 580 based on the Fermi architecture. This is one of the very few tests in which new boards based on a chip with the Kepler architecture are significantly inferior to the best representative of the Fermi architecture. And in the case of this test, the comparative results are most likely explained by the fill rate indicator, because its peak value is higher in the GTX 580.

But if we compare the speed of the GeForce GTX 670 with the performance of its main rival, then the new Nvidia product is expected to be ahead - and by 30%. In the two synthetic tissue and particle simulation tests from the 3DMark Vantage benchmark suite, which make extensive use of geometry shaders, little has changed - although Nvidia's new boards are limited low performance PSP and fill rate, the GTX 670 is still faster than the Radeon HD 7950.

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 the pixel shader. Each color channel uses its own noise function to put more stress on the video chip. Perlin noise is a standard algorithm often used in procedural texturing, it uses a lot of mathematical calculations.

In a purely mathematical test from the Futuremark package, showing the peak performance of video chips in extreme tasks, we see a different distribution of results compared to similar tests from our test package. In this case, the performance of the solutions does not fully correspond to the theory and is at odds with what we saw earlier in mathematical tests from the RightMark 2.0 package.

It is clear that video cards based on AMD's GCN architecture chips cope with such tasks perfectly, Radeon cards always show best results in cases where relatively simple but very intensive mathematics is performed. Therefore, it is no wonder that AMD’s top solution became the leader, and the Geforce GTX 670 model being reviewed today was inferior to its direct rival Radeon HD 7950 by more than 20%.

Both models based on GK104 chips have clearly too low efficiency in this task. In theory, the GTX 680 should be twice as fast as the GTX 580, but no such difference was noted in reality. Even in a relatively simple test, the Kepler architecture chip shows reduced efficiency when executing shader programs, since there are no other explanations - the same lower bandwidth should not affect this.

Direct3D 11: Compute Shaders

To test Nvidia's new solution on tasks that use new DirectX 11 features like tessellation and compute shaders, we used samples from the SDKs and demos from Microsoft, Nvidia, and AMD.

First we'll look at tests that use Compute shaders. Their appearance is one of the most important innovations in the latest versions of the DX API, they are already used in modern games to perform various tasks: post-processing, simulations, etc. The first test shows an example of HDR rendering with tone mapping from the DirectX SDK, with post-processing , using pixel and compute shaders.

Although this is not the most successful example for computational shaders, it shows the difference in performance in one specific task. There is practically no difference in the speed of calculations in computational and pixel shaders for Nvidia boards with video chips of the Kepler architecture, as well as for AMD boards. Judging by previous tests, the results in the problem clearly depend not only on mathematical power and not only on the efficiency of calculations, but also on something else, like bandwidth.

Moving on to specific solutions, we note that the new Nvidia product in this test lags behind the older GTX 680 quite a bit, but the Radeon HD 7970 is inferior. Unfortunately, we do not have the results of the competing Radeon HD 7950 in DX11 tests. But there is the Radeon HD 7870, which is cheaper and clearly slower in these tasks. The second compute shader test is also taken from the Microsoft DirectX SDK and shows a computational N-body gravity problem - a simulation of a dynamic particle system that is subject to physical forces such as gravity.

In this test, the results are completely different, and the difference between the GeForce GTX 680 and the new GTX 670 was more than 18%, which is very close to the theoretical differences between these models. But in general, Kepler does an excellent job, since the Geforce GTX 670 is not only faster than the HD 7870, which is not its competitor, but also outperforms the top AMD board. So this test is unlikely to measure the speed of simple mathematical calculations.

Most likely, its speed is highly dependent on the efficiency of complex calculations, in which Nvidia was previously strong. In general, the result of the new product is very good, especially considering the significant gap even from the best board of the competitor. Performance tests in tessellation tasks will most likely show another strong side of the new GeForce.

Direct3D 11: Tessellation Performance

Compute shaders are very important, but another important innovation in Direct3D 11 is hardware tessellation. We looked at it in great detail in our theoretical article about the Nvidia GF100. Tessellation has been used for quite some time in DX11 games, such as STALKER: Call of Pripyat, DiRT 2, Aliens vs Predator, Metro 2033, Civilization V, Crysis 2, Battlefield 3 and others. Some of them use tessellation for character models, others use tessellation to simulate realistic water surface or landscape.

There are several different schemes for partitioning graphic primitives (tessellation). For example, phong tessellation, PN triangles, Catmull-Clark subdivision. Thus, the PN Triangles partitioning scheme is used in STALKER: Call of Pripyat, and in Metro 2033 - Phong tessellation. These methods are relatively quickly and easily implemented into the game development process and existing engines, which is why they have become popular.

The first tessellation test will be the Detail Tessellation example from the ATI Radeon SDK. It implements not only tessellation, but also two different pixel-by-pixel processing techniques: simple normal map overlay and parallax occlusion mapping. Well, let's compare AMD and Nvidia DX11 solutions in different conditions:

As we have already seen, parallax occlusion mapping on video cards from both manufacturers is much less efficient than tessellation, and tessellation does not significantly reduce performance - compare the upper and lower columns in the diagram. That is, high-quality imitation of geometry using pixel calculations provides worse performance than real geometry with tessellation and displacement mapping.

