Introduction

After the presentation of the nVidia GeForce 7800 series it looked for a long time dismal for ATi. The nVidia GPU was not only extremely fast, she offered also a good image quality, thanks to the transparency AA a feature, not to offer the ATi had. Although rumors of the X 850 successor circulated shortly thereafter, more than speculation, you should get to see “Wunderkind” long time but not from the new. In October, ATi then presented the X 1000 series with a bang and wanted to regain lost ground by an aggressive appearance. To achieve this goal, you put on a new architecture as well as significantly higher clock rates, and an angle-independent texture filtering (even if it is logically not completely).

The result was not only from the point of view of the performance is very good, because it surpassed the former top dog, the GeForce 7800 GTX at times quite significantly. On both in terms of image quality the Canadians put it commendably a shovel, which ATi could outstrip the competition (–_hiermit_möchten_wir_einen_kleinen_Appell_an_nVidia_richten,_in_einem_neuen_Chip_ebenfalls_eine_winkelunabhängige_Texturfilterung_einzusetzen.) (The GPUs of the Californians are because for this fast enough once!)

This setback was nVidia apparently prepared and presented 7800 GTX 512 a few days later the new flagship, the GeForce. This is how the top model of the X 1800 series on a 512 MB VRAM and increased exorbitantly, which is why we again the performance Crown with a small ledge in itself could tear the clock rates compared to the previous version – the disadvantage of poorer image quality remained but still exist. Again relatively soon after the development cycle when ATi turns continues and a “new” chip is coming: the Radeon-X 1900 series. That would bring back back the winner’s trophy in the speed to ATi. Aim shall be achieved this goal Interestingly not by significantly higher clock speeds, but due to some changes in the chip architecture.

The R580 chip has per pixel pipeline three shader units, which is why Silicon at 4 pixel quads logically 48 such beer available are the new piece. The predecessor of R520 had to make do, however, still with an ALU per pipeline. Still, ATi improved smaller details in the R580, which we will dwell on in the technology section of the article. By new X 1900 series there currently three different graphics cards, on the name “Radeon X 1900 XT”, “Radeon X 1900 XTX” and “Radeon X 1900 CrossFire Edition” will hear and are available according to the Canadians from today.

I am delighted that reached both the Radeon X 1900 XTX as the Radeon X 1900 CrossFire Edition our test lab, and it is going to show whether the new high-end models against the nVidia GeForce 7800 GTX and whose SLI configuration can points. Also will be seen, whether ATis may be much praised Avivo technology to improve the quality of the video or the video acceleration against NVIDIA “PureVideo”. With the game is for the first time the “HQV benchmark”, with which a direct comparison of the DVD playback is ensured.

Bookmark

• nVidia GeForce 7800 GTX 512 SLI
• nVidia GeForce 7800 GTX 512
• nVidia GeForce 7800 GTX (SLI)
• nVidia GeForce 7800 GT
• ATi Radeon X 1800, X 1600 and X 1300
• Club3D Radeon X 1800 XT CrossFire Edition
• Radeon X 1000 vs. GeForce 7
• ASUS EAX1800XT TOP
• ASUS extreme N7800GT dual
• Sapphire Radeon X 800 GT, X 800 GTO and X 800 GTO m²
• nVidia GeForce 6800 GS
• ATi CrossFire – power of two graphics cards

Technical data

< div >

Technology in detail

At first glance the new R580 chip differs from ATi not of the “veteran” R520, which was first presented in October of last year. So again 16 pixel pipelines with ever a texture mapping unit (TMU), as well as a raster operation processor (ROP) are in the X 1900 series used, which you, as it is used with the nVidia G70 competitor – apart from the reduction of ROPs in terms of the number of pixel processors – holds still the traditional architecture. Also the 256-bit wide memory interface including the ring-bus memory controller remains the same. Also like its predecessor, the R580 puts on eight vertex shader, in the scope of the function has nothing – changed where so the difference is?

