Parhelia volt. mod. nearly finished! ( with a pencil )

[Indiana]

Originally posted by |Mehen|
wow, im amazed at how close the P was, even beating the 9700 in a bunch of tests, impressive. it just goes to show what a shame it was that matrox didnt have a higher clock speed on the P.

Stop looking at the game tests. They don't tell much about gfxcard performance once a certain level is reached, because they're mostly CPU-limited. There was no real loss due to downclocking the R300 in the Lobby + car chase high and low detail and the Dragothic high detail tests. Raising / lowering CPU clockings has a much higher effect here.

An exception here is the Game4-Nature test, this scales fairly well with gfx-card performance, hence here you gain / lose considerably when varying gfxcard clockings.

The theoretical test of 3DMark are much better as those (Vertex- + PixelShaders, BumpMapping, Fillrates) nearly don't depend on the CPU at all but solely on gfxcard performance in the respective field. Here you can see that while the Parhelia has impressive MT fillrate and decent polygon-processing power, it somehow fails to translate this into real performance.
The shaders in special seem to be a very weak point of the current Parhelia (or the current drivers?), thus the bad result in the Nature benchmark. Unfortunately the use of shaders WILL be raising in the next game titles, so if this is a driver issue, it should be fixed ASAP.

P.S.: This strong CPU-limitation of most of the game tests is also the reason why the GF FX simply cannot reach 25000 3DMarks, because the score unfortunatelx is only build from those game-tests, the much better theoretical tests don't count in. faster than a GF4 Ti4600 can only really get higher scores in Nature and Dragothic low.

[WyWyWyWy]

Indiana how do you capture longer than one screen?

Thanks.

[Indiana]

Originally posted by WyWyWyWy
Indiana how do you capture longer than one screen?

I don't - I make the screen big enough... I captured the above using an oversized 1900x1440 desktop on my 1280x1024 TFT.

[Novdid]

Originally posted by Indiana
An exception here is the Game4-Nature test, this scales fairly well with gfx-card performance, hence here you gain / lose considerably when varying gfxcard clockings.

The theoretical test of 3DMark are much better as those (Vertex- + PixelShaders, BumpMapping, Fillrates) nearly don't depend on the CPU at all but solely on gfxcard performance in the respective field. Here you can see that while the Parhelia has impressive MT fillrate and decent polygon-processing power, it somehow fails to translate this into real performance.
The shaders in special seem to be a very weak point of the current Parhelia (or the current drivers?), thus the bad result in the Nature benchmark. Unfortunately the use of shaders WILL be raising in the next game titles, so if this is a driver issue, it should be fixed ASAP.

The nature test is more of a video card bandwidth test and less of a shader bench. Try keeping the core frequency and play with the mem settings. You'll find interesting results. Could also confirm that Parhelia has an awfull mem controller (or bad drivers)...

You know, 3dmark sucks, the tests doesn't even test what they are supposed to test.

[Indiana]

If you know to "read" the results 3DMark is not so bad. Of course it has its biases towards certain CPU and gfxcard manufacturers, but at least the theoretical tests are of some worth if you keep that in mind.

And nature DOES test the PixelShaders, it's just that all those nifty pixelshader apps seem to eat bandwidth like nothing (take a look at the Codecreatures benchmark here as well) Once the bandwidth is there - as it is with the R300 - the GPU performance does matter.
My normal 3DMark lists 114.6fps for nature; the downclocked result on the last page gave 94.5fps. Those tests ran at the exact same memory speed (337MHz), only the GPU-clock was different (385MHz vs. 275 MHz), that makes 18% lesser performance due to a 29% GPU downclock at the exact same Ram clocking - that's not so bad scaling.

[Novdid]

But you have to agree that if it was supposed to only test pixelshader performance bandwidth would not be as big of issue as it is in the nature test. Just look at the advanced pixelshader test, it's much more of a pixelshading test than nature because it isolates itself more to pixelshading and not mem bandwidth or CPU speed.

[Chrono_Wanderer]

how many pixel shaders are there on the Parhelia?

it's quite intersting that Parhelia does bad on Nature. I thought there were not so much pixel shader instructions integrated into it.

it's quite intersting that Parhelia does better at 8 lights than 1 light. It does look like the Parhelia drivers aren't very optimized. btw are those R9700 benches done with DX9 specific drivers? (catalyst 3.0)

[Novdid]

Yes they are.

[William]

In the Nvidia released benchmarks of the FX; doesn't the 9700 beat the FX ever so slightly in Nature as well? Maybe that's the Radeons strong suit.

[Android604]

Nature test is about driver tweeking... when the Radeon 8500 was first released is had a nature score of 15 FPS or so... after about the thrid driver release is was humming at 50 FPS.

I figure if Matrox get to some good driver tweeking once the bug fixing is done, we will see some better scores, if that ever happens?

Andrew

[drslump]

If 3DMark developers release the source code of the tests we would see if it's a fair benchmark or not.
While we can not see the code I'll assume that it's a "3DMark score" but not something showing the real power of a video card.

I don't think that is a good idea to optimize the drivers for a specific benchmark util, IMO it's a waste of time and resources. Matrox should try to optimize the drivers for speed when using common programming algorithms, if it is easy to develop fast effects for the parhelia lots of developers will add more eye-candy to the P. path, if they have to invest lots of time and money in optimizing for the P. at the end we will end up with a card which can not show it's real power in real world apps/games.

[P5ycho]

noone will develop for the parhelia, look at the amount of Parhelia's versus the amount of Geforces/Radeon's at the market right now and the conclusion has to be no developer will spend time developing for the parhelia.

[drslump]

I agree, most of them will not even have a P. in the office to develop for it.
The thing is that most of the games are based on a few engines (Quake, UT, Lithtech), so Matrox should never optimize for 3Dmark but for that engines and make sure that upcomming engines (doom3 for example) take full advantage of the card power.

[Scratchi]

The most important stuff now also in english. Starting from chapter 3.

http://www.combu.ch/Parhelia/Parhelia_Voltmodding_english.html

cu Itchi

[WyWyWyWy]

Thanks Scratchi keep it up!! Very nice