iXBT Labs - Computer Hardware in Detail

Platform

Video

Multimedia

Mobile

Other

NVIDIA GeForce FX 5900 Ultra 256MB Video Card Review

May 14, 2003






CONTENTS

  1. General information
  2. Features of the NVIDIA GeForce FX 5900 Ultra 256MB card 
  3. Testbed configuration, drivers settings
  4. Test results: briefly on 2D 
  5. RightMark3D synthetic tests: philosophy and description of the tests 
  6. Test results: RightMark3D: Pixel Filling 
  7. Test results: RightMark3D: Geometry Processing Speed 
  8. Test results: RightMark3D: Hidden Surface Removal 
  9. Test results: RightMark3D: Pixel Shading 
  10. Test results: RightMark3D: Point Sprites 
  11. Test results: 3DMark2001 SE and 3DMark03 synthetic tests 
  12. Summary on the synthetic tests
  13. Test results: 3DMark2001 SE: Game1 
  14. Test results: 3DMark2001 SE: Game2 
  15. Test results: 3DMark2001 SE: Game3 
  16. Test results: 3DMark2001 SE: Game4 
  17. Test results: 3DMark03: Game1 
  18. Test results: 3DMark03: Game2 
  19. Test results: 3DMark03: Game3 
  20. Test results: 3DMark03: Game4 
  21. Test results: Quake3 ARENA 
  22. Test results: Serious Sam: The Second Encounter 
  23. Test results: Return to Castle Wolfenstein 
  24. Test results: Code Creatures DEMO 
  25. Test results: Unreal Tournament 2003 DEMO 
  26. Test results: AquaMark 
  27. Test results: RightMark 3D 
  28. Test results: DOOM III Alpha version 
  29. 3D quality 
  30. Conclusion

General Information

It's too banal to start a new review with  the words like the spring has come and new accelerators are making their way in life. It's clear that every new product is followed by dozens of beautiful words, lofty matters of 3D functionality and screenshots from perfectly looking games... from the future. 




Actually, I'm very sorry for those who spent their money on the GeForce FX 5800 Ultra, which is often priced over $500. NVIDIA let them down again. Or not? The NV35 is much spoken about, but the company didn't promised it officially, and they didn't mention that it wasn't worth buying the NV30 since the improved versions were expected soon. 

So, when the GeForce FX 5800/Ultra based cards promised yet before the New Year appeared on the market, the company made one more announcement. In the speech of NVIDIA's President a week ago he promised the release of the NV35 and called the NV30 their mistake. But what about those who have already bought this poor product? No one would replace it with the NV35 for nothing. That is why I feel very sorry for those who were ensnared by the elf girl. 

The NV35 represents a line again. The rumor has it that the line will have three cards, but we have trustworthy information only about two cards: 

  • GeForce FX 5900 Ultra - 450 MHz chip, 256 MB 425 MHz (DDR 850) 256bit local memory bus ($499); 
  • GeForce FX 5900 - 450 MHz chip, 128 MB 425 MHz (DDR 850) 256bit local memory bus ($399); 
  • GeForce FX 5??? - 425? MHz chip, 128? MB 400? MHz (DDR 800?) 256bit local memory bus ($299); 



The specs of the GF FX 5900 can change, but still, the GeForce FX 5900 Ultra, earlier codenamed NV35, is NVIDIA's fastest accelerator. 

The number 5900 looks similar to 5800, and it makes me think that there are not many changes done. But it's not simply overclocked judging by the frequencies, - moreover, the clock speeds are even lower than those of the NV30. Is there only a 256bit bus added? Well, we'll see that. 




One thing is clear - the NV35 is not simply a new line dethroning the previous products, - it replaces the NV30! This is the key difference from the previous tandems: NV10-NV15 (GeForce256-GeForce2), NV20-NV25 (GeForce3-GeForce4Ti). That time NVIDIA also aimed to replace the GeForce256 and GeForce3, but the process was gradual and delivery and sales were normal. Now its the first time when a new product is announced in a month. Certainly, the manufacturers were well aware of it, and the cards were in very short supply. Hence the high costs of the GF FX 5800/Ultra. Here is the list of the reviews of the GeForce FX based cards where we discussed all these subjects: 

Specification

  1. 0.13 micron fabrication process, copper connections
  2. 135 M transistors 
  3. 3 geometrical processors (each exceeds the specs of the DX9 VS 2.0) 
  4. 4 enhanced pipelined pixel processors. Their functionality markedly exceeds the specs of the DirectX9 PS 2.0, they are twice as powerful as the shaders in NV30, and floating-point operations have the true 128bit precision. 
  5. AGP 3.0 (8x) system interface 
  6. DDR2 256 bit local memory interface 
  7. Intellisample HCT (High Compression Technology), is the extension of the Intellisample supported in the NV30, local memory optimization technology. 
  8. Tile optimization: caching, compression and Early HSR, Early z Cull 
  9. UltraShadow technology accelerates rendering of stencil shadows (the stencil buffer is used a lot for their generation) by limiting depth of objects in the scene. 
  10. Support of precise integer-valued formats (10/16 bit per component) and precise floating-point formats (16 and 32 bits per component) for the textures and frame buffer. 
  11. Through accuracy of all operations - 32 bit floating-point arithmetic (128bit color depth supported) 
  12. Being activated, the new algorithm of optimized anisotropic filtering reduces the performance drop (fps) without significant quality degradation 
  13. Anisotropic quality up to 8x of the usual bilinear texture, i.e. up to 128 discrete samples per one texture fetch 
  14. Hybrid AA modes - 8x and 6xS 
  15. Frame buffer compression markedly reduces the performance drop with FSAA enabled
  16. Two integrated 10 RAMDACs 400 MHz 
  17. Integrated interface for external TV-Out chip 
  18. Three TDMS channels on chip for external DVI interface chips

Now let's have a closer look at the list items. 