Of course, the slightly higher clock rates with XTX-version in the fall one, the difference is so small that he will be responsible only for a small performance gain. More than the R520 then finally shows a deeper look into the “inner life” the reason why the R580 chip, which again is made in the 90-nm process from TSMC, about 384 million transistors is almost 60 million. The R580 boasts three equivalent shader units per pipeline instead of just one like the R520, which is why one X 1900 video cards to a total of 48 “arithmetic logical units”, short ALU, may have recourse. Theoretically the shader processing power at the same time has increased accordingly by a factor of three. This 3: 1 ratio between beer and text ureinheiten ATi with the always anfordernderen Shaderanweisungen in modern games, while the text urberechnungen rather stagnant at a level. The field test ATi has started already 1600 with the RV530 in the form of the Radeon X with just that 3: 1-ratio is equipped.

However, what makes an ALU at all? An ALU can perform various (arithmetic) operations such as add (addition), SUB-(subtraction), emergency (Negierungen) or MUL (multiplication) statements that are the calculation of shader needed. In comparison per pipeline offers nVidia GeForce 7800 series two shader units, the each a MADD (multiply-ADD: multiplication and addition) can run. Thus the G70 chip 48 has beer in 24 pixel pipelines (GeForce 7800 GTX) and 40 beer at 20 pixel pipelines (GeForce 7800 GT) on. A NV40, however, has only 16 beer, although on this there are two shader units. The first shader unit can charge however no MADD but only an ADD. Thus, this shader unit is missing the MUL command and is generally not considered full ALU. When an ALU are now described as a “real” shader unit can, as so often unfortunately unclear. We therefore speak of a full ALU, if they can process a MADD function, since this is currently the crucial operation for the placed on the market performance.

Each of the 48 shader units in the R580-chip dominated the following calculation functions:

• 1 Vec3 ADD + input modifier
• 1 Scalar ADD + input modifier
• 1 Vec3 ADD / MULL / MADD
• 1 Scalar ADD / MULL / MADD
• 1 Flow control instruction

There is interesting news about the “dispatch processor”. This the shader pixel splits large block in the R520 in a 4 x 4, which is why not only the latencies are kept extremely low, but in addition the “dynamic branching” (jump statement by if or else commands in a shader) cannot be executed extremely fast. By the higher number of beer in the RV530 (as with the R580 a 3: 1 ratio) the shaders are divided large block in a 4 x 12, which the chip is somewhat inefficient working. ATi informed us on request but that this is not the case at the R580. This used a 4 x 4 pixels of large thread; the GPU renders with the same effectiveness as the R520. But away from the shader units, there is detail enhancement. The R580 offers compared to the predecessor of R520 with an improved HierarchicalZ function. HierarchicalZ is the first stage of the calculation of non-visible pixels.

First, each pixel is divided smaller pixels in a block be individually examined for visibility. While blocks, which are completely invisible, are discarded directly, are entirely or partly visible pixel blocks further directed in the pixel engine. This approach can see other important memory bandwidth savings are and higher resolutions are more feasible. HierarchicalZ calculation, an on-chip memory is required, but quite small. If this memory is too small for extremely high resolutions, the depth test early works much more slowly. To counteract this effect, the Canadians on the R580 ruining a twice as large on-chip memory, which even in high resolutions such as 2560 x 1600, HierarchicalZ can work with full effectiveness.

The last significant change is a feature called “Fetch4″, which is used at shadow maps. Shadow maps are used in many games to represent shadows. A disadvantage of this method are the hard edges of the shadows, which do not occur in a real environment; There, the shadows are mostly soft. To work around this problem, the GPU can take a certain number of samples from the shadow map and filter them in the pixel shader bilinear. Thus, the course of the shadow is soft. This method but also has a disadvantage, so decreases performance on many samples in the basement because many commands must be executed (per sample a so-called TEX command). Now, Fetch4 comes into play, speeding up this process.

The R580 exploits at Fetch4 the fact that each color value consists of four values – red, green, blue, and the alpha channel (transparency value). Fetch4 can now those values at the same time sample in a single pass, as well as calculate and requires only a TEX command; much faster you can create a shadow map. Fetch4 however has a disadvantage because the shadow maps are not filtered and thereby produce often unsightly shadows. To filter to the data in the pixel shader sent and charged there, what takes some CPU time to complete. NVIDIA procedure called PCF (percentage closer filtering) goes one step further and filters the shadow map in a single pass in the TMU. Fetch4 and PCF are used currently, for example, in the 3DMark06, as well as the games of far cry, Battlefield 2 and age of Empires 3.

