Megadrive1988
Veteran
That no doubt seals Kepler's release date.
At GTC? I thought that it would be more a technology/architecture announcement than a release of a product.
iMacmatician
Regular
It's possible. They like to give devs/customers a chance to see the architecture before they release it. So maybe a late Q2 launch?
DarthShader
Regular
You got it wrong. They will be announcing the announcement of a release date of Kepler.
Are you ready? You got it wrong. They will be announcing the announcement of a release date of Kepler.
You forgot about the teaser 2 days earlier on Twitter and/or Facebook.
NVIDIA's quarterly earnings conference is on August 11, so perhaps they'll seize this opportunity to tell us at least in what quarter we can expect Kepler.
GTC is an HPC oriented event. First Kepler-based GPU launch isn't tied to GTC in any way. New Tesla generation is though.
Well, yes, that's why I suspected it'd be more of a technology announcement than a product announcement. I had a hard time reading up on what they announced at previous GTC's, though. Anybody have some info on that?
It was probably mentioned already but:
So delayed 'til Q2 next year. It smells like the rumors of Kepler facing trouble due to manuf. process and underperformance were true. The good thing is they seem to have learned their lesson from woodscrew gate and they don't want egg on their faces again
It will most likely be a product announcement with new Teslas being that product. And new Teslas will most likely come after Kepler-based GFs.
Well, since it was delayed, that makes a lot of sense.
DarthShader
Regular
http://www.xbitlabs.com/news/graphi...o_Start_Shipping_Kepler_GPUs_by_Year_End.html
So the 3x improvment claim is upheld.
So the 3x improvment claim is upheld.
So, assuming power consumption stays at about the same high level, we're seeing a roughly 2x improvement for process tech alone.
Now given that Gipsel (who's referring to Volkov's research) is right, that Fermi is stuck at 65% utilization for large matrix operations and Nvidia is able to remove that bottleneck, reaching the same efficiency as current Radeons of above 90%, then you could jump to the conclusion that there'll be nothing exciting about Kepler.
That, of course, doesn't take into account that a higher utilization usually makes power go up. So something's left to be improved by Nvidia still.
Now given that Gipsel (who's referring to Volkov's research) is right, that Fermi is stuck at 65% utilization for large matrix operations and Nvidia is able to remove that bottleneck, reaching the same efficiency as current Radeons of above 90%, then you could jump to the conclusion that there'll be nothing exciting about Kepler.
That, of course, doesn't take into account that a higher utilization usually makes power go up. So something's left to be improved by Nvidia still.
Yes, but power typically doesn't go up linearly with utilization. This is fundamentally because processors just aren't all that good at saving power when parts of the chip aren't in use. Sure, strides have been made in this area. But in general if you want good performance/watt, you want maximum utilization of the chip.That, of course, doesn't take into account that a higher utilization usually makes power go up. So something's left to be improved by Nvidia still.
DarthShader
Regular
TSMC mentions only 45% improvement: http://www.tsmc.com/english/dedicatedFoundry/technology/28nm.htm
Will nVidia improve the utilization for large matrix operations to over 130%?
Sorry, maybe I was being dense to assume you'd be with me naturally interpreting that as density improvement for going to 28 from 40, as TMSC states themselves in the link you gave.
Chalnoth,
No, not linear, but as more utilization happens, power would go up to some degree nonetheless. And especially in their higher end products Nvidia only has so much wiggle room left before they hit the 300 watt wall.
Chalnoth,
No, not linear, but as more utilization happens, power would go up to some degree nonetheless. And especially in their higher end products Nvidia only has so much wiggle room left before they hit the 300 watt wall.
Well, if 2x-3x HPC perfomance isn't exciting then I don't know what is.you could jump to the conclusion that there'll be nothing exciting about Kepler.
Plus Kepler should differ from Fermi more than just in density improvements. It's supposed to be more than a die shrink of Fermi.
How about 2x-3x game performance?
Who cares about DGEMM? Even in HPC it's far from being the end of it all.
I bet they are comparing Kepler to the old (broken) GF100, not GF110, which is 20% more power efficient. Move Fermi 20% down in the chart and Kepler won't be 3× more efficient, only 2,5×.TSMC mentions only 45% improvement: http://www.tsmc.com/english/dedicatedFoundry/technology/28nm.htm
Will nVidia improve the utilization for large matrix operations to over 130%?
Anyway, if I understand it well, these promises don't mean anything. Just like with Fermi - they promised some power consumption numbers before they had working silicon in their hands…