  1. 135 M transistors on chip: The number of transistors in the chip is not lifted much, that is why the changes in the chip based on the NV30's architecture are slight. 
  2. 256bit local memory interface: This is the most expected new feature in the NV35. Since the NV30 release the enthusiasts of the 3D graphics couldn't imagine the NV35 without the 256bit local memory interface when they predicted its specs. 
  3. 4 enhanced pipelined pixel processors: Instead of 8 pixel processors expected NVIDIA left 4. As before, each processor features two texture filtering units, two integer ALUs (for fixed-point numbers) and one ALU for floating-point operations. Such configuration allows for up to 12 pixel operations at a clock. It's declared that the new chip is twice as powerful as the NV30 regarding pixel processors when they execute DirectX9 PS 2.0 with 128bit precision. 
  4. Intellisample HCT technology: local memory optimization technology: frame buffer lossless compression including color information and information from the Z buffer. Compression ratio up to 4:1. The Intellisample technology was improved after the NV30 - the technology determines whether compression of the data blocks transferred will be effective, but now the probability of data compression is higher. The Intellisample HCT also covers the following techniques that boost the local memory performance: local memory controller with switch, texture caches, z-culling, and fast z-clear. With all these optimizations the performance doesn't fall so much with the FSAA enabled. 
  5. UltraShadow technology: This technology accelerates rendering of stencil buffer shadows (this technique is used in the Doom III). It's declared that the NV30/NV35 chips can double the fillrate providing up to 8 pixels per clock, if only the Z buffer and/or stencil buffer is filled, against 4 pixels in case of ordinary rendering. The other interesting feature is the depth range an object can cast its shadow in. Look at the picture: 

  6.  





    If the pixel depth value does not fall into the range specified when shadows are rendered, the stencil buffer for this pixel is not updated. It's possible thus to save much on the fillrate. It seems that the unit controlling early z-cull is improved, and now it compares the value stored in the Z buffer with the current value interpolated according to the triangle coordinates and with two additional values. As a result, the gain will be twice at least as compared to other chips of the same number of pipelines (8 x-buffer values against 4 color values) and four times (considering that early z-cull can cull up to 16 pixels per clock), even if we assume that the bus load is not lowered. 

    Below you can see how large the area taken by the shadow buffers rendered is. 










The technology, however, has downsides as well. 

  1. UltraShadow will be available only in the OpenGL with the extension "NV_depth_bounds_test" 
  2. NVIDIA is expecting the patent for the UltraShadow technology
  3. The technology can't be enabled with a couple of code lines (like it was with N-Patch aka Truform) - the programmers have to analyze a scene rendered and set an effective range for shadow rendering. 
  4. The technology is announced only for the NV35 by now; the extension in the current drivers is available also only for the NV35. It's not clear whether the extension will be available for the rest of the NV3x based cards, but it might be so because the extension "NV_depth_bounds_test" was available for the NV30 in the emulation mode in "Buzz/emulated" in the OpenGL driver dated yet August 2002. 

Card

The card has AGP X2/4/8 interface, 256 MB DDR SDRAM in 16(!) chips on both PCB sides. 
 








 
Hynix HY5DU283222-AF22 memory chips of BGA form-factor. Maximum clock speed is 450 (900) MHz, 2.2 ns access time. Memory clocked at 425 (850) MHz, GPU at 450 MHz by default.


We can see that 256 MB occupy twice more chips. They sum up to 512bits, i.e. if the chip and PCB had a hardware 512bit memory, the card could have such a high throughput. But there is a 256bit bus, that is why the interleave is used. The memory uses the two-bank scheme, which raises the exchange rate a little. That is why we can affirm that the 128MB GeForce FX 5900 card will be a little less efficient than the 256MB models, just like the RADEON 8500 64 and 128 MB. 
 

NVIDIA GeForce FX 5900 Ultra 256MB






NVIDIA GeForce FX 5800 Ultra 128MB












The card is very long and almost equal to the PCB of the Voodoo5 5500'. The chip arrangement reminds the Matrox Parhelia: 




The PCB is quite sophisticated, but there are only 8 layers plus shielding (against 12 of the NV30): 





The cooler consists of two parts: for the processor and for the memory (a single cooler can be dangerous for less heating elements as it can carry heat to them from the hotter components). The central cooler is a copy of the cooler on the Quadro FX or GeForce FX 5800 (non-Ultra), with the heatsink lid removed for quieter operation. Like on the GeForce FX 5800-non-Ultra, the cooler rotates at a low frequency switching to a higher one when necessary (when the temperature exceeds a certain value). 