HQV benchmark details

The HQV (Hollywood quality video) benchmark from Silicon Optix is on a standard DVD and you want to test the image quality of the playback device on the basis of different image sequences – a perfect test for the Avivo-and PureVideo technology from ATi or nVidia. There are a total of 18 test or test videos, you should evaluate themselves with 10 points, and finally the addition of points results in the overall result. Thus, the HQV benchmark is naturally somewhat subjective, because a human eye is not the other. Nevertheless, the HQV benchmark are a good impression in the playback quality of a video card.

The first test to listen to the name “color bar/vertical detail” and shows a still image with several different color strips. This test is to determine whether the video processor can distinguish a picture of a moving picture, because these are handled differently by the integrated de-Interlacer. If the de-Interlacer the related image correctly detect all strips are presented in detail. If not, flare these strips or disappear even beyond recognition.

The second test is called “Jaggies Pattern 1″ and do see how well the video processor can detect moving objects. This sequence consists of a circle with a pointer that rotates 360 degrees. This, depending on the angle, the pointer begins to flare. The more this to the flickering tends, the worse the video processor works. This needs to discard some pixels at a moving object to reduce flicker. Depending on the which pixels are selected, add to, or to reduce the error.

“Jaggies Pattern 2″ builds on the previous test. Three close together lying bar move it alternately upward and downward. Because this happens again in a circle, both the direction and the speed and the angle of the pointer change. The video processor must again the pixels to discard and add see because the beams otherwise flicker.

The test of “Waving Flag” includes a – who would have thought it? -waving American flag with a building in the background. This sequence is to show how well the video processor with a cope moving and at the same time a related object. The real problem is caused by the many strips moving in different directions and angles at the same time. If it comes into the movement phase to visible noise in the form of aliasing effects, the video processor is in a weakness.

The sequence “Detail Enhancement” shows a picture of a landscape with a road, a high staircase, as well as a bridge and thus represents a complex challenge for the player, because all objects must be presented in detail. These must be brought not only in the correct sharpness, but also without edges and effects of flicker on the screen.

“Noise Reduction” to show exactly what the name implies: noise, caused largely by a poor compression or poor-quality editing. Often, this problem occurs with images with a high colour spectrum, which is why it is “Color noise” called. The test result is obtained from various images such as a sunset or a yellow flower in the bright sunlight.

The next section called “motion adaptive noise reduction” different “noise reduction” provides techniques to the test (for example, “temporal noise filtering”). This filter works on the concept that changing pixels in the image with color noise equivalent and therefore must be edited. However generated pixels also changing a moving object in a related image and processing by the noise-reduction-filer would the image “mutilated”. For this reason, any good video processor can tell the difference between two States.

The test of “3: 2 detection is to come inter alia applies as the”moire effect”on the go. The most films are recorded at 24 frames per second. At the port of footage on a DVD, for example, the data need to convert from the current format in the format of the DVD. One such possible technique is called 3: 2 pulldown”, in which an extra frame is being reviewed. Image errors can arise however, that a high-quality video processor must detect and remove.

“Film Cadence” shows several such methods such as 3: 2 pulldown”, which are used according to the footage. The video processor must identify as this, because otherwise applies.

The last sequence “3: 2 Film Mixed with added video titles” generates a horizontally and vertically running text. If, for example, a film shot with 30 FPS, and then cast in standard 24 frames per second with the 3: 2 pulldown”technology, it can quickly come to quality scrollendem text aimed at preventing a good video processor.

HQV benchmark

We test the HQV benchmark on an LCD TV so that you can rather detect the error in the presentation and give a realistic result. The graphics are via a DVI cable as well as the corresponding adapter connected to the HDMI connector of your TV. Being as additional comparison to a proper DVD player Panasonic S97EG used, is also connected via the HDMI connector with the display device.

As demonstrated by the HQV benchmark, ATi has promised not too much with its Avivo technology and presents a really good video processor, which is well above the video technology of the G70 chip with regard to image quality. nVidia, there still has to catch up. The test run shows but something else: to reach a good playback quality to a relatively large TV, eventually not step on a “real” DVD player passes – the Panasonic has the video chips in the graphics card in its place. Nevertheless ATi Avivo technology in indicates the R580 that something positive has complained and that more and more people use the video capabilities of modern graphics chips. There is no differences between the R520 and the R580, however. This was however not to expect, because ATi has announced no improvements to the Avivo-part.

