LOSTCIRCUITS

Drivers and Performance

The first encounter with the FireGL 8800 was using the standard drivers shipping on CD together with the card itself, revision number 3024. The performance using these drivers was, er, not exactly convincing, moreover, we encountered some more or less severe stability problems on the P4 DDR platform. There is really nothing fundamentally wrong with that, the drivers were about 10 generations behind the current release and mostly geared towards stability and optimal compatibility at the launch of the FireGL, while sacrificing some performance. How much performance was sacrificed compared to the latest driver revision? The answer to this depends on the application. Now, we all know it is one thing to run a benchmark, leave the room because watching it again for the gazillionth time only causes migraines and come back to read off the final score. It is another thing which requires certain self-destructive character traits to sit and endure watching the benchmarks trying to find the anomalities in performance.

Initial Observations

AWadvs-04

One of the results we were initially concerned about was AWadvs-04 from the ViewPERF 6.1.2 suite. With a standard GeForce3 Ti500 card, this benchmark yields average scores of some 90-95 fps on either platform used. The Ti4600 cards are capable of improving these scores to roughly 100-105 fps. The 3DLabs Wildcat 6110 achieves up to 185 fps and the nVidia Quadro holds the record with some 200 fps. So why did the FireGL 8800 come out with only 42 fps?

The answer is, or the answers are that first of all, we are looking at different versions of AWadvs, the 03 version runs much faster than the 04 version and it is really necessary to pay attention which version is used for the benchmarks.

More important is to look at the individual scores of the subroutines because they are telling a rather interesting story:

Raw fps of the ATI FireGL8800 on the ASUS A7M266D dual AMD platform, using either the 3024 or the 3035 drivers.
Looking at the raw summary scores we see a doubling of performance between the original and the latest drivers but that by itself only tells a very poor story of what is going on. A very simplified version is that the first few runs involve more texturing and, therefore are more sensitive to correct driver implementation regarding the interaction with the AGP interface whereas the later runs appear to depend more on GPU-internal T&L than on memory bandwidth and, therefore reflect more the hardware capabilities.

What is very obvious from this close-up of results is that the new drivers yield up to 450 % of the performance of the older drivers. Granted, there are some issues with this version of AWadvs-04, we have seen some mainboards fail miserably on the first iteration of this benchmark and showing outstanding performance if the benchmark was simply repeated. That is, until the next shut down when the game started all over. In the case of the 3024 drivers, repetition did not improve the situation at all. However, the issue was observed on all platforms we ran this benchmark on. More interesting is that we also observed stability problems, however, those could be solved almost completely by either running the memory at higher speed or else reducing memory latencies, as counterintuitive as this may appear at first glance.

A few years back, we did observe a very similar sitation with the introduction of the AMD K6-2 CXT core where improper placement of the Write Allocation and Write Combine buffers over the local frame buffer of the graphics card caused similar problems. In this case, the issue was remedied either by proper use of the setK6 utility or later with the pioneering drivers from 3DFX that not only solved stability problems but also almost doubled performance. There are many possible issues with drivers and this is just one possibility but the similarity between the scenarios is rather compelling.

So why are we even talking about this? Ati have made themselves a reputation for their original release drivers that "ain't pretty" but that is really not the point. On the contrary, the issue at hand is that with newer driver revisions some real performance improvements can be expected. From some other benchmarks we were running, it does appear as if there is still some margin or potential for even higher performance increases, thus it is worth to check for driver updates.

Next Page:    => The Devil in the Platforms =>