Today, NVIDIA will be announcing the GeForce FX 5900 Ultra, which is based on the 0.13-micron NV35 code-named graphics processing unit. With delays and limited production of their initial high-end 0.13-micron part, which debuted as the GeForce FX 5800 Ultra (NV30), NVIDIA needed to quickly follow-up in order to strengthen their high-end graphics product line. While NVIDIA's dominant market share in the desktop PC market continues to rise, they had been losing ground in the areas of performance and image quality to long-time industry rival ATI.
NVIDIA's GeForce FX 5900 Ultra (NV35)
Retail boards might be similar in length to a GeForce4 Ti 4600
Outfitted with a 256-bit memory interface, the GeForce FX 5900 Ultra increases the amount of memory bandwidth, which results in significantly faster 3D graphics performance. Immediate benefits are realized at high resolutions, and when image quality enhancements such as antialiasing and anisotropic texture filtering are enabled. For those of you who have waited for NVIDIA to deliver the "next big thing", the NV35 is a breakthrough product and will provide the gaming enthusiast with unprecedented levels of image quality.
Other enhancements of the NV35 over the NV30 include:
CineFX 2.0 Engine - doubles the performance of floating-point pixel shader operations.
UltraShadow Technology - accelerates computations associated with shadows that originate from multiple light sources and interactions of those shadows with objects.
High Compression Technology (HCT) - up to 50% increase in the likelihood for successful compression of color, texture, and z data. Improvements in antialiasing techniques and anisotropic filtering algorithms.
Anisotropic Filtering - significant increase in performance when Quality (trilinear) is used.
Runs Quieter - the FlowFX cooling system has been eliminated.
UltraShadow Technology is designed to accurately render shadows by maintaining the interactions of multiple lights sources and game objects. UltraShadow performs optimally when used in conjunction with software and can enhance the performance of upcoming games such as Abducted and Doom 3.
An OpenGL extension that will become useful for increasing the performance of stencil shadows is NVIDIA's GL_NV_depth_bounds_test. This extension is similar in concept to occlusion culling as it rejects objects that don't need to be rendered. More specifically, the GL_NV_depth_bounds_test extension allows the developer to compare the frame buffer depth value against a specific range like the depth bounds that are marked in the above diagram.
In hardware, UltraShadow improves performance by rendering shadow volumes in a single pass and ensures that shadow edges are antialiased.
TEST SYSTEM SPECIFICATIONS
In this preview I chose to concentrate on game play, which is ultimately why the enthusiast purchases a high-end graphics card. But first, a series of performance tests with two highly regarded, and optimized, graphics engines will help determine the "sweet spot" for the GeForce FX 5900 Ultra.
AMD Athlon XP 2700+ @ 2.17GHz - Thouroughbred Revision B
ASUS A7NX8 Deluxe nForce2 Motherboard With DualDDR Memory
Maxtor DiamondMax Plus 80GB 7200RPM ATA-133 Hard Disk Drive
Maxtor DiamondMax Plus 40GB 7200RPM ATA-100 Hard Disk Drive
Sony Multiscan E500 CRT Monitor - 21-Inch
NVIDIA GeForce FX 5900 Ultra - 256MB - 450MHz/850MHz
NVIDIA Detonator XP Beta Driver Version 44.03
32-Bit Color / Sound Disabled / Vsync Disabled / 75Hz Refresh Rate / 60HZ at 2048x1536
Windows XP Professional with Service Pack 1 / DirectX 9.0a
Expert System Performance
166MHz Front Side Bus
Memory Timing 4-2-2
CAS Latency 2.0
APIC Mode Disabled
Age Of Mythology - v1.05
Dungeon Siege - v1.11
IL2-Sturmovik - v1.2
Morrowind - v1.2.0722
Need For Speed 2 Hot Pursuit Demo
Quake 3 Arena - v1.30
Raven Shield Single Player Demo - v1.0
Return To Castle Wolfenstein Single Player Demo - v1.0.1
Star Trek Elite Force 2 Single Player Demo - v1.0
Unreal Tournament 2003 Demo - v2206
Unreal 2 The Awakening Demo - v.1000
Warcraft 3 - v1.05
Game play tests were conducted with sound enabled.
Note that on the GeForce FX 5900 Ultra, the default image setting is Quality, which delivers the highest quality of anisotropic filtering. The High Performance and Performance settings are compromises between texture quality and performance and are based on adaptive techniques since they limit anisotropic filtering to textures that require it.
GeForce FX 5900 Ultra Performance & Quality Settings
The following images are from a program that was specifically developed to measure anisotropic filtering quality, and can be downloaded from Beyond3D. The program renders a three-dimensional cylinder and contains a black and white texture that's applied to the interior.
1X Anisotropic Filtering
The black and white area illustrates the amount of texture detail that anisotropic filtering is delivering. More color represents lower image quality.
8X Anisotropic Filtering
The following thumbnails from Quake 3 are examples of 8X anisotropic filtering on the NV35 using the three Image settings.
