Big Kepler to have 7 billion transistors??

[CAT-THE-FIFTH]

It seems to be mentioned on this webpage:

https://registration.gputechconf.com/?form=scheduleBySession

"S0642 - Inside Kepler
Stephen Jones ( CUDA Developer, NVIDIA )
Lars Nyland ( Senior Architect, NVIDIA )

In this talk, individuals from the GPU architecture and CUDA software groups will dive into the features of the compute architecture for “Kepler” – NVIDIA’s new 7-billion transistor GPU. From the reorganized processing cores with new instructions and processing capabilities, to an improved memory system with faster atomic processing and low-overhead ECC, we will explore how the Kepler GPU achieves world leading performance and efficiency, and how it enables wholly new types of parallel problems to be solved.

Topic Areas: Parallel Programming Languages & Compilers
Session Level: Beginner"

 

[Jokester]

What's full fat Fermi (580)?

 

[CAT-THE-FIFTH]

The GF110 was meant to have 3 billion transistors although the GF100 was meant to have 3.2 billion transistors. I am not sure which figure is correct as some websites say the GF110 has 3.2 billion transistors too.

 

[555BUK]

GF100 (GTX480) had 3.2bn, GF110 (GTX580) was slightly smaller at 3bn.

I cannot imagine Kepler GK110 having 7bn transistors. That will require a bigger die than GF100 and would be extremely expensive to produce. I think anywhere between 4.5bn to 5.5bn is much more realistic.

If it does end up with 7bn transistors (almost twice as many as the GTX680) it will be an absolute monster, in every sense of the word.

 

[lsg1r]

To be honest I don't think it will make much difference to 90% of the people on this forum.

From everything I'm reading it's looking like nvidia are now making a concious decision to nerf the cuda processing of the geforce cards in an effort to push tesla and quadro (and bigger profits per gpu's). So the likelihood is that this 'big kepler' won't even be in geforce cards.

It used to be a simple case of pick a geforce and you'd get the same sort of cuda performance per cuda core at the same clocks as a tesla/quadro, it just run a bit hotter and cost about a 1/3 of the price.

As shown with the gtx680, it's lost a lot of performance compared with the older fermi generation in cuda performance. In most cases they're saying buy a geforce 580 over the 680 if you want to use cuda as it performs better.

 

[LtMatt]

Its going to need at least 8 billion transistors imo.

 

[Rroff]

They may be (from the topic) talking about a compute orientated part (tesla, etc.).