The GT300/Fermi Thread - Part 2!

[Sly Fox]

Not true. It's PC as well.

Although in fairness id has been pretty clear that they want it to run excellently on all platforms.

It's just too bad Carmack is so busy with his aerospace stuff, or Doom 5 would probably be out by now!

Who cares about space anyways, video games where you blow stuff up are far more important.

[saaya]

It's just too bad Carmack is so busy with his aerospace stuff, or Doom 5 would probably be out by now!

aerospace stuff? :o
like what? didnt know he was into that...

and carmack... his genius is completely wasted on getting "ok-good" graphics on consoles... im not even sure hes good at that, hes good at pushing graphics to the limits, not on getting the job done with the least possible resources...

Fermi GPU costs more than a 5870 (price can be anything)

yes, but depending on yields it kinda doesnt matter anymore...

rv870 = 100% yields
150/150 = 25us$

GF100 = 33% yields
33/100 = 100us$

GF100 = 50% yields
50/100 = 70us$

GF100 = 66% yields
66/100 = 53us$

at 50% yields the difference in chip cost is already acceptable... 50$ isnt much, and consiering that fermi is 50% bigger and SHOULD be at least 20% faster, nvidia can easily charge 50$ more for fermi than an rv870 part... so that means if fermi hits 50% yields AND perf is 20% better or even more, nvidia will make as much money selling GF100 as ati does selling rv870...

nvidia clearly never aimed for high yields with fermi, they knew it wouldnt be good, but they didnt expect the claimed single digit yields, which is a huge deal breaker for them...

40% more Transistors, ~64% more die size.
Yields aren't comparable between two differenz chip designs, so costs are also not comparable. Different contract details with TSMC.

why arent costs comparable?
you take the amount of money transferred from nvidia to tsmc and divide it by the functional parts nvidia received, and the same for ati... and voila... was it that hard?

and different contract details... afaik thats not true, both nvidia and ati buy wafers and not functional chips. at least thats what an article i read mentioned... was it anandtech?

[speculation] about the manufacturing contracts, i have a feeling that tsmc is giving nvidia a better deal per wafer than it's giving amd. tsmc knows that amd will be going to gf fairly soon, so there is no insentive to lower it's prices to retain the customer. nvidia however, has the option to go either way, so i beleive it's safe to assume that nvidia is getting some good prices in an attempt to keep them happy with tsmc. [/speculation]

yeah but how much of a difference would that make? lets say tsmc gives them 50% off, which is a huge discount, even then they need ~20% yields to get to a point where they can get the same margins as ati... (again, assuming GF100 is at least 20% faster than rv870 so they can sell GF100 for 50$ more)

if the rumors of single digit yields are true, then even a massive discount of 50% wouldnt help nvidia...

Given my background is in competitive first person shooters...

...I actually play them with a 360 controller

0_o mmmmkaaaaay... hehehe

[Sly Fox]

aerospace stuff? :o
like what? didnt know he was into that...

and carmack... his genius is completely wasted on getting "ok-good" graphics on consoles... im not even sure hes good at that, hes good at pushing graphics to the limits, not on getting the job done with the least possible resources...

Check out http://en.wikipedia.org/wiki/Armadillo_Aerospace

If there's ever been someone in the game business who I'd call a legitimate genius, it's Carmack.

Which is to say, that I agree wholeheartedly with you.

[xdan]

700+ MHZ frecuency of the core is a myth..
GEFORCE GTX 480 will have 48 ROP not to "have more than Ati " but to compensate lower GPU frecuency. And the power consumption will not be so bad as predicted.
625-650 core frecuency more reality.

[annihilat0r]

700+ MHZ frecuency of the core is a myth..
GEFORCE GTX 480 will have 48 ROP not to "have more than Ati " but to compensate lower GPU frecuency. And the power consumption will not be so bad as predicted.
625-650 core frecuency more reality.

What about SP count?

