Articles: Graphics
 

Bookmark and Share

(14) 
Pages: [ 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | 16 | 17 | 18 | 19 | 20 | 21 | 22 | 23 | 24 | 25 | 26 | 27 | 28 | 29 ]

Performance in Theoretical Benchmarks

As we have already told you, DeltaChrome S8 Nitro differs from the predecessor only by the clock frequencies. But nevertheless, we decided to carry out a full set of theoretical tests, in order to find out how greatly the performance of the new solution grew up due to the working frequencies increase. As usual we will start our theoretical part with the fillrate investigation.

Synthetic Benchmarks: Fillrate

As always we will start the fillrate test with the enabled color writes and Z writes:

What do we see? And we see the same architectural drawbacks, which we have already revealed in the previous version of DeltaChrome. However, the increase in the working frequencies of the chip would never actually cure them, that’s evident. When working only with the Z buffer, S3 chips perform just perfectly, but they immediately lose this advantage when there is one or two textures to be laid. This definitely indicates that the texture caches are two small. After that the graphs levels out and the performance stops its dramatic reduction. Moreover, the solution turns out even faster that the competitors when we have to lay three or four textures. Certainly, 8 pipelines is an indisputable advantage.

Now let’s disable Z writes:

Graphics processors from ATI and NVIDIA have hardly reacted to disabled Z writes. However, DeltaChrome did: the fillrate with 0 textures dropped down a lot. In fact this is not at all surprising, as it lost one of its major trumps: efficient work with the Z-buffer.

It is interesting but this time the transition from one to two textures appeared much sharper than in our previous investigation of the DeltaChrome performance, where the fillrate decreased smoothly with the growth of the number of textures. There is only one reasonable explanation to that: new drivers.

Now we are going to change the testing conditions again: let’s enable Z-writes and disable the color writes.

Usually, when we disable color writes all graphics adapters demonstrate the maximum possible performance, because they do not have to work on textures any more. However, this is not the case for DeltaChrome, which acts like that only if there is not more than one texture involved. As soon as there appears the second texture, the S3 processor slows down a lot. Moreover, the graph in this case looks completely different from what we saw with enabled color writes. Besides, this graphs also has nothing in common with the graph you can find in our previous review of S3 DeltaChrome S8: if you remember, the performance dropped even more when we had two textures to be processed. If you remember, at that time we supposed that this performance drop was caused by the drivers and could be eliminated in the new driver version. Time showed that this was a correct supposition: the situation has been significantly improved with the release of the new driver version, although it is still far from ideal. We assume that the bug in the drivers and/or the architectural peculiarities of the DeltaChrome architecture make it continues attempting to perform the color writes, even if this feature is formally disabled. That is why the performance results turn so poor. Since color writes need to be disabled pretty rarely, S3 software guys could have simply omitted it. The situation with enabled Z-writes and disabled color writes is very unlikely to occur in any game or application, that is why the results obtained in this mode are interesting mostly from the theoretical point of view.

As for the performance gain ensured by the higher working frequencies of the chip, it is most evident in case of 0 or 1 texture. As the number of textures grows, the chip gets more and more loaded with work and the architectural flaws start mattering more than higher working frequencies.

 
Pages: [ 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | 16 | 17 | 18 | 19 | 20 | 21 | 22 | 23 | 24 | 25 | 26 | 27 | 28 | 29 ]

Discussion

Comments currently: 14
Discussion started: 03/10/04 04:59:23 AM
Latest comment: 02/28/08 10:49:42 PM

View comments

Add your Comment