We deliberately outside left the PureVideo Decoder from nVidia, although this would have possibly achieved an improvement. In contrast to Avivo, PureVideo is charged, which is why the Avivo competitor is not tested.

Impressions

ATi Radeon X 1900 XTX

The top of the line with the new R580 chip from ATi listens to the name of X 1900 XTX and thus heralds a new name shortcut for ATi graphics cards. The suffix XT PE (Platinum Edition) for a minimal later switched mode flagship seems passé to be final. The graphics card to go according to the Canadians from today for $ 649 of the counter, which it adjusts not only the level of the GeForce 7800 GTX 512, but sets up a new record price for ATi GPUs also. The PCB is XT strongly reminiscent of that of the Radeon X 1800 and is in fact largely identical; only few members were replaced.

The map measures 23 cm and fits easily into almost all housing. The dual-slot cooler remains absolutely identical 1800 XT to the X. The cooling system to bring the R580 to pleasant temperatures and thus a 6.5 cm large radial fans is used. This works under Windows pleasantly quietly, but the propeller in a 3D application produces an unpleasant noise level, which is accompanied by a high-frequency singing – here is still catching up to do, because a GeForce 7800 GTX 512 by nVidia is considerably quieter!

In addition, the high weight of the graphics card falls on what is caused by a large copper cooling plate that is attached to the GPU. The fan sucks the cool air from the enclosure in the cooling system, forwards it on the graphics core, which is cooled down to, and again blows out the heated air from the tower. The converters are protected by a passive heat sink from the overheating, while the back without any cooling has to make do. As usual two dual-link enabled DVI ports and a HDTV output, which has for ViVo features are installed on the bracket.

The R580 chip on the Radeon X 1900 XTX added 650 MHz, which produces a very high performance of geometry in combination with the eight vertex shaders. The fill rate compared to the Radeon X continues to 1800 XT only slightly as there have been no changes to the structure of the pipeline. The shader performance is, however, significantly higher at the X 1900 XTX because it has the triple number of shader units per pipeline. The 512 MB of GDDR3 memory comes again from Samsung, is equipped with an access time of only 1.1 ns – something surprising, at the “low” memory clock. Accordingly, there is still plenty of room for a spacious overclocking. The memory bandwidth is so much lower than with a GeForce 7800 GTX 512. To save power in 2D mode in addition, added to the Radeon X 1900 XTX download 594 MHz and 500 MHz, respectively.

ATi Radeon X 1900 CrossFire Edition

The ATi Radeon X 1900 CrossFire Edition a clone of the X 1900 XTX looks like. The fan system is identical, works but logically something louder than just a map with due to the additional graphics card. Nevertheless the propellers on Windows of non-negative, while disturbs the noise level in a 3D application. Although the temperature values of the CrossFire team are very high, we have seen any crashes or image errors during the load test.

The CrossFire Edition works with the same clock speed as the X 1900 XT, talk to 625 MHz for the chip and 725 MHz for the memory. If the video card 1900 XTX is combined with an X, accordingly performance is lost, which is why with the planned purchase of a dual-GPU system the cheaper X 1900 sufficient XT. Speaking, that is at a suggested retail price of $ 599 (549 dollars at a conventional X 1900 XT) and thus about $50 below the level of an individual X 1900 XTX. The video card should be available like all other models of the X 1900 series now available in trade. Also on the X 1900 CF a 512 MB large VRAM with an access time of 1.26 ns is used. To save power in 2D mode in addition, added to the Radeon X 1900 CrossFire Edition Download 594 MHz and 500 MHz, respectively.

On the bracket the same outlets such as with an X are 1800 XT CrossFire Edition, say, a normal, dual-link-enabled DVI output and the CrossFire connection, where the CF cable with the slave card must be connected. The compositing engine, which blended the images of the second video card, remains identical 1800 XT CF Edition for the X. You can find a detailed explanation in the article Club3D Radeon X 1800 XT CrossFire Edition.

