thoughts about the review at Tomshardware/Rivastation

[CaineTanathos]

You can find the review here : http://www.rivastation.com/parh-review.htm

in that review on rivastation they also use sharkmark:
the scores
GF 4600 - 111
Parhelia - 166
which makes actually sense because the 4600 has 2 vertex shaders at a speed of 300 mhz and the parhelia's has 4 vertex shaders at 220 mhz


So how come that the parhelia's has low vertex shader scores in 3dmark , my guess is that 3dmark only uses 2 shaders:
Vertex Shader scores:
GF 4600 - 102,1
Parhelia - 86,5

And the "low polygon count" scores:
GF 4600 - 51,0 (1 light) - 12,6 (8 lights)
Parhelia - 25,2 (1 light) - 11,1 (8 lights)

?? pretty weird , this looks more like a lack of driver optimization

and the pixelshader :
GF 4600 - 44, 6
Parhelia - 24,9
advanced pixelshader:
GF 4600 - 123,3
Parhelia - 58,0

as you can see the difference between the normal pixelshader and advanced pixelshader is more than a double with the GF 4600 and the Parhelia double his score , why not quadriple ?

again my guess is that 3dmark only uses 1 pixelshader in the normal PS benchmark and 2 pixelshaders in the advanced pixelshader benchmark

and another lack of optimization of the driver , I think

[Jazzz]

Isn't that review the same as the THG one? AS its the same guy doing the review

[CaineTanathos]

whoops, DOH

[WayneHu]

I think this is the main short coming of the current Parhelia. Because there was never 4 pixel unit hardware design, it is possible that most softwares were optimized to issue at most two pixel calculations one time in the rendering pipeline. In those cases, the fact can be that some of Parhelia's execution units are IDLE at most time or just with 50% utilization.

The problem is here ... The core clock of current Parhelia is 220MHz vs 300 MHz GF4 Ti4600. In the case when GF4 does not have the external memory bandwidth problem, it can easily beats 220 MHz Parhelia with just 2 pixel units or 2 texture units in work.

I do not have any knowledge about Parhelia's execution unit design. If it can be programmed as Voodoo SLI mode, however, there might be a very interesting result coming out...

[Wombat]

I do not have any knowledge about Parhelia's execution unit design. If it can be programmed as Voodoo SLI mode, however, there might be a very interesting result coming out

I predict absolutely no advantage from something like that. I doubt P's true problem is fill rate.

[CaineTanathos]

I think that current yields of producing the chip are low too.

Because there were a few reviewing who had a dead card or which card died after a day (because of bad wireing of the fan, or something)

as we all know there will be new revisions just like Nvidia and Ati did.

I think its the same proces as with Intel and AMD, by improving the proces clockspeed improves too, the chip stays the same.

[WaltC]

Re: thoughts about the review at Tomshardware/Rivastation

Originally posted by CaineTanathos
You can find the review here : http://www.rivastation.com/parh-review.htm

in that review on rivastation they also use sharkmark:
the scores
GF 4600 - 111
Parhelia - 166
which makes actually sense because the 4600 has 2 vertex shaders at a speed of 300 mhz and the parhelia's has 4 vertex shaders at 220 mhz


again my guess is that 3dmark only uses 1 pixelshader in the normal PS benchmark and 2 pixelshaders in the advanced pixelshader benchmark

and another lack of optimization of the driver , I think

Something like this is definitely possible. Although Mad Onion denies it, I've always felt there was a lot proof to suggest that 3D Mark is heavily optimized for nVidia architectures. This goes way, way back to when TNT2's were getting much better scores in 3D Mark than V3s while 3dfx V3s were blowing them away in real 3D games.

For the last couple of years a lot of game developers were consciously or otherwise optimizing around nVidia hardware simply because that hardware has been ubiquitous among gamers. People no less than John Carmack have pushed the hardware zealously. Some of the 3D Mark code may contain nVidia-specific optimizations of which the 3D Mark programmers are unaware, believe it or not.

But lately it looks as though things are changing for the better, with renewed competition from ATI (I never knew ATI could be this competitive), to newcomers in the 3D market like 3D labs. And Matrox is chiming in with P. Indeed, even Carmack senses a sea change as he is now investigating and actively promoting architectures other than nVidia, which I for one am glad to see him do.

I think nVidia is aware of what's going on as well as their bid to grab the OpenGL spec through Cg, an in-house nVidia-specific tool which nVidia is pushing as an "open standard" is no less than a bid by nVidia to get back in the driver's seat so that most software will continue to be developed on nVidia products and thus will run better on nVidia products as a result. It's very possible that nVidia may indeed be usurped in the coming months and that developers will be much more open minded about developing on a variety of products other than nVidia as time goes on.

Meantime, take your 3D Mark numbers with a large grain of salt...

[WayneHu]

Originally posted by Wombat
I predict absolutely no advantage from something like that. I doubt P's true problem is fill rate.

Which part of Fill rate? Single Texture or Multi-Texture?

According to the number of Multi-Texture Fillrate in 3D Mark 2001 SE, it even beats GF4 Ti4600.

However, the Single Texture Fill rate of Parhelia is much lower than GF4 Ti4600's (and slightly less than Radeon 8500's). That is why I guess some of execution units are idle or useless with those kind of operations.

The Pixel Shader test in this benchmark tells the same thing. The Advanced Pixel Shader can draw more Triangles than the regular Pixel Shader. This behavior is totally different from Radeon 8500 or GF4 Ti 4600.

[WaltC]

Originally posted by WayneHu


Which part of Fill rate? Single Texture or Multi-Texture?

According to the number of Multi-Texture Fillrate in 3D Mark 2001 SE, it even beats GF4 Ti4600.

However, the Single Texture Fill rate of Parhelia is much lower than GF4 Ti4600's (and slightly less than Radeon 8500's). That is why I guess some of execution units are idle or useless with those kind of operations.

The Pixel Shader test in this benchmark tells the same thing. The Advanced Pixel Shader can draw more Triangles than the regular Pixel Shader. This behavior is totally different from Radeon 8500 or GF4 Ti 4600.

Yes, but single-texturing in 3D games has been "out" for a relatively long time now, hasn't it? Heck, I recall as far back as the Voodoo Banshee what a flop the card was because it only did single texturing compared with other products at the time which multitextured.

But I think it's statisitics like these that make 3D Mark such a chore to interpret. I do think the aggregate total mark is "OK", but the fact is that every single test in 3D Mark is weighted in the determination of the final score. How much weight in the aggregate score does the "single-texturing" benchmark pull as compared to the multitexturing bench? Obviously the multitexturing bench should pull a lot more of the aggregate weight in the final score simply because few games today do much in the way of single texturing. Does it, though? Indeed, which tests are are weighted to what degree in the final aggregate score? Is this known? I'm not sure that anyone knows the weight that each of the tests in 3D Mark pulls relative to the aggregate score, and that would be my biggest criticism of it. Obviously the weights applied to each test are determined subjectively by the programmers of 3D mark according to what they feel are the "important" tests versus the "relatively unimportant," so there's a lot of room for error as well as bias. Hopefully, though, all the tests are not given the same weight and then averaged, as that would make 3D Mark's final scores completely irrelevant to current 3D gaming software (in which some technologies are used much more than others.)

[TdB]

actually, everybody who bothered to read the readme file that comes with 3dmark knows that only the gametests influence the score, the theoretical tests don´t contribute anything to the final score, madonion made those extra test solely for fun.
for example the radeon8500 gets no credits what so ever, for having pixelshaders 1.4.