The memory' cooler consists of two massive aluminum heatsinks painted black. Heat is transferred through special thermal elements attached to the heatsinks, which also ensure a good contact between the chips and heatsinks. The cooler's mount was carefully worked out - there are no any plastic clips but only spring bolts. 











 

The card heats less than the NV30. You can't touch the GeForce FX 5800 Ultra (even the PCB) in an hour of operation in 3D, while this one is just a little hotter than the pain barrier. 
 
 

One side of the card is used entirely for the power supply system, - there is also an external power supply connector. 




As for the TV-out, the card incorporates the Philips 7108 codec that potentially supports the VIVO. 




See the review by Aleksei Samsonov and Dmitry Dorofeev on operation of TV-out with such codec. 

Overclocking

Overclocking is locked in the drivers: you can change the frequencies but the new settings will be changed back then. Low-level overclocking is not possible either because the latest versions of the RivaTuner do not support the NV35. 

Installation and drivers

Testbed: 

  • Pentium 4 based computer: 
    • Intel Pentium 4 3066 CPU (HT=ON); 
    • ASUS P4G8X (iE7205) mainboard; 
    • 1024 MB DDR SDRAM; 
    • Seagate Barracuda IV 40GB HDD; 
    • Windows XP SP1; 
    • ViewSonic P810 (21") and ViewSonic P817 (21"). 

NVIDIA drivers v44.03, VSync off, texture compression off in applications. DirectX 9.0a installed. 

Cards used for comparison: 

  • Reference card NVIDIA GeForce FX 5800 Ultra (500/500 (1000) MHz, 128 MB); 
  • ATI RADEON 9800 PRO (380/340 (680) MHz, 128 MB, driver 6.307). 

Drivers Settings

 



















The IntelliSample's modes are renamed again: Quality instead of ApplicATIon, Performance instead of Balanced, High Performance instead of Aggressive and Performance. AA 8x was added (not 8xS only for D3D, but supporting both D3D and  OGL). 

All the other parameters are the same as before. 

Test Results

2D graphics

In spite of the 400 MHz RAMDAC and other quality components, 2D quality can depend on everything but  the specs. Moreover, it can change from sample to sample, even if they come from the same factory. Besides, monitors do not work equally with all videocards. That is why our estimation of 2D is subjective. 

The 2D quality was tested on the ViewSonic P817-E monitor and Bargo BNC cable. In my opinion, the card has excellent 2D quality! No visual problems at all in 1600x1200 at 85Hz and in 1280x1024 at 120Hz. 

RightMark 3D (DirectX 9) Synthetic Tests

The test suite from the RightMark 3D (which is under development now) includes the following synthetic tests at this moment: 

  1. Pixel Filling Test; 
  2. Geometry Processing Speed Test; 
  3. Hidden Surface Removal Test; 
  4. Pixel Shader Test; 
  5. Point Sprites Test. 

The philosophy of these synthetic tests and their description are given in the NV30 Review.

Those who are eager to try RightMark 3D synthetic tests can download the "command-line" test versions which record the final XLS file in the XML format accepted in Microsoft Office XP: 

Each archive contains description of the test parameters and an example of a .bat file used for benchmarking accelerators. We welcome your comments and ideas as well as information on errors or improper operation of the tests. 

The first beta COMPLEX version of the packet of the above RightMark 3D synthetic tests is available at http://www.rightmark3d.org/d3drmsyn/. This site is entirely dedicated to the RightMark 3D packet. You can also find there the beta version of the RightMark Video Analyzer v0.4 (14.8Mb) packet we are currently using for the video cards tests. 

Mailto: unclesam@ixbt.com

Practical Tests

The accelerators we are going to test now are meant for enthusiasts in opinion of the manufacturers: NVIDIA GeForce FX 5900 Ultra (today's hero), NVIDIA GeForce FX 5800 Ultra and ATI RADEON 9800 PRO. The NVIDIA GeForce FX 5800 Ultra which has the closest architecture to the GeForce FX 5900 Ultra is used as a reference model. The scores of the ATI RADEON 9800 PRO are given to find out whether NVIDIA is able to outpace ATI's flagship. 

Pixel Filling

  1. The test measures a frame buffer fillrate (Pixel Fillrate). Constant color, no texture sampling. Scores are given in million pixels per second for different resolutions in the standard mode and in 4x MSAA: 

  2.  




    The fillrate in the AA mode corresponds to the frequency difference between GeForce FX 5900 Ultra and NVIDIA GeForce FX 5800 Ultra (450 vs 500). In spite of the higher clock speed of the GFFX chips, the GeForce FX 5900 Ultra loses to ATI in all cases. For the GFFX chips the bus throughput is not a limiting factor in case of ordinary filling. Radeon 9800 Pro takes the lead at the expense of 8 pixel pipelines able to record up to 8 color and depth values per clock, and the maximum values demonstrated by ATI are limited exactly by this throughput. FX can record only 4 full pixels (color + depth + Stencil buffer when needed). But FX's 4 pixel processors have an interesting capability - if when the shader is processed or a triangle is filled, we do not save the pixel value but change only the Z or Stencil buffer values, the pixel processor can have two results per clock. That is, it can record 8 values of the Z or Stencil buffers per clock. NVIDIA openly declares this advantage of the GeForce FX 5900 Ultra NVIDIA as a part of the UltraShadow technology. Such optimization will be of much help in games with stencil shadows, like DOOM III. It can accelerate scene rendering almost 1.5 times. But our test deals with color values as well. That is why the results demonstrate 4 pixels rendered per clock. 

