D9P/G94: 9600 GT on 19th of February

I believe that G9600 GT will deliver the on-par performance with 3850.

However, it depends on product variance.

G9600GT seeems to be a successor to 7600GT.


Happy Christmas to everyone.

 

Oops, missed the die size on the first page. Ok, so how have they managed to squeeze ~equivocal performance to G92 out of the same manufacturing process with only ~60% the die size?

 

Is it possible with only 64SPs without some architectural improvements?? Hmm unless SPs will be clocked about 3Ghz but i doubt it.

 

You seem a little confused right now. You just posted a question in reference to 8800GT, which Arun answered to the point. Now you are asking (and with it, changing the subject) how they managed that same level of performance with an ASIC that is a lot smaller in size than G92. The answer to both questions is that for the performance of a 8800GT, you don´t need a core with a die size of 320mm² (G92), which should be pretty obvious right now.

The margin thing, remember?

 

Presumably higher clocks (I'm willing to bet 500MHz really applies to 8800GS, not 9600GT) and simply being smarter. And you also need to consider G92/G98's die sizes and transistors counts. VP2/NVIO/extra TA should be cheap, so 754M vs G80's 681M despite fewer ROPs is fairly screwed up.

I've got a fairly simple and efficient theory to explain this, but I have got no idea whether it's right or not: several non-shader parts that were double-pumped in G8x are not in G92/G98. Case in point: G98 is rumoured to have a SKU with half the TMUs disabled, G86 never had that.

So why would D9P be that much denser? More time to do more parts in custom. Why would it clock higher? More time to do a better synthesis with experience from G92/G98. Why would it achieve 8800GT levels of performance with less texturing power? Because texturing obviously isn't the primary bottleneck on G92 and they made tweaks elsewhere.

Also, if it's on 55nm, then that's 'only' 20-25% smaller than G92. Which would imply the impressive part isn't the scaling but the IPC/unit improvements. However, I wouldn't even exclude the possibility of 65nm, given that D9E is apparently on 65nm and the timeframe is mostly similar... That'd make it 35-40% smaller.

 

Thanks Arun. Good point about the abundance of texturing ability with G92. NV could go back to G80-style TUs (at least unit count-wise) and save some die space which could be spent on DP & SM 4.1, but I still don't see how they could get such a significant reduction in size unless they've also cut ALU count in ~half and just clocked them higher.

 

Well, once again this is all assuming HKEPC's performance estimates are correct, which might not be the case. The question there is whether misc. architectural tweaks will improve IPC/unit enough, and whether the core clock is indeed 500MHz or if it's higher. What I'd expect personally is 32TA/32TF and 96 SPs, but who knows how things will turn out.

 

32TA/TF 96SP is exactly what I was thinking, but the die size still seems too small for 65nm...

 

Has 9600 a 200mm^2 die size? For 64sp it's quite big, i suppose. G92, with 128sp, has about 300mm^2 die size...I'm still thinking 9600 could have 96sp...
About D9E, 2*9600GT look quite good. That kind of cards , cheap for Nvidia, could beat R680 both in perfomance and in price.
Later in spring, maybe in April, they could launch G100: 256SPs

 

I believe that next generation high-end will be the Dual-core solution.(G100 or R700)

 

http://64.233.179.104/translate_c?h...conline.com.cn/topic.jsp?tid=8000957&pageNo=1

Do you think these benchmarks are true?? They said it`s GF9600GTS and if it really is i can say it`s performance isn`t impressive (non overclocked HD3850 level)

 

Maybe HDR/SM3.0 Score is to low for 64GB/s and 16ROPs@650MHz or?

But something other:

http://www.pcpop.com/doc/0/258/258133.shtml

Package with new lettering and different* look than the one from Expreview.

*other resistor layout and bigger die

 

Hmm, not very impressive benchmarks if true I guess, but makes sense given the die size if it's 65nm (which it probably is). Oh well. I wonder if GPU-Z's A11 & 650MHz have any basis on reality?

 

DigiTimes is saying D9P will be fabricated on 65nm from TSMC and 'feature' 64 stream processors.

 

Right now, I can't get the page to load in the original language, but what I want to know is whether the GTS is suppose to be better than the GT, or vice versa? I can't keep up with all these crazy names that change every generation between 2 different companies.

I know he's not sure if it's a GT or GTS, but until I can get the original language to load, I can't say much about it.

 

NV's nomenclature is ordered numerically and alphabetically. The higher the number, the higher the performance. Same goes for the suffix.
GS<GT<GTS<GTX<Ultra
It's really not as hard as some make it out to be, at least if you stop and think about it anyway.

 

*shrugs*

Then ATI suddenly comes in and shows my a X1900GT when I thought I finally figured out the placement of their products.

Anyway, I finally manage to load up the site and read the messages through. They're saying it's fake... the girl's breasts are fake that is. I had to go through 7 pages of nonsensical crap focusing on the girl than the benchmarks. Some claim she's a shemale. What little talk there is does not yield much in the way of information. What I did notice is that no one out right claimed the benchmarks are fake in all 7 pages. There was one guy saying how the benchmarks look just like the 8800GT.

My conclusion so far is that whoever posted those benchmarks could very well be a reliable source, thus no one would challenge his claims. After all, many of these cards are manufactured in China (just don't lick them), and someone being able to get their hands on hardware so early might not be so far fetched. Either that or they were too busy looking at fake shemale breasts to care whether the guy's numbers are real or not.

 

Who said anything about ATi? No one, until you did just now

Their nomenclature is even simpler than NV's. Purely a number system where higher number = more performance.

 

http://www.chilehardware.com/foro/informacion-exclusiva-sobre-t133896.html

No D3D10.1 and performance far behind 8800 GT? No good D9P...

 

If that slide true performance can be close to hd3850 according to this review
3 months later than hd3850, 9600gt not sound any good, looks like its just a lower cost competative product against hd3850.