Test system

Test system:

• Processor
  • AMD Athlon 64 4000 + (San Diego core, 90 nm, SSE3, 1024 kB level 2 cache)
• Motherboard
  • Gigabyte GA-K8NXP-SLI (nVidia nForce 4 SLI, Socket 939)
  • Sapphire pure CrossFire PC-A9RD480 (ATi RD480, Socket 939) for X 1800 CF
  • ASUS A8R-MVP (ATi RD480, Socket 939) X 1900 CF
• Memory
  • 2 x 512 MB crucial BallistiX DDR400 (2-2-2-5)
• Graphics cards
  • ATi Radeon X 1900 XTX (650 / 775)
  • ATi Radeon X 1900 CrossFire Edition (625 / 725)
  • ATi Radeon X 1900 XT (625 / 725) (simulates through down cycles of the Radeon X 1900 XTX)
  • ATi Radeon X 1800 XT CrossFire Edition (621 / 720)
  • ATi Radeon X 1800 XT (625 / 750)
  • ATi Radeon X 1800 XL (500 / 500)
  • ATi Radeon X 1600 XT (590 / 690)
  • ATi Radeon X 1300 Pro (600 / 400)
  • ATi Radeon X 850 XT PE (540 / 590) (simulates by overclocking the Radeon X 850 XT)
  • ATi Radeon X 850 XT (520 / 540)
  • ATi Radeon X 850 Pro (507 / 520)
  • ATi Radeon X 800 XL (400 / 490)
  • ATi Radeon X 800 GTO (400 / 490)
  • ATi Radeon X 800 GT (475 / 490)
  • ATi Radeon X 800 (400 / 350)
  • ATi Radeon X 700 Pro (425 / 430)
  • nVidia GeForce 7800 GTX 512 (550 / 850)
  • nVidia GeForce 7800 GTX (430 / 600)
  • nVidia GeForce 7800 GT (400 / 500)
  • nVidia GeForce 6800 Ultra (425 / 550)
  • nVidia GeForce 6800 GT (350 / 500) (simulates through down cycles of the GeForce 6800 Ultra)
  • nVidia GeForce 6800 GS (425 / 500)
  • nVidia GeForce 6800 (256 MB) (350 / 300)
  • nVidia GeForce 6600 GT (500 / 500)
• Peripherals
  • AOpen AAP-1648Pro-DVD drive
  • Samsung S-ATA 2 HDD 200 GB disk space (NCQ enabled)
• Driver versions
  • nVidia ForceWare 78.03
  • nVidia ForceWare 82.12 (GeForce 7800 GT, GeForce 7800 GTX, GeForce 7800 GTX 512)
  • ATi catalyst 5.8
  • ATi catalyst 5.13 (Radeon X 1600 XT, Radeon X 1800 XL, Radeon X 1800 XT)
  • Sample_8-203-3-060104a-029367E (Radeon X 1900 launch driver)
• Software
  • Microsoft Windows XP Professional SP2
  • Microsoft DirectX 9.0 c

Benchmarks

The following benchmarks were used during our test:

• Synthetic benchmarks:
  • 3DMark05 version 1.2.0
  • 3DMark06 version 1.0.2
  • Aquamark 3
• Game benchmarks:
  • Age of Empires 3
  • Far cry version 1.33
  • Splinter Cell: Chaos theory
  • Fear
  • Serious Sam 2 demo
  • Doom 3
  • The Chronicles of Riddick
  • Call of duty 2
  • Battlefield 2
  • Quake 4
  • Half-life 2: Lost coast

All benchmarks run with maximum details, so that the graphics card will be charged as high. As settings we chose this 1280 x 1024 and 1600 x 1200. So we pay tribute to the modern high end Accelerator, make the resolutions by their computing power lower than 1280 x 1024 CPU limited. In addition to the pure resolutions we can go through the benchmark course with two anti-aliasing and 16-fachen anisotropic filter, since these are frequently occurring quality settings.

Attention:After careful consideration and multiple analysis of even recorded game sequences, we have come to the conclusion in the ForceWare driver for nVidia cards increase the quality settings on high quality, because we can effectively combat the texture flickering only with this setting. In addition, this mode is similar to the setting “catalyst a.i. standard” on the ATi counterparts, resulting in largely a tie in image quality.