In a simple bump mapping test, it is clear that the boards most likely run into memory bandwidth. This is why the Radeon HD 7970 is so strong and outperforms even the GeForce GTX 680, not to mention today's new product. The second subtest with more complex pixel calculations showed that the efficiency of performing complex mathematical calculations in pixel shaders for GCN architecture chips is still higher than for Nvidia GPUs. The top video card of the Radeon HD 7000 family showed the best result in the parallax mapping test, and even the relatively cheap HD 7870 outperformed the GTX 670 presented today in this subtest.

And in the most interesting tessellation subtest, we can note the results of Radeon video cards - they are very strong. In this tessellation test, the division of triangles is moderate (we know from our previous materials that the subtest is not entirely limited to geometry processing speed), and therefore AMD boards do not lose too much performance, and their speed reserve is enough to show the best results. Particularly surprising is the Radeon HD 7870, which, due to the high clock speed of the GPU and optimizations in drivers, has overtaken even the top video card.

The second tessellation performance test will be another example for 3D developers from the ATI Radeon SDK - PN Triangles. Actually, both examples are also included in the DX SDK, so we are sure that game developers create their code based on them. We tested this example with different tessellation factors to understand how much impact changing it has on overall performance.

In this example, we see a more plausible comparison of the geometric power of various solutions. All modern chips handle light to medium geometric workloads quite well, but Nvidia GPUs remain unsurpassed in the toughest conditions. Unfortunately, we do not have results with the maximum tessellation level (tessellation factor = 19) for the Radeon HD 7870; we wrote about the reasons in the corresponding review.

All chips from the Fermi and Kepler architectures are good at such tasks, but the top solutions in the GeForce GTX 600 line are the best in these indicators. The GTX 670 solution presented today is only slightly inferior to the top-end GTX 680 model, and outperforms the Radeon HD 7970 in average conditions by almost two times, and in the most difficult conditions by more than three times. Although the GCN architecture chips were noticeably faster in tessellation, this did not allow them to catch up with the GK104.

Let's take a look at the results of another test, the Nvidia Realistic Water Terrain demo, also known as Island. This demo uses tessellation and displacement mapping to render realistic-looking ocean surfaces and terrain.

Let us recall that Island is not a purely synthetic test for measuring only geometric performance; it contains complex pixel and computational shaders, too, and such a load is closer to real games that use all GPU blocks at once, and not just geometric ones, as in previous benchmark.

We tested the demo at four different tessellation ratios, in this case the setting is called Dynamic Tessellation LOD. And if at the very first triangle partitioning factor, when the speed is not limited by the performance of geometric blocks, AMD video cards are very strong, then when the geometric work becomes more complex, Nvidia cards begin to win seriously. As the partitioning factor and scene complexity increase, the performance of both Radeons drops quite significantly, even despite obvious software optimizations in the case of the Radeon HD 7870.

The presented new Nvidia product is inferior to the GTX 680 in heavy modes by up to 9-11%, but it wins too much over the top competing Radeon HD 7970. Although the number of geometric blocks in GK104 did not increase, the increased clock speed of the video chip and greater mathematical power allowed the boards of the new family to show even stronger results. The latest solutions from AMD have seriously improved geometric performance and in real applications, where there are no ultra-high degrees of triangle partitioning, they are practically not inferior to Nvidia solutions, but in synthetic geometry tests the winner is always known in advance.

Conclusions on synthetic tests

Based on the results of synthetic tests of the new model of the Geforce GTX 670 video card, based on the “cut down” GK104 graphics processor of the Kepler architecture, as well as the results of other models of video cards produced by both manufacturers of discrete video chips, we can conclude that the new near-top solution from Nvidia will become one of the most profitable video cards in the upper price segment. Even in terms of technical characteristics and synthetic tests, the gap between the new model and the GeForce GTX 680 turned out to be not so great, which means that excellent results will be shown in games.

We have already reviewed the GK104 graphics processor earlier; it is the first-born of the new Kepler architecture and is made using the most advanced technical process to date. The new architecture chip has several improvements, the main ones of which are aimed at increasing energy efficiency. Our set of synthetic tests showed that the GeForce GTX 670's performance in almost all tasks is close to the GeForce GTX 680 and Radeon HD 7970, and is often higher than that of the competing Radeon HD 7950 model. Not to mention geometry tests, where Nvidia solutions have an advantage just huge.

Among the shortcomings, we again note the reduced efficiency of the new GPU in some tests of complex pixel shaders, such as Parallax Occlusion Mapping and Fur, in which the new board is inferior to its rival in the form of Radeon HD 7950, as well as limited video memory bandwidth and its size. In some games, these restrictions may not allow the new product to show all its capabilities. On the other hand, these shortcomings extend back to the GTX 680, and in the GTX 670 it was simply impossible to install more fast memory larger volume - they left everything the same, and thanks for that.

Thanks to architectural changes in Kepler, Nvidia has improved the energy efficiency characteristics of its top solutions, including the Geforce GTX 670. The new GK104 chip has very high performance, while consuming not much energy - according to official data, less than the Radeon HD 7950. This has improved consumer characteristics, and the Geforce GTX 670 video card should become a very good offer for those enthusiasts who do not want to buy the most expensive top solutions.

We are confident that the good results of the new GeForce GTX 670 model in most synthetic tests will be supported by excellent performance in gaming applications from our test set. The new product should show high speed in games compared to its rivals and become one of the most attractive video cards in its price segment.