No matter what we did, the fill rate did not show more than 280 Mtexel/s.
|
Advice Beginners BIOS Guide CPUs Links Mainboards Memory Network Storage Video/Sound Cards Contact Forum SiteMap Sponsors WebNews Home |
. | . |
Prices: Mainboards ABIT ASUS Chaintech Shuttle Soyo Tyan CPU Intel P4 2.4C-800 P4 2.6C-800 P4 2.8C-800 P4 3.0-800 P4 3.2-800 AMD AthlonXP XP 1700+ XP 2000+ XP 2400+ XP 2500+ XP 2700+ XP 3000+ XP 3200+ Athlon64 Athlon64 3200+ Athlon64 FX-51 Opteron Opteron 240 Opteron 242 Opteron 244 Opteron 246 Memory Corsair Crucial Kingston Mushkin OCZ |
LOSTCIRCUITS |
|
| 3DLabs Wildcat VP990 512 MB LFB but... Whatever Happened to the Specs | |
| (Review by MS, August 1, 2003) |
Fill Rate
We used 3DMark2001SE and 3DMark2003 to measure the sustained fill rate for both single- and multi-texturing in Direct3D and the results were, er, interesting in that both benchmarks reported 270 and 280 MegaTexels for single and multitexturing, respectively. Both benchmarks show the same results, that is, approximately one quarter only of what 3DLabs specifies, however, they are also both part of benchmark suites released by the same company and, by extension, could carry the same "flaws" that is, lack of exposure of the actual pixel / texture pipelines. Aggravating in this case is certainly that the Wildcat VP990 is not necessarily geared towards Microsofts Direct3D but rather towards a high end OpenGL environment.
GL_EXT_reme as an OpenGL-specific benchmark solved the conundrum by showing results identical to those obtained with 3DMark.
No matter what we did, the fill rate did not show more than 280 Mtexel/s.
OpenGL T&L Performance
GL_EXT_reme appears also a very useful tool to measure the actual T&L capabilities of any GPU by generating data showing not only the raw vertex performance but also the impact of two or eight light sources on the vertex engine. 3DLabs advertises the VP990 as capable of setting up 225 Million vertices /second, we reckon that this refers to raw transformation of geometry data rather than the more complex combined T&L.
Our results in GL_EXT_reme were somewhat disappointing, pure transforms amounted to some 20 million vertices per second, for reference, an ATI 9700 is about 2.5 x faster in all disciplines listed here than the VP990. To be fair, we need to state that we don't know the exact metrics applied here to generate the data.
T&L performance at 1024x768x32 / 70Hz refresh (vSync disabled) to show the toll the different methods of lighting take.
Springmark
As additional test, we used Springmark which generates a spring/coil composed of 1 million vertices per loop. Each additional loop adds another 1 million vertices. The number of vertices times fps gives the number of vertices/sec generated in a large scale model.
Click for larger picture
Vertex generator performance dependency on the number of vertices / model. Numbers shown are the fps with increasing number of loops, the amount of geometry memory used by the model and the number of vertices (in millions / sec) depending on model size. Springmark uses one directional light, backface culling, depth test and smooth shading are enabled. Moreover, the object is centered such that no clipping is necessary (the full object always stays within the 2D Viewpoint Space). According to Springmark, the Wildcat VP990 can generate 50 million vertices / second, however, with increasing model size, the performance drops to ~ 40 million vertices. The Quadro4 GXL980 only manages 28-29 million vertices / sec, however, there is no performance degradation with increasing object size. Also keep in mind that only one single directional light is used, which, according to the numbers shown above, implies that these numbers are extremely close to the pure vertex performance.
Next Page: => SPEC.org ViewPERF 7.0 =>