    In the AA mode the fillrate changes. As compared to the GeForce FX 5800 Ultra the 5900 card goes far ahead, and the ATI's card can fight only in low resolutions. In higher resolutions the GeForce FX 5900 Ultra is an unexampled leader in spite of twice fewer pixel pipelines. 
     

  3. Frame buffer fillrate with simultaneous texturing. One simple bilinear texture sampling added. We will estimate how the competitive read stream from memory reduces the rendering effectiveness. Million pixels per second for different resolutions in the standard mode and at 4x MSAA. Since all latest ATI Catalyst drivers (3.1 and higher) together with our synthetic test provide quite strange results on some ATI cards (lower than expected probably because of the failed attempt to optimize the test in the texture sampling modes), the Pixel Filling results with textures are not given for the RADEON. 

  4.  



    In general the situation is the same, but the peak values are lower. Let's see whether the reality goes along with the theoretical limits based on the core clock speed and number of pipelines: 
     

    Product Theoretical maximum, M texels/sec Measured maximum (no texture), M texels/sec Measured maximum (one texture), M texels/sec
    RADEON 9800 PRO 3000 2486 -
    GeForce FX 5900 Ultra 1800 1757 1739
    GeForce FX 5800 Ultra 2000 1957 1848

Geometry Processing Speed

The results are sorted according to the complexity degree of the light model used. The lower group represents the simplest variant which corresponds to the accelerator's peak vertex throughput. 

  1. Fixed TCL performance (for NV30 and R300 - performance of the shader that emulates it): 



  2. The scores' difference is equal to the frequency difference - the GeForce FX 5900 Ultra is 10% slower than the GeForce FX 5800 Ultra (450 vs 500 MHz) 

  3. Vertex Shaders 1.1: 



  4. In the simplest case (Shader Level 1) the GeForce FX 5800 Ultra card has much lower results than when the fixed TCL settings, functionally equal, are used. The GeForce FX 5900 Ultra has comparable results with the fixed TCL and VS 1.1, - since the driver version is the same, the microcode couldn't be optimized. In case of the more complex shader the difference makes 10% again. 

  5. Shaders 2.0 with loops: 



  6. Again, the difference between NV30 and NV35 is equal to the frequency difference. In case of compilation of the Shader Level 1 for VS 2.0 the NV30 has the true results as compared to the lowered ones in case of the VS 1.1. 

    The ATI RADEON 9800 PRO processes VS 2.0 with loops as efficient as VS 1.1 though earlier the Radeon 9700 Pro was less speedy. Probably, the driver unwinds the loops into one big shader program when compiling the shader into a microcode. 

    The NV3x series chips have to emulate loops with dynamic transitions with the microcode developed for dynamic execution (dependent on data of the current vertex). That is why the loop overheads are considerable for the NV3x! Well, the instruction flow dynamic management must cause noticeable delays in loop execution. Speed is the cost of flexibility. 

    But if it's true that the ATI RADEON 9800 PRO unwinds shaders, this optimization has the reverse side. A microcode of such shader will take much more space in comparison with loops, the driver will be able to store fewer shaders in the chip's cache, and time taken for loading a new shader into the chip can bring to naught all the optimizations. If the constant that sets the number of loops is changed, the driver has to regenerate and reload the shader's microcode. 

  7. Cross dependence on geometry detail level and shader's complexity: 






  8. When the detail level is low (i.e. the geometrical unit is less loaded as compared to the pixel pipelines) the GeForce FX 5900 Ultra outdoes the 5800 Ultra probably due to the much higher bus throughput and the optimized cache of the frame buffer. The 5800 Ultra takes the lead when the geometrical unit gets more load. 

Hidden Surface Removal

  1. HSR support and maximum performance (%) vs. resolution and number of triangles, no-texture scene (early Z cull is not accounted for): 



  2. The HSR is less effective on the NV35 because of the higher bus throughout as compared to the NV30 because even in the worst case (back to front sorting) the bus is able to pump through the data volume increased. But the R300 has the best HSR effectiveness anyway since it uses the hierarchical structure and often makes clipping at higher levels, while the NV30 has only one decision-making level combined with tiles used for depth information compression. In 1600x1200 the HSR becomes much less effective on the R300 probably because of the limited cache of the hierarchical buffer which is not used in this resolution, and decisions are made only at the lowest level combined in the compressed units in the depth buffer, like in the NV3x. 

    HSR effectiveness vs. scene complexity




    In the NV3x with only one tile level, the fewer the polygons - the more effective the HSR. R300 keeps to the golden mean in this aspect. Its HSR has just started spreading its wings on the scenes of average complexity. 

  3. HSR support and maximum performance (%) vs. resolution and number of triangles, textures on, early Z cull: 






  4. All the chips have more effective operation thanks to the early Z cull. With the textures accounted for, all the chips prefer scenes with fewer polygons. In all other respects, the situation is similar to the above one. 

