Let's talk about Double-Precision calculation speed.

Brace yourselves... I am going to post what appears to be a disturbing fact to me, about Fermi.

http://www.techreport.com/articles.x/18332/5



I do realize that there are not that many apps that actually use double-precision calculations, but I am disappointed at the news. I like to play with various GPU accelerated apps, and you never know when having outrageous double-precision speed might come in handy.

I wish they would reconsider their decision on this, especially if our gamers Fermi, is actually capable of running them at a much greater speed. (Only deactivated, sitting there idle.)

I do hope this is bad info.

I have heard F@H is single precision only...  ??

I'm just sick about it.

Folding is single precision.

Talonman I still am excited about Fermi too...

Really? I hadn't noticed...
 
;-)

Come on guys, with all reason and logic, they had to do something to differentiate Tesla and Quadro from GeForce line. Why would anyone buy those expensive pieces otherwise?
That and also think about the possibility (fact?) that sustaining a high level of DP performance would both cost the card some performance overhead for gaming AND require "better" chips and /or higher Q&A cost and /or other overhead to Nvidia. We've been talking about how DP is not needed for gaming (this time around at least). There you go, DP is attenuated...

Good observation, but DP performance should not be underestimated. There is always a possibility that ATI or Nvidia might use this to gain more performance out of their GPUs, but I am up for it, if you ask me.....hehe

Ya a bit of a casualty I think in Nvidia's attempt to grow their their Tesla business.  I had a feeling they might disable or purposefully cripple some features to differentiate their GeForce and Tesla parts as they have been doing for years with their Quadro parts.  Some features detailed in the Fermi previews, like ECC, I knew wouldn't be on the GeForce as there's no need for it and ECC memory carries significantly higher costs and ultimately runs slower than non-ECC.  I also thought they might disable some of the L1/L2 on the GeForce parts for both binning/performance reasons, but thankfully they are not.

Hehe ya I know its a bit of bummer for the folks interested in CUDA apps, but in the end Nvidia is delivering on all important features which relate to gaming for GeForce parts.  They're only cutting back in areas that don't benefit gaming at all, like ECC and DP performance.  Also from the TR write-up, it sounds like you need the ECC anyways to really benefit from the DP speed-up, as one of the major reasons for ECC to begin with was memory coherence when reading/writing to memory.

Thanks again guys...  I do indeed feel better now!  

artificial crippling?  I'll wait for the release to make a decision, but artificial crippling of this scale, on a feature I was excited about, is enough to make me consider buying an ATI card.  I'm a long-time Nvidia fan, but damn, this sucks.  For those who care, I understand the importance of ECC.  If I was going to do real scientific research with the card, there's no question but to buy something from the right line/branding with ECC.  However, the closest I'd get to that with my own machines would be either distributed computing, which generally has fault-tolerance built into the entire system already, or my own messing around, in which case, I'd either not care about the correctness of the results, or I'd be ok with running things 4 times to confirm (an example equivalent to the crippling factor, yet with lower probability of identical error).  I like PHYSX, CUDA and friends, but if I can do the same math on an ATI card at 4x the speed, **** are they thinking with this crap?

Talonman

I have heard F@H is single precision only...  ??

I'm just sick about it.

umm its definitely single precision otherwise only gt200 would run it.

RBIEZE ...I feel like I just got sucker punched ,this is a tragedy of epic proportions that few of you seem to grasp.
As a Beta tester I need a fully DP capable card that doesnt cost Uber cash.

Many thousands of us have used gaming cards for Cuda apps.
Now that D.P is being implemented in upcoming apps and apparently D.P is hobbled on the game line,this means w'ell have to spend major $$$ for a tesla card or be relegated to the slow boat..

Wow ,I may be a switching army's

Nvidia made supercomputing affordable and placeable. Now they're offering fault tolerance and ECC accuracy to go with it.
 
They never promised to give it away. Never even hinted at it.
 
Your project's budget must be awfully slim...