The monster is out !!! (who can stand against the mighty NV20?)

[Fozzie]

ARRG. You obviously don't understand what I am saying and if you choose not to read my posts through thats your problem. Ask Dave if what I am saying is correct or not. In summary, if you had a tiler with similiar technical specifications it would be more complex then a traditional of the like design(assuming no one has radically more efficient pipeline designs and whatnot). The tiler would be faster then a Traditional without external frame-buffers and "HSR" methods in many situations however.

"Actually GP-1 is from TNT/Banshee days. So even the raw spec was pretty much the same."

I'm aware of that Dave, my last statement wasn't very clear. I was trying to get across WHY all current tilers have a smaller gate count compared with Traditionals of the same timeframe. I was not saying anything in regards to performance of tilers in general. I think 3dfx has some hot technology in the basket now, if they can bring it to market in a resonable time frame(eg before nVidia and ATI can field 3-4MB+ integrated frame-buffers) and at a competitive cost it will be quite incredible.

 

[pidge]

Yeah. I remember seeing the GP1 at GDC this past year. It was about 10-30 FPS faster than a Geforce DDR on Quake III Arena. I saw a lot of promise in that chip, especially in the laptop sector since the slower memory and core speeds would allow it to consume a lot less power and cost less to make. It was a very nice chip.

 

[Blackhawk2]

Fozzie, I understand what you are saying and I've read your posts. I don't agree. GP-1 is much more powerful (going up against a Voodoo5/Geforce2) than a TNT/Banshee and has a much lower transistor count.

<<...3-4MB+ integrated frame-buffers...>>

Fozzie 3-4MB isn't enough, this isn't the Playstation2 that plays at 640x480 or less(I heard it was at an interlaced resolution of 320 by 240 or something like that, it was a very low resolution). At least 16MB of embedded ram would be needed to fit the entire frame buffer, back buffer and some game textures in memory, assuming we are not playing at the Playstation2 resolution . If this was not done you would be limited to external memory speed/bandwidth used to access the missing information therefor negating the effect of the embedded memory->aka speed/bandwidth. 16MB or more of embedded memory would be expensive, that is why Sony could only afford 4MB in the PS2 and they didn't need to go to a resolution of 1600x1200.

NV20->$500+, 50-60 million transistors, 15 watts of heat (might require special cooling system says one of the board manufacturers)

GP-1 (note GP-3 with T&amp;L announced, possibly GP-4 in production) ->$Unknown, 3 million transistors, unknown wattage...3 million vs 50-60 million transistors. It is beating Voodoo5 &amp; Geforce2 with 4XFSAA and the card is about as old as the TNT/Banshee.

You figure it out.

 

[DaveB3D]

The only way for a traditional architecture to match a deferred renderer in performance is to store the entire frame-buffer in embedded memory. So consider when they could possibly do this, let us assume 1600x1200 with 4x AA and 64-bit color. You are going to need a 64 MB frame-buffer. So 64 MB on a chip.

 

[ElFenix]

ATi will have 1 transistor sram embedded in the gamecube's graphics chip. wonder how that will perform?


as for dave working for 3dfx... he hasn't been there very long. he was an extremely knowledgable outside observor prior to that. if he didn't think that 3dfx was going to survive (thrive?) he wouldn't have gone to work there. what is the point of working at a company that you are pretty sure is going to be dead within a year? so dave's employment should be seen as a vote of confidence in 3dfx's future.

 

[Fozzie]

&quot;The only way for a traditional architecture to match a deferred renderer in performance is to store the entire frame-buffer in embedded memory. So consider when they could possibly do this, let us assume 1600x1200 with 4x AA and 64-bit color. You are going to need a 64 MB frame-buffer. So 64 MB on a chip.&quot;

Dave, keep thinking that Embedded ram will be impossible in the next few years for Traditionals. Keep thinking that, I hope all of 3dfx does with you.

Blackhawk, again you apparently DO NOT understand what I am saying in regards to complexity. There is no more point in continuing this discussion as far as I am concerned. I say at this point to Dave if you understand what I was saying, do you agree or disagree?

 

[DaveB3D]

I never said embedded memory was impossible.. I said the needed ammount of embedded memory is impossible unless you want your chip to cost 200 and your board to cost $1000. Sure, boards will have a few megs of emdedded memory as a cash.. but that won't do anything close to what deferred rendering can do. So yes, I will keep thinking that because it is an unquestionable fact.

and yes, I know what you are saying...

 

[ElFenix]

isn't quite a bit of the vsa-100s 13 million transistors cache?

 

[DaveB3D]

I have no idea the exact amount, but we aren't talking about the kind that is associated with emdedded memory.

 

[Tung]

doodie forever!