By John Grabski and Jeff Bailey - December 18, 2003
Competing with the 3D Blaster 5 FX5900 in this review will be a recently purchased 128MB GeForce FX 5600 Ultra from XFX. The GeForce FX 5600 Ultra is at a disadvantage since its limited to a 128-bit memory bus and based on the original implementation of NVIDIA's CineFX Engine.
XFX GeForce FX 5600 Ultra
I ran performance tests on two Athlon based systems. System A has some components that are a few generations behind.
Arithmetic: 7241 MIPS, 3004 MFLOPS - close to the reference Athlon XP 2400+
Multimedia: 17482 Integer, 17967 Floating Point - between the reference Athlon XP 2200+ and 2400+
Memory Bandwidth: 2235 ALU, 2087 FPU - adequate, but KT266A memory bandwidth isn't a strong point
Cache Memory: near the reference Athlon XP 2400+
System B is a mid-range PC with an up-to-date chipset and 512MB of memory. It is owned by Jeff Bailey, who co-authored this review, and was primarily used for synthetic benchmark testing. A majority of the gameplay tests were done on my system (System A).
Arithmetic: 8388 MIPS, 3471 MFLOPS - right around the reference Athlon XP 3200+
Multimedia: 20216 Integer, 20340 Floating Point - right around the reference Athlon XP 3200+
Memory Bandwidth: 3005 ALU, 2805 FPU - where it should be for a dual-channel setup
Cache Memory: right at the reference Athlon XP 3200+ levels until the 512KB block tests
SYNTHETIC BENCHMARK RESULTS
Performance testing begins with synthetic benchmarks from Futuremark, SPEC (Standard Performance Evaluation Corporation) and NVIDIA. Futuremark's 3DMark2001 and 3DMark03 are DirectX 8 and 9 game-like benchmarks and provide an overall score referred to as the number of 3DMarks.
SPECviewperf measures CAD related application performance under OpenGL. Results are in frames per second and are based on a weighted geometric mean.
Tests were conducted at resolutions of 1024x768 and 1280x1024 without antialiasing (AA) and anisotropic filtering (AF), no AA and 8X AF, and 4X AA and 4X AF.
In 3DMark2001, the 3D Blaster 5 FX5900 shows improved processor scalability over the GeForce FX 5600 Ultra. Without AA and AF, the increase in 3DMarks from System A (slower) to System B (faster) on the GeForce FX 5600 Ultra was 13% at 1024x768 (from 10324 to 11709) and 19% at 1280x1024 (from 8002 to 9551).
With the 3D Blaster 5 FX5900, performance improved by 25% at 1024x768 (from 11487 to 14390) and by 31% at 1280x1024 (from 9581 to 12591).
3DMark2001 Benchmark Results
With 4X AA and 4X AF enabled, the 3D Blaster 5 FX5900 blew past the GeForce FX 5600 Ultra on both systems courtesy of its 256-bit memory bus. On System A, the 3D Blaster 5 FX5900 beat the GeForce FX 5600 Ultra by 25% at 1024x768 (from 6996 to 8748) and by 55% at 1280x1024. On System B, the number of 3DMarks increased by 41% at 1024x768 (from 7802 to 10971) and by 61% at 1280x1024.
Processor and memory capabilities are less of an influence in 3DMark03 as Futuremark focused their attention on developing a benchmark that targeted graphics card performance. These results mirror the previous statement as the difference in performance of each graphics card between System A and System B only ranged from 2% to 4%.
With no AA and no AF, the GeForce FX 5600 Ultra achieved an overall score of 3064 on System A and 3137 on System B, which is a difference of 2%. The difference in the performance of the 3D Blaster 5 FX5900 between System A (4889) and System B (5041) was also 2%.
3DMark03 Benchmark Results
However, when the performance of both graphics cards is compared on each system, the difference between them is significant. On System B with 4X AA and 4X AF, the 3D Blaster 5 FX5900 outperformed the GeForce FX 5600 Ultra by 61% (3055 vs. 1901). Again, the improvement comes from the 256-bit memory bus and more powerful shader processing units of the 3D Blaster 5 FX5900.
NVIDIA's ChameleonMark was specifically developed to showcase the new features of the GeForce3 and measures the performance associated with DirectX 8-based pixel shaders. Pixel shaders are used to generate the three skins of the Chameleon.
A real skin.
ChameleonMark - Real Skin
A glass skin.
ChameleonMark - Glass Skin
And a shiny skin.
ChameleonMark -Shiny Skin
The ChameleonMark tests were run without AA and AF.
ChameleonMark Benchmark Results
Noticing that there is a minimal difference in the performance between System A and System B without AA and AF, it is unlikely that the greater memory bandwidth of the 3D Blaster 5 FX5900 gives it a significant advantage in ChameleonMark. Therefore, we are left to assume that the advantage is either a result of the superior shader performance of the 3D Blaster 5 FX5900, or more than likely, that the shader performance of the GeForce FX 5600 Ultra was lacking.
The Specviewperf tests, which run at a resolution of 1024x768, reveal that the capabilities of the CPU and memory subsystem play a much greater role in determining performance than the GPU.
SPECviewperf 7.1 Benchmark Results
With either graphics card, System B outperformed System A by 58% in Lightscape (~10 to ~16 fps), 80% in Data Explorer and Pro/ENGINEER 2001, and 100% in DesignReview.
DirectX gameplay performance was conducted using the Teenage Mutant Ninja Turtles demo, Unreal 2, Halo, and Gun Metal, which makes use of NVIDIA's Cg (C for graphics language). OpenGL gameplay performance was analyzed using Serious Sam: The Second Encounter.
FRAPS was used to generate the average frame rate during gameplay. My primary objective during gameplay was to stress the graphics subsystem by enabling the highest quality in-game graphics options. There were cases where I might not have actually played with a specific setting enabled due to its impact on performance. With the exception of the TMNT demo, all games were played on outdoor levels for added stress.
Gameplay Results - 1024x768
Halo confirms that the shader performance of the GeForce FX 5600 Ultra is lacking. Halo uses a mix of pixel shader 1.1, pixel shader 1.4, pixel shader 2.0 half-precision and full-precision, all of which perform significantly faster on the 3D Blaster 5 FX5900.
The second area where the 3D Blaster 5 FX5900 clearly outperforms the GeForce FX 5600 Ultra is when AA and AF are enabled. Here we find increases ranging from 13%in SS:SE to 48% in Gun Metal. The increases would have been greater had I been able to test gameplay on System B.