Driver settings: nVidia video cards

• Performance: High quality
• Vertical sync: off
• Force MipMaps: no
• Trilinear optimization: from
• Anisotropic mip filter optimization: from
• Optimization of anisotropic pattern: from
• Negative LOD bias: clamp
• Gamma adjusted AA (G70): A
• Transparency AA (G70): Off

Driver settings: ATi video cards

• Catalyst a.i.: Standard
• Mipmap detail level: high quality
• Wait for vertical refresh: always off
• Adaptive Anit-aliasing: off
• High quality AF: off
• Truform: Always off

Chip efficiency part 1

The theory is all very well, just what uses the higher Shaderleistung one, if the GPU can transform them into a higher performance? Nothing! For this reason, we have the Radeon X 1900 XTX clock rates a Radeon X 1800 XT (625 / 750 MHz) lowered and the R580-against the R520 GPU seek accordingly clock normalized compete. 1600 X 1200 and 1600 x 1200 with two anti-aliasing, and 16-time anisotropic filtering is used as a quality settings as high capacity so the two ATi graphics cards.

Chip efficiency part 2

As we can see clearly on the charts, the triple number of shader units in most games have a significant speed increase – 25 percent profit is not uncommon. In addition to the two 3DMark-Programmen from Futuremark, Battlefield 2 is a big winner, which apparently puts on more Shaderprogramme, as one might first think. The Radeon X can win 1900 1800 a huge advantage is compared to the X in age of Empires 3. This is the new feature “Fetch4″ to explain, because so the complete shadows no longer must be computed in the pixel shader. How otherwise was to be expected, can obtain the shaderlastige also fear of the 48 beer a profit and similar to it is taken also half-life 2: lost coast. Also you slowly set to 1900-series to several percentage points in the year of killed Mr Riddick and splinter cell 3 with the X. Logically, texture-based benchmarks such as, for example, Aquamark 3 can not benefit from the improvements in the R580 chip.

Theoretical benchmarks

FillRate Tester

• This useful little program is used to measure the fill rates of a graphics card. In contrast to the computer or fill rates tests integrated in the 3DMark that primarily measure the bandwidth in the case of single-texturing, this program can be quite differentiated information on different types of fill rate, including the PixelShader fill rates, which we want to look at.
  Because the shader used are sometimes quite short and bandwidth intensive, we have increased the resolution as far as possible to shift the focus more on the fill rate. As several mathematical operations per pixel are necessary, the fill rate will be charged more by increasing the resolution as the bandwidth.
  Refreshrate has been tested in 1600 x 1200 in 32-bit with 24 bit Z and 8 bit buffer and 60 Hz.

• Download: Fillrate Tester

VillageMark

• The VillageMark was developed by PowerVR and was used to illustrate the benefits of KYRO, because in that benchmark of Overdraw by a factor of up to 10 is particularly large 2. Many, especially older graphics cards, calculate also the surfaces that are hidden by others and mean therefore only wasted bandwidth and fill rate, so that this benchmark graphically actually not very expensive but more often than you first thought is to a stumbling block. That is why it is of the utmost importance in this benchmark, a well-functioning technique to remove hidden surfaces hidden surface removal (HSR) own.
  Has been tested using the following command line: [InstallDir] D3DVillagemark.exe benchmark = 1 – width = xxxx-height = xxxx-bpp = 32 ”

• More information: PowerVR.com
• Download: PowerVR.com

Fablemark

• The Fablemark was developed as the subsequent Templemark of PowerVR and serves despite a very high share in Overdraw of the display of the strengths of the KYRO chip regarding the stencil buffer.
  Of course the performance of the stencil calls also on all other cards strongly so that this test can be an indication of upcoming games, insert a Z / stencil-only pass prior to actual rendering, to prevent advance any Overdraw.
  Has been tested using the following command line: [InstallDir] D3DFablemark.exe benchmark = 1 – width = xxxx-height = xxxx-bpp = 32 ”

• More information: PowerVR.com
• Download: PowerVR.com

Templemark

• The Templemark is from PowerVR been just like the previous Fablemark, originally a demonstration programme. Because it supports a lot of current features, such as Hardware TnL, bump mapping, and up to six text urlagen in one pass however, he is the guaranteed not to NVIDIA also well suited as an independent benchmark, or ATi chips is optimized.
  “Has been tested using the following command line: [InstallDir] empledemov1-0-6.exe – benchmark”

• More information: PowerVR.com
• Download: PowerVR.com