NVIDIA Kepler speculation thread

Graphs are supposed to be informative based on their appearance. They are literally graphical representations of data. Bar graphs, in particular, are meant to show relative proportion. If you're going to start at a non-zero origin you need to at least add the squigglies showing you've chopped out the middle per CAD, etc.

 

The drivers for the memory interface are significantly larger compared to Pitcairn. IIRC Dave Baumann hinted they are meant to allow higher memory speeds and may be actually overspec'd to get there (it is actually quite easy for Tahiti to reach 7+ Gbps even with the 6 Gbps memory chips used).

???
Where did you get that from? The SRAM per CU is basically the same (registers, LDS, vD$) as long as we don't factor in ECC. Pitcairn even has relatively (to the CU count) slightly more sD$, I$, and L2 cache as well as relatively more (the same in absolute numbers) ROP (color and Z) cache and GDS.

 

Yeah I was being nostalgic (and my memory may be spotty), 9500->9700pro, x800 pro-->x800 XT, 6800-->utlra, 6200-->6600. Man that was just great times back then. I miss that. Very good value.

 

Definitely. I remember taking a X800 Pro VIVO and converting it to a XTPE for a friend of mine (I foolishly just bought the XTPE). It may be the the 6950 -> 6970 may be the last we see of something like that. But who knows, maybe something will fall through the cracks again in the future.

Regards,
SB

 

At least over here the X800 Pro Vivo was almost as expensive as the real X800 XT PE and they often struggled to reach XT PE speeds, mine was one of those, although it ran fine after the pencil mod. 9500/9700np to 9700 Pro and 6950 to 6970 were great stuff.

 

One more guess as to density differences. A larger chip might require an increase in the port count of the on-chip switch(es) connecting everything together. If those are cross-bars, then my guess would be that associated transistor count grows about linearly with port count, but that wiring grows quadratically. Assuming that's not way off, then perhaps the effect is non-negligible?

 

Ryan Shrout ‏ @ryanshrout
Stay tuned, we'll have a quick preview of Kepler GPU performance on the mobile site coming later tonight!!

https://twitter.com/#!/ryanshrout/status/179708075253567488

 

Nvidia GeForce GT 640M Review: Kepler Arrives For Mobile

http://www.pcper.com/reviews/Mobile...iew-Kepler-Arrives-Mobile/Performance-Verdict

 

384 CCs, 625 MHz core clock, and 1250 MHz shader clock? So hot clocks are in Kepler?

EDIT: BSN says "unified clock of up to 625MHz."

 

So it beats the outgoing Fermi in BF3 too? Suddenly a lot of stuff starts to make sense. Remember the "faster in dx11 titles". BF3 has been mentioned a lot obviously.

The point on the heat is most worrying.

Ivy Bridge is also supposed to be struggling with heat issues so what are both gonna be like together?

When is Nvidia going to learn that these are real issues? I can totally believe that Apple told them to GTF because of too much heat. It's just not acceptable in 2012.

 

This was an ultrabook review. It seems Acer decided to be aggressive with GPU clocks, but not with their cooling solution. I don't see how that's Nvidia's fault.
This review doesn't evaluate performance per watt directly, which is the key information you'd need to extrapolate to other laptops (including Apple's). An inadequate cooling solution would make any GPU look hot.

 

I doubt Acer was the one deciding the GPU clocks. It was the one responsible for cooling it though.

Still, this is ridiculously hot.

 

So it's up to the OEM to decide final clocks?

Well, an ultrabook chasis, even if it is 15", might a bit too little to put such a "beast" inside. But it sure looks like a common issue, if cooling isn't adequate.

 

The Nvidia rep said they choose the thermal limit, and the GPU throttles clocks to meet it. Acer decided to be aggressive with the thermal limit on the GPU, but not with the cooling. So, maybe this laptop isn't the one for you, but it's hardly time for overwrought doom and gloom about Kepler's perf/watt.

 

I don't know if anyone else caught this, but did Tim Sweeny just leak kepler floating point spec?

I can't think of anything else that would line up with that unless hes talking about SLI...

 

GK104 Transistor Count and Exact Die-Size Revealed

 

^^ it's posted on previous pages..

another 640M preview by anandtech confirmed 384CC@625MHz,(no hotclock madness), 128bit 1.8GHzDDR3

http://www.anandtech.com/show/5672/acer-aspire-timelineu-m3-life-on-the-kepler-verge/1

 

Hmm.. a bit cryptic.. what numbers are they aluding to? The 384 cores?

 

That's pretty good performance from what I assume is a 30GB/s part. Would be interesting to see how it fares with GDDR5.

 

http://www.brightsideofnews.com/new...ia-launches-kepler-meet-geforce-gt-640m-.aspx

Theo says the drivers provided were invalid (???) and showing wrong clocks. There is indeed a uniform 625mhz. Wonder what is up with those performance boost tricks.