NVIDIA GT200 Rumours & Speculation Thread

from TG daily, i linked above:

referring to the memory controller

GT200 looks to be a monster! i just somehow know they will break 1 TFLOPs with their top performer
.. it is marketing-driven to do so, as a "milestone" for a single GPU! [imo]

 

Calculate based on what if you don't mind me asking?

 

I remember when the term MT/s used to mean millions of triangles per second. Amazing it now means millions of textures processed per second

 

I think that was supposed to be a joke.....

Given that Nvidia promised a teraflop I'm thinking this is feasible:

16 clusters
2 x 8 SP multiprocessors / cluster
3 flops / SP (explains the +50%)
8 TF / cluster
500-600Mhz core clock
1300Mhz shader clock

Geforce products with 15 clusters enabled - 930 Gflops
Tesla products with 16 clusters enabled - 1 Teraflop

Of course we still have no idea how the clusters are really configured or how clocks will turn out....

 

Uhhhmm.....

I definitely need sources to speculate of any X amount of anything

 

I'm extremely bad with any bets, but my gut feeling tells me that the [10*(8*3)] theory makes more sense.

 

just throwing it out there but how different is the 260 from the 280 besides the 448 bit bus...
and when they shut down the cores couldn't you theoretically "turned on" them or do they do something to prevent this?

 

Search for softmod. If the circuits are physically disabled via laser, then there is nothing you could do.

 

dude did you join just to answer that?, anyway your talking software solutions aren't there any hardware solutions to undo it?, like physically going to the graphics card



(btw junior member FTW!!!!!!!)

 

He already answered your question...

If it's physically disabled, you aren't going to re-enable it. If it were simple, it would be done. There's a reason 8800 GTS 640 cards couldn't be magically transformed into essentially 8800 GTXs with less RAM.

 

ummm first off no not entirely he gave me a software option and never said anything about hardware
and second i know that im just talking could it be done period i don't remember saying that i wanted it to be easy....

 

Physically disabling units with a laser (or blowing an e-fuse by other means) rules out hardware methods of modding the hardware back--or at least it does for those without an electron microscope and the hardware lab needed to redraw lines in the silicon.

 

there we go thank you thats all i wanted to know so it costs more to fix the problem than it does to just buy the higher end product

 

The other post mentioned the use of a laser, or at least did when I read it.
I just elaborated a little.

Chip manufacturers went to greater lengths to prevent softmods and things like jumper or resistor changes in later chips.

 

yes but the little bit of elaboration you provided answered my question completely

yep technology has come a long way
i still remember when 16mb of ram was godly

 

Those answers tend to come out after the cards are sold and people are able to test them. Hence, the soft mod or laser cutting answer. Also, you are very patronizing in your replies.

 

yeah i feel bad about that... little frustrated this morning didn't me to take it out on you guys

 

http://forums.vr-zone.com/showthread.php?t=278769

 

those specs are pretty good seem pretty realistic
1.1 TFlops sound right

 

Hmm... So, it's 24 SPs per eight TMUs, e.g. dual 12-wide SIMD per cluster?