[xdan]

What about SP count?

More probably after or during CEBIT. NDA is very respected.
ROP's are a good sign, GTX 480 will have 50% more ROP than a 5870 which may compensate very much other lacks of performance.

[mindfury]

More probably after or during CEBIT. NDA is very respected.
ROP's are a good sign, GTX 480 will have 50% more ROP than a 5870 which may compensate very much other lacks of performance.

GT200 has 100% more ROP than Rv770/rv790,but not much faster...

[saaya]

ROP's are a good sign, GTX 480 will have 50% more ROP than a 5870 which may compensate very much other lacks of performance.

may? so are you speculating or do you actually have a source?
and according to several articles i read nvidias rops are chewing through data about half as fast as atis?

Still i consider that 1600SP's of 5870 suffered a bit of too few ROP's...that's way it didn't beat GTX 295.

then how come the rop castrated 5830 scales pretty well according to sps?

My guess was that this was due to the fact that Fermi wasn't released. Of course I could be wrong.

oh enough with this bs speculation already!!!

hahahha

he builds rockets

wow.. thats cool...

living under a rock? Carmack loves the challenge of pushing as much pixels as possible, no matter what the hardware
http://www.gamasutra.com/php-bin/new...hp?story=23997
:-)

yes jmke, im living under a rock... sigh...
and thanks for proving my point by linking to carmack spending time to get (how many years old?) doomI running on a 5" screen mobile phone instead of cooking up a crysis eye candy killer

Anyone know the length of the 480?
Id hate to buy a case it doesn't fit in..

i heard nvidia is doing some special fermi verification process for cases? check nvidia.com or google, shouldnt be too hard to find infos... i remember silverstone and coolermaster launched new cases and they said it was fermi certified...
whatever that means...

Too many baseless claims...

When i tried to cast a light on this problem, several chose to use personal attacks to offense me. I'm wondering why it is so hard to keep this thread at a level which this forums reputation would deserve. This is something that can affect the credibility of the entire forum.

I don't know if they (obviously faked fermi benchmarks) are real, but the guy is calming that nVidia has provided the info, and in case they matter a lot . Who told you they are fake? source please.

stop! please stop!
your killing me man!

http://www.youtube.com/watch?v=MOZDG9UgzkQ

[jmke]

aerospace stuff? :o

he builds rockets

and carmack... his genius is completely wasted on getting "ok-good" graphics on consoles... im not even sure hes good at that, hes good at pushing graphics to the limits, not on getting the job done with the least possible resources...

living under a rock? Carmack loves the challenge of pushing as much pixels as possible, no matter what the hardware
http://www.gamasutra.com/php-bin/new...hp?story=23997

:-)

[weston]

yes, but depending on yields it kinda doesnt matter anymore...

rv870 = 100% yields
150/150 = 25us$

GF100 = 33% yields
33/100 = 100us$

GF100 = 50% yields
50/100 = 70us$

GF100 = 66% yields
66/100 = 53us$

at 50% yields the difference in chip cost is already acceptable... 50$ isnt much, and consiering that fermi is 50% bigger and SHOULD be at least 20% faster, nvidia can easily charge 50$ more for fermi than an rv870 part... so that means if fermi hits 50% yields AND perf is 20% better or even more, nvidia will make as much money selling GF100 as ati does selling rv870...

nvidia clearly never aimed for high yields with fermi, they knew it wouldnt be good, but they didnt expect the claimed single digit yields, which is a huge deal breaker for them...

where did you get your pricing on gpu dies? each wafer can fit roughly 94 dies and costs $5000. if fermi get's 40% yield it's $131/gpu. at a similar defect rate, ati would get roughly 70% yield which is about $48/gpu. Add in extra cost for PCB of fermi needing more memory chips, wider bus, more connections, etc...you end up with well over $100 more per complete card.

http://www.brightsideofnews.com/prin...-analysis.aspx