  5. HSR for unsorted and sorted scenes, with and without textures: 






  6. Correlation of the HSR effectiveness of the different chips is the same, but the gain with the HSR is lower. But anyway, the gain is more than twice when the textures are used. 

So, even if the scene is originally unsorted, the gain is great. If you want to benefit from the HSR you should sort scenes before rendering. The speed will rise several times! 

Pixel Shading

  1. Pixel shader performance of NV35 vs. NV30. 



  2. The test using PS 2.0 demonstrates queer results for the NVIDIA's chips, especially because the pixel shaders of DirectX9 are twice as powerful in the NV35. Taking into account that the clock speed of the GeForce FX 5900 Ultra is 10% lower we have the following performance gap at the same frequency in different tests: 
     

    Test 1 3 4 5 7
    GeForce FX 5900 Ultra / 5800 Ultra +15% -19% 0% -32% +46%

    If you look into the shader code you will see that shaders 1 and 7 (as well as 6) use much more texture samples (moreover, shaders 6 and 7 also sample data out of 3D textures) relative to the arithmetic shader commands in comparison with the other tests. So, the performance grows due to the higher bus throughput and improved caching algorithms. 

    But what about the performance drop in other tests if the growth is announced to be twice! The matter is that the GeForce FX 5900 Ultra executes floating-point operations in PS 2.0 with the true 32bit precision (128 bits per register), contrary to the GeForce FX 5800 Ultra where the precision is limited by 16 bits (64 bits per register) in the driver. Later we will test more carefully the GeForce FX 5900 Ultra chip in PS 2.0. 

    As to the Radeon 9800 Pro, it remains a leader. 

Point Sprites

    1. Lighting on/off, size dependence: 






    As expected, lighting influences only small sprites; as they grow the performance gets limited by the fillrate (approximately at the size of 8). So, for rendering systems comprising of a large number of particles the size should be less than 8. By the way, size 8 is optimal for the NVIDIA's chips as they outedge the Radeon 9800 Pro. This is probably connected with the size of tiles cached in NVIDIA. The GeForce FX 5900 Ultra also works better with the frame buffer: with light (i.e. the load of the geometrical unit is increased) over 500 the Ultra goes on a par, it takes the lead when the load is mostly on the frame buffer unit. 

    Still, the Radeon 9800 Pro is ahead in most modes thanks to 8 pixel pipelines rendering 1 pixel per clock with one texture applied. 

3D graphics, 3DMark2001 SE, 3DMark03 synthetic tests

All measurements in 3D are taken at 32bit color depth. 

Fillrate in 3DMark2001 SE

 




The GeForce FX 5900 Ultra performs better in the AA mode, though the Radeon 9800 is still more efficient in spite of its lower bus bandwidth. Contrary to RightMark 3D, in 3Dmark 2001 the GeForce FX 5900 Ultra outruns the 5800 Ultra even without the AA since the 3DMark 2001 needs a higher bus throughput. 

Multitexturing: 




In case of multitexturing the NVIDIA chips have a higher speed than the Radeon 9800 Pro. As the number of textures grows up, their sampling and filtering are getting more important, and they mostly depend on the core clock speed, rather than on the number of pixel pipelines or bus bandwidth. In this case the chips have 8 texture units each, and the chip's clock speed becomes the determining factor. That's why the best one is NV30 (500MHz), then goes NV35 (450 MHz) and then R350 (380MHz). 

In the AA mode the performance of the NV3x chips falls down more than that of the R300 in spite of the frame buffer compression in MSAA in all chips and greater memory bandwidth of the NV35. 

Fillrate in 3DMark2003

 




The picture is similar to the 3DMark 2001; the absolute values are a little lower - the test is less precise, and the figures are farther from the limiting values. 

Multitexturing: 




Similar to the 3DMark 2001. 

Pixel Shader 2.0

Simple variant: 




In this test the GeForce FX 5900 Ultra outdoes the 5800 Ultra, and in higher resolutions it outpaces the Radeon 9800 Pro. This test is functionally close to the RightMark 3D, Pixel Shading in Shader Level 6 and 7 which also use procedure textures. That is probably why it much depends on the texture sampling speed, and the GeForce FX 5900 Ultra runs faster than its predecessor. 

If we keep on comparing to the Shader Level 6 and 7 of RightMark 3D we will see that in RightMark 3D the Radeon 9800 Pro beats both NVIDIA chips, while there is no such an advantage in the 3DMark03. Well, optimization is obvious, especially if you look at the GeForce FX 5800 Ultra with the driver v43.45. 

Vertex Shaders 1.1

 




Although the VS test used in the 3DMark03 differs a lot from the RightMark3D, the scores are similar: GeForce FX 5900 Ultra loses about 10 % to the 5800 Ultra, and the Radeon 9800 Pro keeps on the same level with the NVIDIA's chips. 

Anisotropic Filtering

The anisotropy algorithms haven't changed. There are the same three modes mentioned above (see the driver settings). A bit later the quality section will demonstrate the screenshots. 

