Parhelia - no flame - makes me wonder

[notagain]

since they are some really sensible people in this forum , I chose to post my wondering here. I have great hope that i will find real answer:


from what i've gather there and there (leaks leaks leaks)

a) -we have the Parhelia 10 to 20% slower (fps wise) than a (let say) GF4 4600 when playng without AntiAlising
b) -on the other hand we have a Parhelia runing with 16x FAA , 30% to 60% a GF4 4600 with 4 FSAA.


Now:
given than the pharhelia has about 1.8* the memory bandwith of the Gforce. a) makes me wonder about the "suposed" advanced architecture of the pharhelia ?

Then it also strikes me that the pharhelia is using a very clever way of AA.
does this FAA involve less calculation than the FSAA ?
wouldnt it be itresning to compare Pharhelia with 4*FSAA and the GF4 4600 with 4*FSAA

is it insane to thinks that FAA algorithm could be done on other chip ? => assuming all the above would it be wrong to imagine that A GF4 chip with a FAA algorithm will be faster than a Pharhelia.



Please I'm not posting to start a flame neither for you to tell all they other pocibilitie offred my the Pharhelia, making a fps only comparation a nonsence. I 'm really enthousiastic about matrox, I just had those cold questions without answers running beheind my back this afternoon...

[thop]

Re: Parhelia - no flame - makes me wonder

Originally posted by notagain
a) -we have the Parhelia 10 to 20% slower (fps wise) than a (let say) GF4 4600 when playng without AntiAlising
b) -on the other hand we have a Parhelia runing with 16x FAA , 30% to 60% a GF4 4600 with 4 FSAA.

i'm just curious, but where did you extrapolate this numbers?
and boy, you are gonna be sooo flamed just joking

but that is a interesting question. will ATI with the R300 f.e. license (or whatever you do; edit: i meant steal ) FAA from Matrox?

[Cheesekeeper]

The parhelia does indeed use very clever methods to perform FAA, and memory bandwidth really doesn't come into it. In Full-Screen Anti-Aliasing, the entire image is rendered at the higher resolution (in this exaple 4x display res), and is then scaled down to match screen res. The card has to perform almost as many calculations as would be required to display at the higher resolution.

The Parhelia's Fragment Anti-Aliasing does not render the entire frame at the higher resolution. It works out which parts of the image represent the edges of objects (the only places where stepping is likely to be a problem), and only those parts are rendered at the higher res (16x display resolution). This not only allows it to perform its AA with a minimal performance hit (only a tiny amount of the image is affected by the AA algorithm), but also avoids blurring the textures appearing within objects - one of the biggest problems with FSAA (since only the edges are re-rendered, the surfaces aren't touched).

If I recall correctly, the Parhelia still takes a pretty major performance hit at 4x FSAA - same as any other card, and for the same reasons. 16x FAA is a nice bit of lateral thinking, allowing them to sidestep the problem, rather than hit it head on by applying more horsepower......

[Jammrock]

The final fps numbers for the Parhelia are unknown. Most of the reviewers that have mentioned they have one, have posted screenshots of a Beta board (I think some have a final board?). Either way, the drivers are still very beta, as the BBz (Beta Boyz) have mentioned they are still very engrossed in testing and the drivers get better every revision. So any fps number you get now is beta. Whether the final numbers will break Ti 4600 numbers or not is unknown. We will have to wait for June 25 at the soonest to see official numbers.

The advanced memory arcitechure is designed so your can run your games at high res, all eye-candy turned on, 16x FAA and still have the game run smoothly. Matrox simply chose to use a wiser hardware arcitechure instead of hardware 'tricks' that other companies use. People that don't care about the looks of their games and only want fps should stick with nVidia. So far the only thing the BBz have said about performance is that it runs 'smooth' at high res with all eye-candy turned on. That's something even the Ti4600 struggles with.

Jammrock

[Pace]

Parhelia offers a lot more than FAA to increase framerates. We'll have to wait on benches to see, but I'm sure if you try a game at UXGA or higher res', Parhelia will come out the winner compared to GF4. R300 however, I'm not so sure!

P.

[Technoid]

Who care's about FPS if its smoth?

[notagain]

technoid

60fps today is fine i agree
but 60fps today also means 30 fps in 3 or 4 month.

ie my real fear is that: in fact the only stuning fact about the card is in fact a "Algorithm".

that R&D &spy from other compagny will 'easyly' implement as well ...

to put those into words: you take a Gforce 3 you ask it to FAA inseat of FSAA.. and you got a Gforce 3 16*faa faster than a Gforce 4 4*fsaa...

MY BIG Q is: IF the required processing power required to do FAA was higher than the on to do FSAA (and therefor FAA could NOT be done one any Gforce chip) THE Parhelia Should be faster than the Nvidia chip in all cases.
As it doesnt seems to be the case... my conclusion is that FAA could be performed on Nvidia chips and will outperform Matrox's.

