What will replace NV50?

You mean with regards to ROPs?

 

erased

 

Why would they do that?
History has shown us that ati is into having huge fillrates for turning up the AA and AF, so if they did that, they'd need to have a nice clockspeed advantage, unless they would have more pipelines than the R520, but same number of ROPs, which would give them a shader speed advantage, but not so much a fillrate advantage

 

I prefer ATI's graphic cards to NVIDIA's
do you think singel 6800 ULTRA is better than X850?
do you think FX5950 ULTRA is better than 9800XT?
NVIDIA's graphic card's price always higher than ATI's

 

what the...

 

Doubtful. This configuration is killing AA performance which is not important for 128 bit NV43 but is important for 256 hi-end chips. It is possible however that NV48 (or 47 or whatever) will feature different from NV40 ROP configuration.

 

My thoughts exactly..

 

He's right on everything cept price.

 

Well..I think better choice to NV is when they will decide to push 16x2 architecture in NV48.
Why??As we know NV43 is 4x2, 2xGF6600GT(8x2) have performance little higher than X850xt pe and then you can imagine how fast could be GPU which will have 16x2 and clocked about 500mhz and memory about 1,6-1,8ghz???

 

Not gonna happen. It's a waste of transistors actually.

No it's not. It's 8x1 with the ability to write only 4 pixels to framebuffer per tact (8 in case of Z/S).

 

NV48 is likely to have 24 pipes and isn`t 24 pipes need the same number of transistors like 16x2???
I read somewhere that 32x1 needs about 20% more transistors than 16x2..

This is only my opinion and maybe i`m wrong....

 

16x2 marchitecture will be significantly worse than 24x1 in almost every pixel shader programm.

 

If it's a 16 pipe chip, I expect a similar ROP configuration as in NV40. If it's a 24 pipe chip, I stillexpect a similar ROP configuration as in NV40

How can questions be right? Besides, the quote is... weird.

Not if x2 and x1 relate to the number of (full) ALUs, not TMUs.

 

I believe we were talking about textures not ops. If there would be two full FP32 ALUs then it's all become a question of balance - do we really need that much PS processing power per clock on one pipe able to output one pixel or are our PS programs still rather simple and would benefit most from more simplier pipelines able to output more pixels per clock? There's no clearly better solution i think.

 

Those were rhetorical questions.
He is a fanatic..

 

Uh...no. Two 6600 GTs working in unison produce about the same performance as an X850 XT while running at low resolutions and no AA. Bump them up and the results become drastically different.

 

"We know" that up until NV3x there were two TMUs assigned per SIMD channel. Past that and more specifically on anything NV4x, there's only one TMU assigned to each SIMD channel.

NV44 = 4*1
NV43 = 8*1
NV40 = 12/16*1

2*NV43@SLi are more or less equal to a single 6800GT (give or take); the GT's real competitor is a X800PRO. I've not a single idea where it comes from that it could have a higher performance than a X850XT PE.

Example:

http://www.beyond3d.com/reviews/ati/r480/index.php?p=9

Albeit those are strict theoretical numbers, they DO NOT take under account that the maximum performance increase of two NV43 boards in SLI is up to 85% and not under all occassions:

X850XT PE = 540*16 = 8640 MPixels/sec
6600GT@SLi = 2*500*8 = 8000 MPixels/sec

(would the NV43 have 2 TMU's per channel, in SLi the maximum fill-rate would be at 16 GPixels/sec)

X850XT PE = 37.6GB/sec
6600GT@SLi = 2* 16GB/sec = 32GB/sec