  1. Quality - corresponds to the parameters set by a user or an application. The algorithms are similar to those used since the GeForce 3 
  2. Performance - the driver tries to ascertain the filtering algorithms optimal for quality. You can look at the examples in the Aleksei Nikolaichuk's review. The GeForce FX also uses the anisotropic filtering that needs fewer texture samples for most pixels. The trilinear algorithm is altered as well - the NV3x can work in the mode when sampling and filtering from two MIP levels are supported only for the surface part closest to the transition between the MIP levels. For the other part of the surface sampling and filtering are made for only one MIP level, which makes the texture filtering speed rise twice in this place. 
  3. High Performance - the anisotropic degree is lower than in the Performance mode, and the trilinear filtering is simplified as well. 

Summary on synthetic tests

Let's sum up the estimation of various units of NV35 in the synthetic tests. 

  1. The NV35 gains mostly from the higher memory throughput and from the redesigned and optimized memory controller and chip units controlling the frame compression and, probably early z-cull due to the UltraShadow technology. The higher memory throughput told upon the tests in AA modes. In the Fillrate tests the gain runs into dozens of percents if not hundreds. 
  2. No considerable changes in the geometry processing unit. 
  3. Pixel processors are also optimized; now pixel shaders using floating-point operations with 32-bit precision (128 bits per register) have the speed comparable to 16bit precision operations in NV30. However, the Radeon 9800 Pro is still faster is processing PS 2.0. 

3D graphics, 3DMark2001 game tests

Anisotropy was set to 16x for ATI's cards and to 8x for NVIDIA because algorithms of this function considerably differ (we discussed it in NV30 Review). The criterion is just one: maximum quality. The screenshots were shown several times earlier. Besides, it's interesting to compare NVIDIA's different anisotropy modes (Application, Balanced, Aggressive) with the ATI's high-quality mode; our readers can estimate how speed and quality correlate from the screenshots from the NV30 Review that demonstrates anisotropic quality. 
 
 

3DMark2001, 3DMARKS














3DMark2001, Game1 Low details














 

3DMark2001, Game2 Low details














 

3DMark2001, Game3 Low details














 

3DMark2001, Game4













  1. As expected, the 256bit memory bus helped a lot the NV35 in the AA modes. The memory bandwidth is higher (256 / 8 * 850) / (128 / 8 * 1000) = 27200 / 16000 = 1.7 times. It helped much the card win in the heavy modes with AA and anisotropy, though the GPU's speed is lower (450 against 500 MHz), as well as the memory's one (425 against 500 MHz). 
  2. The anisotropy algorithms are the same though the guys at NVIDIA keep on working on this function (read more about the same in the OpenGL in the Aleksei Nikolaichuk's review (AKA Unwinder)). By the way, they are working on the whole GeForce FX family, not just NV35; and with the latest drivers v44.03 the  NV30 and the NV35 sometimes have identical anisotropic efficiency. It's interesting that NVIDIA loads anisotropy mostly on the processor, and we could expect lower scores from the NV35 because of the lower GPU's clock speed, but the processor is well optimized for this function. 
  3. AA 8x, a hybrid of MSAA and SSAA, takes a huge part of the resources; its quality will be discussed later. 

3D graphics, 3DMark03 game tests

 

 
 


3DMark03, 3DMARKS













3DMark03, Game1

Wings of Fury: 

  • DirectX 7.0; approx. 32000 polygons on the scene, 16 MB memory used for textures, 6 MB for buffers for vertices and 1 MB for indices. 
  • All geometrical operations are based on Vertex Shaders 1.1 which can be emulated with CPU (if there is no hardware support). 
  • All planes have 4 texture layers, that is why accelerators able to process 4 textures in a pass will benefit. 
  • Fire and tail effects are made with the point sprite and other techniques. 












3DMark03, Game2

Battle of Proxycon: 

  • DirectX 8.1; Approx. 250 000 polygons on the scene with Pixel Shaders 1.1 (and 150 000 polygons on the scene with Shaders 1.4), 80 MB memory used for textures, 6 MB for buffers for vertices and 1 MB for indices.
  • All geometrical operations are based on Vertex Shaders 1.1 which can be emulated with CPU (if there is no hardware support). 
  • All heroes are "dressed" with the vertex shaders as well. 
  • Some light sources made dynamic shadows with a stencil buffer. 
  • All pixel operations are carried out with shaders 1.1, and if possible with shaders 1.4. 
  • Calculation of per-pixel lighting for haze effects and other components. 
  • Accelerators supporting pixel shaders 1.1 use one pass for determining Z buffer, then 3 passes for each light source. If an accelerator supports shaders 1.4, it needs one pass for each light source.












3DMark03, Game3

Trolls' Lair: 

  • DirectX 8.1; approx. 560 000 polygons on the scene with Pixel Shaders 1.1 (and 280 000 polygons on the scene with Shaders 1.4), 64 MB memory used for textures, 19MB for buffers for vertices and 2 MB for indices. 
  • All geometrical operations are based on Vertex Shaders 1.1 which can be emulated via CPU (if there is no hardware support). 
  • All heroes are "dressed" with vertex shaders as well. 
  • Some light sources made dynamic shadows with a stencil buffer. 
  • All pixel operations are carried out with shaders 1.1, and if possible with shaders 1.4. 
  • Calculation of per-pixel lighting for haze effects and other components. 
  • Realism of the heroine's hair is achieved with physical models and anisotropic lighting. 












