NVIDIA Kepler speculation thread

My bet on the die-size is something in the range of G80 and GT200b, if it is a 512-bit device with 16 multiprocessors.
 
I have the feeling, nothing new there... i think we all wait a "refresh" or new Kepler for September, october so far.

I hope a bit more of 25% still. ( question is more TDP, base/boost clock speed if we speak about a double mem controller )..

Look like the end of the year will be interesting.
 
20-25% faster than GK104 sounds about right. Will probably be a beast at compute though.

I wonder how can it be a beast at compute and end up only =/>20% faster than a GK104. Unless of course NV's engineers are truly a bunch of drunken monkeys (Arun(tm)) and it has a 1:1 DP/SP ratio.

Any rumors on when the 670, 660, 650 and lower GPUs will launch?

May?
 
I wonder how can it be a beast at compute and end up only =/>20% faster than a GK104. Unless of course NV's engineers are truly a bunch of drunken monkeys (Arun(tm)) and it has a 1:1 DP/SP ratio.

Most of the compute speed enhancing features won't help gaming speed much if at all, but take a crapload of space from the die, after which you have only so much extra space on the silicon to get more gaming speed jammed in, too?
 
I wonder how can it be a beast at compute and end up only =/>20% faster than a GK104. Unless of course NV's engineers are truly a bunch of drunken monkeys (Arun(tm)) and it has a 1:1 DP/SP ratio.

I guess you answered the whole GK104 vs. Taihti here.
 
I wonder how can it be a beast at compute and end up only =/>20% faster than a GK104. Unless of course NV's engineers are truly a bunch of drunken monkeys (Arun(tm)) and it has a 1:1 DP/SP ratio.

Well for starters it seems nVidia is much better at separating compute and gaming performance this round. It's very possible that they could address GK104's efficiency issues without a large increase in theoretical maximums.
 
Good for gamers who don't want those monsters which do nothing they care about.
Good for people who want to have massive parallel processing units to do whatever they like.
 
Well for starters it seems nVidia is much better at separating compute and gaming performance this round.
Given the recently rumored performance and die size ballparks of GK110, I still don't see what's supposed to make the relation between GK104/GK110 so much more impressive than what AMD did with their Pitcairn/Tahiti split?

Again: AMD's compute chip is roughly 70% bigger and about 30% faster than AMD's gaming chip - and a lot of people say that this kind of scaling sucks. Now Nvidia's compute chip is rumored to be ~70% bigger and ~30% faster than Nvidia's gaming chip - and people say NVidia is "way better at separating compute and gaming performance this round"? I don't get that point.

You might argue that Nvidia was more lucky with their overall die size targets when it comes to absolute performance positions. But then, you might also consider that AMD's "big chip" was launched more than three months ago - and Nvidias "big chip" won't show up until autumn.
 
Well for starters it seems nVidia is much better at separating compute and gaming performance this round. It's very possible that they could address GK104's efficiency issues without a large increase in theoretical maximums.
It will be very interesting to see if they're going to do the traditional thing of using the same exact architecture for compute and gaming, or whether they've decided to produce a compute-optimized chip that is more like the previous architecture.

Otherwise they'll need to do a huge amount of work on their compilers to get a compute version of the GK104 to compete.
 
It will be very interesting to see if they're going to do the traditional thing of using the same exact architecture for compute and gaming, or whether they've decided to produce a compute-optimized chip that is more like the previous architecture.

Otherwise they'll need to do a huge amount of work on their compilers to get a compute version of the GK104 to compete.

GF100 and GF104 were based on slightly different architectures. I suspect the differences will be even more marked this time.
 
Back
Top