As I promised I keep on tracing new solutions on the professional market. Last time you got quite detailed information on the new solutions but no tests because they weren't adequate to real conditions. In most tests the QuadroFX 1000 turned to be faster than the QuadroFX 2000, while the Quadro4 980XGL showed higher performance than its followers.
The situation was even more complicated because NVIDIA didn't render any help in spite of all my attempts. The conclusions I made were pretty rough, and if such a review had been published two months ago I would have been blamed for the bias towards ATI. But I decided to wait a little to look deeper into this unsophisticated problem.
The conclusions that arose were based on the following facts:
So, the following plan was drawn up to get all necessary explanations
I managed to checked all the ways and now I'm ready to share the results with you. But before we proceed I recommend that you refresh in your memory the April's release on the professional tests.
Well, there are only two cards in the list:
The following cards of the previous generation was used for comparison:
Our aim is to:
NVIDIA QuadroFX 2000
NVIDIA QuadroFX 1000
I used almost all software I selected before for measuring performance of professional cards. Here is it:
As you can see, the list is impressive. We are going to estimate performance of the cards in two CAD programs, two DCC programs and the established professional test SPECviewperf of the 7th version. All that was running under the Windows XP Professional, Service Pack 1, DirectX 9.
The following officially available drivers were used for comparison sake:
The following products were used for comparison of the special driver developed for the 3ds max 5:
Anti-aliasing, anisotropy and vertical sync were forcedly disabled where it was possible. All the other settings were set by default. I didn't use optimizations for applications to find out how clean the card works without additional optimizations in the driver.
Finally, look at the screenshots with settings of some programs.
The tests in the CAD packets were carried out with special benchmarks from http://www.spec.org/. Measurements were taken after installation of the modeling packet and of the test suite. No special settings in the CAD system were used because all necessary ones were made automatically with the test packet used.
We are going to start our examination with the popular benchmark SPECviewperf 7.0. No secret that manufacturers of professional graphics accelerators are guided mostly by this test. Moreover, drivers are often optimized exactly for this test to show an advantage of a new product over an old one though they can have no difference at all (for example, NVIDIA Quadro 900 and NVIDIA Quadro 980). This is a very important thing to know because last time the scores in this test didn't worry at all, and the new version of the drivers shouldn't make problems.
All the cards finish with the expected results. The QuadroFX 2000 works correctly
on the new driver version, being the fastest runner in its line. The proportions
are approximately equal. The difference between the Quadro4 980XGL and the QuadroFX
1000 is about the same for the QuadroFX 1000 and the QuadroFX 2000. The Light-05
demonstrates no difference at all, which means that it strongly depends on the
CPU in spite of two physical processors.
In some tests the absolute speeds are getting higher. But in general, the cards take the same positions in this race. It means that the company managed to make a universal driver which perfectly works with the latest and last generations of the professional cards. The supposition that the wrong results can be caused by the dual-processor system is incorrect because the proportions are the same. That is why it's rather the driver to blame. The BIOS of the card is not guilty because I tested three different samples.
But remember that this test is not cure-all. Its source codes are open, and vendors can easily optimize their drivers for this test. And they do that. I can't say that it's wrong; it's even a good thing if there are optimizations for a certain application on the driver level. If they trigger no artifacts, they are even useful for end-users. But only if they are made for real applications. If they are used only to show how good a new family is, it can be harmful. That is why the tests include real applications as well.
Let's start with the popular 3D modeling packet 3ds max 5. This time I combined the diagrams by the test subjects. For example, the scenes where the geometry is tested are combined on one diagram, the scenes where lighting is checked are on another etc.
Remember that I didn't use the Triangle Strip, that is why the VPU load must be a little higher. Almost all problems in the driver v42.82 are done away with, and the cards work as they should. I must say that the driver optimization for the new QuadroFX line was quite successful, without losses in performance of the previous line. Also, the Maxtreme 4.00.29 driver demonstrated perfect scores, and exactly this driver lets us feel the power of the latest accelerators.
Now let's expand the FSB's bandwidth and add power to the processor and see how it will influence the outcome. Below are the scores for the auxilary platform.
Frankly speaking, I didn't expect so amazing results, especially under the Maxtreme driver. Fast processors and wide system busses are very important for this modeling system, but only if you deal with modeling. If you are going to spend more time on final rendering, a dual-processor system will be preferable.
Now, as we are through with the 3ds max, let's turn to tasks of the other class, in particular, to CAD applications.
For the start I'll take the Solid Works 2001 because it operates with one figure only:
The QuadroFX 2000 rocks! :) But the difference is not great, and it doesn't matter much which card will be used for the SolidWorks 2001. All the cards tested have approximately equal points. Let's wait for a new version of this problem which will employ shaders.
Now the CAD system again, but at a different level - SolidEdge. Below is the summary diagram.
As you can see, the performance mostly depends on a test packet or a central processor. The auxilary platform didn't have a great effect on the speed, that is why we omitted its diagram.
The last big test is the DCC LightWave 7.0.
The difference is actually very fine. Probably, so powerful cards need more than two processors. Another supposition is that the problem is in the OpenGL driver. Last time when we tested ATI's cards the gain on its new products was considerable, and this time it's lacking. But maybe, both these factors are influential.
The scene rendering was flawless.
NVIDIA updated its line of professional accelerators with the QuadroFX solutions. But the procedure didn't go through smoothly. The fact that the driver v42.82 was released simultaneously with bringing the new cards onto the market could disappoint owners of the QuadroFX 2000 if they had a chance to compare the speed with the lower-level solutions. But the time cures everything. The company soon released new drivers where a lot was corrected. Now it works smoothly and quickly.
The new line turned to be faster than the Quadro 4. The reverse of the medal is that software developers do not strive for realizing the shader support in their new programs. But again, time cures. Maybe, in half a year or so we will make real-time clips like the Final Fantasy using all progressive methods and solutions. But by that time we will probably be admiring Next Generation professional accelerators, for example, QuadroFX 4000. :)
Alexander Kondakov (email@example.com)
Write a comment below. No registration needed!
Copyright © Byrds Research & Publishing, Ltd., 1997–2011. All rights reserved.