3DMark03, Game4

Mother Nature: 

  • DirectX 9.0; approx. 780 000 polygons on the scene, 50 MB memory used for textures, 54MB for buffers for vertices and 9 MB for indices. 
  • Every leaf is separately animated with Vertex Shaders 2.0. Grass is animated with vertex shaders 1.1. 
  • Lake's surface is formed with pixel shaders 2.0. 
  • Sky is made with pixel shaders 2.0, sun glints are formed with extra-precision calculations in DX9. 
  • Earth surface is made with shaders 1.4. 












  1. In this packet the anisotropy of the NV35 is sometimes slower than in the NV30 because of the difficult tests (or because of insufficient optimization of the drivers). The lower frequencies of the former make the effect. 
  2. But the 256bit bus helps the NV35 to win. However, this test is not entirely objective because both NVIDIA and ATI can make optimizations here. Thus, in the Pixel Shaders NVIDIA simply lifted up the speed of the NV30 four times with the drivers, while the PS in the NV35 are not that efficient. 

3D graphics, game tests

3D games used to estimate 3D performance: 

    • Return to Castle Wolfenstein (MultiPlayer) (id Software/Activision) - OpenGL, multitexturing, Checkpoint-demo, test settings - maximum, S3TC OFF, the configurations can be downloaded from here

    •  
    • Serious Sam: The Second Encounter v.1.05 (Croteam/GodGames) - OpenGL, multitexturing, Grand Cathedral demo, test settings: quality, S3TC OFF 

    •  
    • Quake3 Arena v.1.17 (id Software/Activision) - OpenGL, multitexturing, Quaver, test settings - maximum: detailing level - High, texture detailing level - #4, S3TC OFF, smoothness of curves is much increased through variables r_subdivisions "1" and r_lodCurveError "30000" (at default r_lodCurveError is 250 !), the configurations can be downloaded from here 
    • Unreal Tournament 2003 Demo (Digital Extreme/Epic Games) - Direct3D, Vertex Shaders, Hardware T&L, Dot3, cube texturing, default quality 

    • Code Creatures Benchmark Pro (CodeCult) test demonstrates operation of cards in the DirectX 8.1, Shaders, HW T&L. 

      AquaMark (Massive Development) test demonstrates operation of cards in the DirectX 8.1, Shaders, HW T&L. 
       

  • RightMark 3D v.0.4 (one of the game scenes) - DirectX 8.1, Dot3, cube texturing, shadow buffers, vertex and pixel shaders (1.1, 1.4). 

Quake3 Arena, Quaver

 













NVIDIA's promise to outscore the RADEON 9800 PRO in the modes with AA and anisotropy by 50-80% becomes a reality. The 256bit bus helps the NV35 to succeed in AA. The anisotropy is good as well: the frequencies are lower and the speed is higher (I mean NV35 and NV30). Remember the review of Aleksei Nikolaichuk. 

Serious Sam: The Second Encounter, Grand Cathedral

 













ATI optimized its drivers long ago, but NVIDIA ignores this popular benchmark to some reason. Well, the NV35 loses to the RADEON 9800 PRO in the no-AA and no-anisotropy modes (except 1024x768). But even AA can't help the NV35. It's only the fast anisotropy which can be a ring-buoy. 

Return to Castle Wolfenstein (Multiplayer), Checkpoint

 













After the driver version 42.82 the performance dropped much yet on the NV30. NVIDIA has done nothing so far to cure it! The NV30 could have had over 120 fps in 1600x1200! And now it has only 70! It's not just 8-10 fps! Obviously, the NV30 and NV35 work incorrectly in this game, the driver wrongly forms the rendering architecture of so flexible processor. It's strange that this game was developed on the base of the Quake3's engine together with id Software, the company for products of which the programmers at NVIDIA optimize the drivers for a long time already.  

Code Creatures








Unfortunately, the anisotropic filtering of the NV30/NV35 does not work in this test, and we failed to test these modes as well as the combined modes including AA and anisotropy. It's also NVIDIA's mistake, like in case of the Elder Scrolls III: Morrowind, where the AA was not supported (only the 4xS mode could work!). 

Unreal Tournament 2003 DEMO

 













Here everything is OK, though the advantage of the NV35 over its rival is not great. 

AquaMark

 














In this test the NV35 in the anisotropy mode works slower than even the NV30 (probably the drivers were not optimized for this test), the advantage with the AA is not great, and the overall breakaway of the NV35 is intangible. 

RightMark 3D

 













The drivers were not optimized for this test because the benchmark itself was not available on the Net. The pixel shaders are very important here, and the NV35 loses to the NV30 (the core clock speeds are again a determining factor). 

DOOM III Alpha

 






Since this test utilizes all innovations of the NV3x, in particular, the accelerated shadow processing, the results are obvious. A great amount of data pumped through the memory require a high throughput, that is why the 256bit bus has come in handy. 

3D graphics quality

Since the NV35 is actually a copy of the NV30 with some optimizations, they do not differ in games support. That is why I'm not going to shows again the screenshots again that demonstrate anisotropy quality in the IntelliSample modes as well as different AA levels. NVIDIA's and ATI's AA and anisotropy modes were already estimated closely in the GeForce FX 5800 Ultra Review (theory, tests), and in the GeForce FX 5600/5200 Review (practical anisotropy comparison). In the RADEON 9800 PRO review you can find some interesting facts as well. 

