AquaMark3: Benchmarking Results
We tested each of the graphics cards two times: in the “pure speed” mode to check their maximum possible performance and in the “eye pleasing” mode with full-screen anti-aliasing (FSAA) and anisotropic filtering (AF) enabled. The level of AF was set to 8, as this is the maximum mode supported by NVIDIA GPUs. As for ATI’s chips, we preferred 8x over 16x, since there is a very slight difference between them in speed as well as in quality. So, it’s time we checked the actual test results.
We have got a new leader, GeForce FX 5950 Ultra. Ultra frequencies, super-fast memory, and much-improved drivers all contribute to the result: the NV38 is everywhere faster than the RADEON 9800 XT by 2-3 frames per second. In the mainstream graphics cards class, NVIDIA is also on the winning side with its GeForce FX 5700 Ultra, and this time NVIDIA’s creation is substantially faster than ATI’s.
After turning on FSAA and AF, we witness a well-known effect: the cards on ATI VPUs easily break away from the competitors, leaving them far behind. The architecture of ATI’s chips coupled with efficient AF and FSAA algorithms helps handling huge workloads.
Now, let’s get to the specific tests from the AquaMark3 package.
Dynamic Occlusion Culling
The GeForce FX 5950 Ultra wins the test in which invisible surfaces are discarded, while the GeForce FX 5700 Ultra goes close to the RADEON 9600 XT. The performance of the two top-end cards seems to be limited by the CPU in this test. The GeForce FX 5700 Ultra is once again faster than the RADEON 9600 XT.
It should be noted that NVIDIA has been implementing more efficient culling algorithms since the previous generations of its GPUs. That’s why it’s not much of a surprise to see the GeForce based cards winning this test. On the other hand, we shouldn’t forget that NVIDIA GPUs work at higher frequencies than ATI’s ones.
Under higher workloads, the GeForce FX loses its ground. High frequencies and software optimizations cannot make up for all the shortcomings of the architecture. As a result, the more efficient culling algorithms can’t help the NV36 and NV38 out in this test.