NV35 Image Settings - 8X Anisotropic Filtering
The Quality setting provides the highest level of texture filtering, while the Performance and High Performance settings are adaptive texture filtering techniques. In this scene, the main area of difference between Quality and Performance is shown in the second set of images of the ground. However, the GeForce FX 5900 Ultra was architected with keeping the Quality setting enabled.
Let's take look at the performance of the Quality, Performance and High Performance Image settings in Quake 3 and Unreal Tournament 2003. For Quake 3, the default high quality graphics setting was enabled, along with maximum geometry and texture detail. In Unreal Tournament 2003, the MaxDetail user and system ini files were used with the handy UT2K3 Benchmark program.
Quake 3 Demo Four - Anisotropic Filtering Performance
With the GeForce FX 5900 Ultra, 8X anisotropic filtering with an Image setting of Quality is a starting point for game play testing. The NV35 has been particularly impressive in this area as it delivers rich texture quality and minimizes the performance hit associated with using all levels of anisotropic filtering.
The crew over at ToMMTi-Systems have developed a slick application that can be used to illustrate antialiasing quality. The thumbnails below show a portion of the full size test image, which can be viewed by clicking on each thumbnail.
Antialiasing is a desirable feature as it attempts to eliminate jagged edges and can also reduce and eliminate texture shimmering. After having analyzed the images below, 4XS does the best job on horizontal edges, 6XS and 8X do the best on angled edges, and 8X does the best on vertical edges.
However, NVIDIA's mutisampling techniques could use some improvement. Especially when you take into consideration that a significant impact on performance with 4XS and higher still exists on the GeForce FX 5900 Ultra.
NV35 Antialiasing Modes Comparison
Here's a comparison of no antialiasing and 4X antialiasing at 1024x768 from Unreal Tournament 2003.
However, I will caution the reader as the image quality captured by in-game screenshots or image capture programs may not be representative of what is actually being rendered on screen. For example, the following two images were taken on the GeForce FX 5900 Ultra using Quake3's in-game screen capture feature. The image on the left was taken with antialiasing disabled and 8X anisotropic filtering. That image is representative of what was being rendered on the screen.
The image on the right was also taken using Quake3's in-game screen capture feature, but it does not represent what was being rendered on the screen. Closer inspection of the two images reveals that the image on the right is blurred. While that was the image captured in Quake 3 with 4X antialiasing enabled, the image being rendered on screen is better represented by the image on the left. The slight blurring of textures that is typically associated with mutisampling antialiasing wasn't present. I'm certainly going to conduct more tests in this area.
Let's check out antialiasing performance. Note that the Quality Image setting will be used for the remainder of this preview.
Quake 3 Demo Four - Antialiasing Performance
UT2003 Antalus Flyby - Antialiasing Performance
The benefit of moving to a 256-bit memory bus becomes apparent as the GeForce FX 5900 Ultra only takes a 30% hit in performance with 4X antialiasing enabled at 1600x1200 in Quake 3 and Unreal Tournament 2003. Therefore, 4X antialiasing is chosen a second reference point for game play testing. Note that higher levels of antialiasing can be used, but are beyond the scope of this initial preview.
And finally, a performance recap, which includes benchmark results with 4X antialiasing and 8X anisotropic filtering enabled.
Antialiasing & Anisotropic Filtering Performance
The Detonator driver control panel includes a temperature monitoring applet, which provides the temperature of the GeForce FX 5900 Ultra processor. This time of year, outdoor daytime temperatures average around 75 degrees Fahrenheit (24 degrees Celsius). Even with central air conditioning, it gets quite warm in the computer room, which is located on the second floor.
Last year I installed a window unit air conditioner and with the room temperature at 74 degrees Fahrenheit (23 degrees Celsius), the GPU core registered 45 degrees Celsius (113 degrees Fahrenheit).
GeForce FX 5900 Ultra Core Temperature
I measured the GPU core temperature under 3D by looping through a Quake 3 timedemo running in a window, which allowed me to instantly see the changes in temperature.
GeForce FX 5900 Ultra Core Temperature - 3D Stress Test
The following table shows the core temperature reading following the completion of selected timedemo loops.
3D Stress Test - Core Temperature Readings
The highest temperature occurred during playback of the 34th timedemo and reached 84 degrees Celsius (183 degrees Fahrenheit). After completing the 3D stress test, the GPU temperature dropped to 57 degrees Celsius after 30 seconds, 53 degrees Celsius after 1 minute, 50 degrees Celsius after 2 minutes, and 48 degrees Celsius after 3 minutes.
The GeForce FX 5900 Ultra contains clock frequency settings for 2D performance (300MHz/850MHz) and 3D performance (450MHz/850MHz). There's also a command button labeled Auto Detect, which contains the following help text:
Auto Detect Help Text
Following a series of brief tests, the auto detect feature automatically adjusted the clock speed to 491MHz/933MHz. Auto detect also set the screen refresh rate to 60Hz, which I manually changed back to 85Hz.
Auto Detect Results
In order to complete all of the Unreal Tournament 2003 benchmarks, I needed to drop the frequencies down to 485MHz/925MHz.
UT2003 Antalus Flyby - Overclocked Performance
At 491MHz/933MHz the system didn't lock up, but would pause a few seconds towards the end of the benchmark, then automatically resume.