Now we will focus on some peculiarities appeared with the drivers v44.03. First of all, the trilinear filtering has new presets. If you remember, Quality-Performance-High Performance modes differ in the trilinear filtering optimization degree:
 

Presets Screenshots from Quake3
Example 1
Bilinear, Quality 


Trilinear, Quality 


Trilinear, Performance 


Trilinear, High Performance 



 
Example 2
Trilinear, Quality 


Trilinear, Performance 


Trilinear, High Performance 



 
Serious Sam: TSE
Trilinear, Quality 


Trilinear, Performance 


Trilinear, High Performance 



 

The Quality mode is flawless, the trilinear filtering is fully executed (4 samples are taken at each MIP level, go twice through bilinear filtering, then the pairs of values are mixed with respect to weighting factors of distance from the BORDER=1 between the levels), in the other two modes only a small part of the pixels endure trilinear filtering, the most remain untouched after the bilinear processing. It saves the GPU resources which can be used for anisotropy calculations. But it's only an assumption about the release of computational resources of the texture units with the trilinear filtering optimization which was expressed by my colleague from F-Center in the NV30 review. The High Performance also forcedly compresses textures, - you can look at the example 2 in Quake3 and in Serious Sam: TSE in the sky. 

What about the anisotropy? The quality is very similar to the NV30. Let's just look at the screenshots from 3DMark2001 with the different presets in the IntelliSample (anisotropy 8x everywhere). 


 
Game 1
Quality Performance High Performance









Animated GIF



Game 2
Quality Performance High Performance









Animated GIF
No difference between the modes! 
Game 3
Quality Performance High Performance









Animated GIF
No difference between the modes! 
Game 4
Quality Performance High Performance









Animated GIF




 

If the scene consists of a great deal of curved surfaces and colorful textures, there is almost no difference between the IntelliSample modes. If so, why to pay more by setting the Quality mode? 

Conclusion

So, what can we say about the fastest game accelerator NVIDIA GeForce FX 5900 Ultra? 

  1. If we sum up all highs and lows, the NV35 will be the fastest game graphics card despite the fact that it loses sometimes to the ATI's card. 
  2. The NV35 does not supplement NVIDIA's current lines but replaces the failed NV30. Even NVIDIA President admits it. That is why you should forget about the NV30 and regard the NV35 as the work on errors. The codename does not agree with what there was earlier: NV*5 used to be scheduled improvement of a previous generation that brought both error corrections and speed growth, as well as improvements in the architecture. Well, on one hand, the 256bit bus is a considerable enhancement, but on the other hand, the core has few changes. 
  3. Anyway, this card will be the forefront solution for mass production of cards (in High-End); the manufacturers being hungry for over a year are looking forward to any sign from NVIDIA. 
  4. I'm sorry for those who spent money on the NV30, but that card is also a fast runner and will serve for a long time. Its downside is noisy operation, while the NV35 is deprived of it. I want to specially thank NVIDIA for this improvement. 



 
 
Andrey Vorobiev (anvakams@ixbt.com)
 

Aleksei Barkovoi  (clootie@ixbt.com


Write a comment below. No registration needed!


Article navigation:



blog comments powered by Disqus

  Most Popular Reviews More    RSS  

AMD Phenom II X4 955, Phenom II X4 960T, Phenom II X6 1075T, and Intel Pentium G2120, Core i3-3220, Core i5-3330 Processors

Comparing old, cheap solutions from AMD with new, budget offerings from Intel.
February 1, 2013 · Processor Roundups

Inno3D GeForce GTX 670 iChill, Inno3D GeForce GTX 660 Ti Graphics Cards

A couple of mid-range adapters with original cooling systems.
January 30, 2013 · Video cards: NVIDIA GPUs

Creative Sound Blaster X-Fi Surround 5.1

An external X-Fi solution in tests.
September 9, 2008 · Sound Cards

AMD FX-8350 Processor

The first worthwhile Piledriver CPU.
September 11, 2012 · Processors: AMD

Consumed Power, Energy Consumption: Ivy Bridge vs. Sandy Bridge

Trying out the new method.
September 18, 2012 · Processors: Intel
  Latest Reviews More    RSS  

i3DSpeed, September 2013

Retested all graphics cards with the new drivers.
Oct 18, 2013 · 3Digests

i3DSpeed, August 2013

Added new benchmarks: BioShock Infinite and Metro: Last Light.
Sep 06, 2013 · 3Digests

i3DSpeed, July 2013

Added the test results of NVIDIA GeForce GTX 760 and AMD Radeon HD 7730.
Aug 05, 2013 · 3Digests

Gainward GeForce GTX 650 Ti BOOST 2GB Golden Sample Graphics Card

An excellent hybrid of GeForce GTX 650 Ti and GeForce GTX 660.
Jun 24, 2013 · Video cards: NVIDIA GPUs

i3DSpeed, May 2013

Added the test results of NVIDIA GeForce GTX 770/780.
Jun 03, 2013 · 3Digests
  Latest News More    RSS  

Platform  ·  Video  ·  Multimedia  ·  Mobile  ·  Other  ||  About us & Privacy policy  ·  Twitter  ·  Facebook


Copyright © Byrds Research & Publishing, Ltd., 1997–2011. All rights reserved.