THE BIG Assuption I make is that the instruction set of NVIDIA chip allow to do the math for FAA. I have absolutly no idea
what are the instruction required for FAA nor FSAA.
Futhermore I have no clue on haw programmable GPU are...

I'm seeking answers.... that was what my post was about.


regards

[Umfriend]

I know nothing (been to Barcelona once), but I think that GF3/4 or any other card to date could only emulate FAA in software, which should be slower. Parhelia would have most/all fo the logic in hardware.

Hence, FAA on a GF3/4 would mean a far greater hit on performance. Without any AA, GF3/4 might still be faster.

Assume for instance a P4 @2.2Ghz, but without the Floating Point Unit. It would be way faster than a n AMD Athlon @ 1.0Ghz in integer operations, which both do in hardware. But as soon as floating point operations come into play, the Athlon will hardly take a hit: it's in hardware. The P4 in this case however would have to emulate this: use integer operations to get the same results. Loads of integer operations to mimick what the other does in HW.

Best I could do

[Wombat]

but 60fps today also means 30 fps in 3 or 4 month.

Not necessarily. My G400MAX was completely usable for <B>2 years</B>, and some people are still running theirs after 3. You can't really compare architectures. I was playing CS at 1024x768x32 at 35-55fps (which looks better and smoother than my GF3 at 70+ fps), and ran games like Deus Ex and Citizen Kabuto just fine. The G400 came out, then outlasted the TNT2, TNT2Ultra, GF1, GF2.....

[Pace]

I know more than Umfriend (been to Milan once and Paris once ).

Is what you really want to know how hard it would be for NVIDIA/ATI to add FAA to their future products? I think only time can answer that one, unless Matrox tell you they have it patented or something. If they don't, then NV/ATI could add a similar feature, possibly with a better implementation. However, Parhelia is the best at FAA for the moment

Plus its HDM, GigaColor, TriH, RSN, DH-HF etc...

P.

[xortam]

Originally posted by Pace
I know more than Umfriend (been to Milan once and Paris once ).

You know Umfriend was speaking as Manuel?!

[notagain]

"I was playing CS at 1024x768x32 at 35-55fps (which looks better and smoother than my GF3 at 70+ fps),"

I'm having hard time to agreeing with this statement

"any other card to date could only emulate FAA in software, which should be slower. Parhelia would have most/all fo the logic in hardware"

this assume that the AA are hard wired function and not 'pieces of program' that can be reprogramed.

This assumtion might actualy be true at least for ATI , as they care to mention that there "new" Chip "R300" will have "programmable" AA filter.

But Then again, I guess we have no evidence that it is the case.. I had the occasion to talk years ago with Nvidia representative. (it was during the 3DFX voddo 5000 launch), he told me that the Gforce (1) chip had from the start FSAA capacity but they werent able to enable it . then the Gforce 2 when it came out did not do FSAA, only years later with new drivers has this option be enable...

this leads me to beleave
1) Anti Alising is not hardwired in the Nvidia chips, Hardwired are set of instruction that are use by the FSAA, Never the less a certain amount of Mhz, and of bandwith is required to perform those heavy calculation.
2) things remain to be seen if the instruction used in the nvidia chips can be use in FAA



BUT as you said the exemple with integer and fpu is spot on...


any one has experiences on haw GPU are programmed ?

[Wombat]

I'm having hard time to agreeing with this statement

Then that's your problem, but you'll find a number of uses that will say the same thing. nVidia knows that people rely on FPS numbers, and does all sorts of things to boost them. Doing a 360 degree rotation in Deus Ex looked a lot better on my MAX at 32fps than it does on my Ti200 at 56fps. Also, my G400 had a much smaller delta between min and max frame rate, and it's definitely minimum frame rate that matters the most. And that doesn't even touch on how much better Matrox's colors are.

[Greebe]

All I have to say is your wrong. Features that don't work can be disabled either in hardware or software and unless you were told it was in there wouldn't know it exsisted. Case 'n Point, the G550's TnL unit. We know it's there but it's not used, there is no reference to it in the drivers...

nV has never claimed to have this ground breaking feature (FAA) and we all know how they like to pimp their stuff. So needless to say this assumption on your part is totally baseless.

[TdB]

any one has experiences on haw GPU are programmed ?

I don´t have any experience, but I assume that you program the gpu with pixel/vertex-shader programs(assembler-like language), and that it is executed IN the gpu, and not in the cpu.
for example: with pixelshaders you can make your own light-reflection-algorithm, and apply it to a texture, and then the algorithm will be executed on the gpu, this eliminate the need for alot of bandwith between gpu and cpu, because it is performed locally in the gpu.

regarding your faa question, i don´t think the any gpu is flexible enough to do this(with the possible exception of the 3dlabs P10), most gpus have VERY limited flexability.