ATI Radeon X800 XT (R420): Extreme force
Problems of tests: "cheating" and "optimizations"
The topic of chronical lack of adequate tools and quick removal of old tests from use is the favorite talk of the town among the employees of test labs at leisure. With the release of R420, two more tests of the good old arsenal have been washed off. For example, this is how tests of ATI Radeon X800XT look in Comanche 4:
The performance of the video card is somehow seen only in the resolution 1920x1440 at the most aggressive quality settings. All the remaining figures characterize only the system CPU and the limit of the game's engine.
Another ancient test - Village Mark (evaluating the hidden surfaces cut-off efficiency) does not perceive results higher than 200 fps, and Radeon X800XT easily jumps over these 200 fps even at 1600x1200.
We've been constantly convincing developers that addition of the benchmarking kit guarantees at least that the game will be surely optimized in NVIDIA drivers. What is more important, the game will immediately acquire immense popularity and absolutely free "citing rate" at hundred and thousands of Internet resources which are in bad need for new tests.
Official attitudes of ATI and NVIDIA to the problems of testing video cards are also different - ATI recommends using FRAPS for the case of lack of a game console or an integrated benchmark, but for NVIDIA the release of every new game with a benchmark included is another headache - it has to be included into the list of "first-priority optimization items". Both these factors are not inspiring. FRAPS is good only for the case of a very serious performance difference for a video card, e.g. like this:
 |
 |
| ATI Radeon 9800XT |
ATI Radeon X800XT |
FRAPS is absolutely useless for comparing video cards of approximately the same rank. Well, who can get absolutely equal ranking while driving NFS:Underground? (Basically, it is possible if you choose the 'Free' and 'no traffic on the road' options).
And the most painful thing - cheating and optimizations by NVIDIA. At that, we have to dwell presenting in more detail. The thing is that "cheating" and "optimizations" are absolutely different notions. So, what is it all about, and where is the problem rooted? Let's try to explain it in as simple as possible ways.
The pipeline architectures of ATI and NVIDIA graphic chips are different, of course. In a word, NVIDIA's pipeline is longer and in theory it is able handling (pushing) more data per cycle than the shorter ATI pipeline can do. But that is possible provided there is the only condition - the input data has got to be optimized (ordered) for sequential data processing just by that pipeline. One error, and all fails. The data is sent back to the beginning, a repeated sampling is done and then sent to the processing again. It's impossible to rule out idle cycles completely, but their percentage must and has to be reduced. This is done using the two factors - by either general improvement of the branch processing mechanism in the compiler, or through hard-coding the name of each particular game thus advising the method for processing its code in the most optimum way. It's just the second technique which is called "optimization" and is absolutely legitimate. Another thing that this sore costs NVIDIA too much. They have a whole staff team (as rumors have it, between 10 to 20 employees) who receive luxurious salaries only for analyzing competitors' code (in all the new games) while firmly keeping their compiler specifics in mind. This is infinite hair of a job which requires exceptional skills. But they have to and ALWAYS will do that! Despite any attempts to make them stop their activities. At that, NVIDIA is absolutely right.
| If all the data is handled by the pipeline properly without turning it back to the beginning and repeated sampling, then potentially NVIDIA should take a convincing lead over ATI's pipelines. But that is only in theory. This never happens in real world - it's like crooked teeth in a watch gear manufactured manually in the 17th century - the probability of their complete match drops by an order of magnitude if the radius of matching increases.
As an illustration, we invite the readers to estimate the probability of successful outcome for the left-hand and right-hand cases.
All the endless new versions of Detonator/ForceWare driver just include more and more new optimizations for NEW GAMES. |
To make their lives easier, NVIDIA invented the Cg language, but that tool-set turned out to be so intricate and complicated that most developers simply don't use it. But even if they decide to learn it, then financial and technological aid from NVIDIA is needed (just remember the marketing initiative called "The way it's meant to be played!" :)
But what about ATI? For ATI, there is no need to do that - the company has taken a different path. They use a shorter pipeline, and due to the fewer number of transistors per chip they were able to increase the chip operating frequency even in R420 as compared to NV40 (525 MHz versus 400). On the one hand, the shorter pipeline handles less data per cycle, but it gets fewer idle cycles and higher operating frequency as compared to the competitor, which makes optimizations for all with their own efforts a useless job.
ATI compiler stands out with its higher predictability of results and offers more stability. Even now, with the beta version of Catalyst 4.5 (for the X800 series), no problems were found. The driver is predictable in its performance and, unlike ForceWare, it never demonstrates any unexplainable performance boosts or strange drops.
 Terry Makedon
"Mr. Catalyst" - head of the
Catalyst development departmentTherefore, 'a la NVIDIA optimization' is a forced measure of the company, otherwise the result will be catastrophic. As a typical example, see the test results in "3DMark2003 build 320 vs 340: a rare moment of truth?" which tells that Futuremark at the culmination of the intrigue simply blocked all NVIDIA's optimizations for 3DMark2003.
Why was that done? And here we have just arrived at the problem of "cheating" which, unlike the "optimizations" is no longer a legal method of competition.
Futuremark has a special version of 3DMark2003 benchmark that demonstrates a map fly-by and once with a version of the Detonator/ForceWare drivers it was found that a whole piece of the map had been simply removed in one of the tests. Actually, this part of the map is not seen to the eye, but this is a fact. Futuremark decided not to sort it all out but simply blocked all the optimizations and cheats, whether legal or not, at one stroke. Both companies were in rage. Futuremark values its own reputation, and they don't seem to lose it at all - even without that they have had a lot of cavils...
The parties agreed upon the following: NVIDIA bound itself not to overstep the "critical threshold" in Futuremark products, and the latter undertook not to make this history public :-).
 |
Top Stories: |
 |
 |
|
MoBo:
| |
|
|
VGA Card:
|
|
|
|
CPU & Memory:
|
|
