NVIDIA Kepler speculation thread

[phenix]

And over-powered, and hot, and...

Well, if it really is 2.5 times more efficient than Fermi, it might turn out to be a cool chip.

 

[Alexko]

And the hardware of their first Tesla generation was ready in November 2006 or April 2007.

This roadmap is about chip codenames, not the Tesla product line. Here, Tesla means GT200.

 

[digitalwanderer]

Well, if it really is 2.5 times more efficient than Fermi, it might turn out to be a cool chip.

That'll be great if it is, and the idea of a new line-up in January from nVidia is a big plus too.

 

[Sontin]

This roadmap is about chip codenames, not the Tesla product line. Here, Tesla means GT200.

"Cuda GPU Roadmap". And which was their first Cuda architecture? Hint: Not GT200.
They used Tesla for G80 GPUs. And before Fermi they never used names for their chips.

 

[CRoland]

So do they plan to abandon the approximately yearly release cycle of the good old days, or will there be some kind of refreshes in between?

 

[OlegSH]

And before Fermi they never used names for their chips.

ORLY?

# nv10 - celsius
# nv20 - kelvin
# nv30 - rankine
# nv40 - curie
# nv50 - tesla

 

[no-X]

"Cuda GPU Roadmap". And which was their first Cuda architecture? Hint: Not GT200.
They used Tesla for G80 GPUs. And before Fermi they never used names for their chips.

In this context Tesla = GT200. G80 cannot be present in DP roadmap ;-)

 

[iMacmatician]

So do they plan to abandon the approximately yearly release cycle of the good old days, or will there be some kind of refreshes in between?

It looks like there will be a refresh between architectures.

Something else of interest:

"Between now and Maxwell, we're going to Introduce features like virtual memory. We're going to ENHANCE the GPU's Ability Thurs autonomously process, so it's less dependent on the CPU, along with a very large Improvement in performance. "

 

[MarkoIt]

It looks like there will be a refresh between architectures.

Something else of interest:

http://www.fudzilla.com/graphics/graphics/graphics/maxwell-to-fermi-is-40-times-faster.

Kepler to be 3x-4x more efficient in DP than Fermi, and 10x faster (probably in SP?) aka almost 10 Teraflops
Maxwell to be 16x more efficient in DP than Fermi, and 40x faster (SP)? aka almost 40 Teraflops.
If that is true, they will have to change an awful lot the architecture.. and beat by more than double both perf/mm^2 and perf/watt of ATi's current architecture.

 

[aaronspink]

I'm happy to learn that Fermi was released in 2009, and that Tesla was released in 2007. Kind of makes me think about how reliable that "2011" will turn out to be for Kepler…

Oh I see why you are confused. That's not release dates. Those are the dates that they realize it will need to be delayed for at least another year!

 

[UniversalTruth]

.. and beat by more than double both perf/mm^2 and perf/watt of ATi's current architecture.

Kepler will have to compete against the successor of NI, not the current one.

 

[CarstenS]

Oh, I love how Nvidia tells us time and again where they will be at some point in the future. Especially when in a situation like the current one.

 

[MarkoIt]

Kepler will have to compete against the successor of NI, not the current one.

If AMD can keep up with the paces it had for the last 3.5 years, by fall 2013 they will deliver something close to a 15 Teraflops SP chip (aka 3 Teraflops DP chip.. similar to Maxwell DP efficiency actually)

 

[UniversalTruth]

I think your calculations are a little bit wrong.

Kepler to be 3x-4x more efficient in DP than Tesla, and 10x faster (probably in SP?) aka almost ? Teraflops
Maxwell to be 15x more efficient in DP than Tesla, and 40x faster (SP)? aka almost ? Teraflops.
If that is true, they will have to change an awful lot the architecture.. and beat by more than double both perf/mm^2 and perf/watt of ATi's current architecture.

I think it looks better now. The other things you can fix yourself...

 

[Blazkowicz]

are those tape-out dates?

 

[PatrickL]

Well at least theses slides are very cheap to make and never a delay

 

[fbuffer]

I think your calculations are a little bit wrong.





I think it looks better now. The other things you can fix yourself...

except (unless my memory/math is truly THAT bad) Fermi increased Tesla DPFP effectiveness by over 4x already..

tesla (GT200b Cuda 1.3 78GFlops at 190W) = .42 GF per Watt
fermi (GF100 Cuda 2.0 515GFlops @ 247W) = 2.09 GF per Watt

heck Tesla -> Fermi increased GF/Watt effectiveness almost 5x (4.98). I do take the slides (and almost anything from nV given their past history) with a ton of salt however.. the performance increase of 3-4x over Tesla is already present. Do I think we'll see another 4x increase in the next year.. nope.. 4 years.. possibly.

 

[MarkoIt]

I think your calculations are a little bit wrong.





I think it looks better now. The other things you can fix yourself...

I just took it from fudzilla.. and they reported it from Jensen's speech (oh wait, that doesn't make it more believable? :smile

 

[DarthShader]

Well, if it really is 2.5 times more efficient than Fermi, it might turn out to be a cool chip.

This might be a good moment to pull out what has been promised for Fermi. You know, 150W, high clocks, etc.

 

[caveman-jim]

are those tape-out dates?

basic architecture announcement dates?