If fermi flops what happens to Nvidia

[Hacp]

What does "double precision" do? Is that the only type of calculation performance that matters for HPC?

Where is ATI when it comes to double precision?

Isn't double precision just 16 numbers instead of 8?

 

[BenSkywalker]

It's not going to do well in the 3d/rendering market because it uses way too much power and produces too much heat for effective an effective server model. If you are building some sort of render farm you would not want to use Fermi cards because you will be looking at performance per watt if you have any sense.

Fermi is now using over 2000 watts? I must have missed that article.

Can someone explain the trade-offs a company must make in total computational power when building more double precision into a product?

It isn't a question of computational power, but one of die space. How much more die space it is going to use is going to vary wildly based on the architecture.

Would it be possible for the entire architecture to be double precision?

Absolutely, but it wouldn't be a very good design. It comes down to the particular exacting piece of math if DP is going to do you any good at all. If your code is 30% DP, 45% SP and 25% scalar then having a chip that is entirely build for DP ops will be oversized and rather power hungry compared to one that has more functional units for SP. Tip the balances of code around and things start to look differently. As a processor engineer, you have to make a trade off. Obviously it is a lot more complex then DP or SP ratio of transistors, but that is one of the many issues you must deal with and every transistor matters.

So other than the event/error handling and ecc memory the cards are kind of 1 to 1.

I would say we will have to wait and see. If you took an abstract cursory look you could say that the PentiumD is 1:1 with the i3 too. Some may skoff at that comparison, but they are likely to be much closer in HPC apps then ATi and nV.

 

[Schmide]

Double precision is for HPC. Single precision in games since it's not critical the calculations are exactly right, game engine code is optimized for approximations anyway since it's faster than finding the real values.

It's 64 bit floating point numbers instead of 32 bit floating point numbers.

There are some trade offs where single precision can be very useful, especially when using inherently error prone methods. (i.e. Euler methods) If you can increase the sample rate by 2-4 times in the same computing time by using single precision, why not? More often than not the error term exceeds the extra precision anyways.

For example say someone is sampling winds/pressure across a large area. Does a thousandth/micro of a kph/bar really matter?

Isn't double precision just 16 numbers instead of 8?

and a 7 fold exponent difference for larger numbers.

 

[HeXen]

Release a game that requires Fermi only for added special graphical technique, Arkum asylum anyone?

 

[tweakboy]

They change their name from nVidia to 3dfx lol lol

 

[vss1980]

One or all of the following:

1) Stock value will go down a little
2) CEO will try to spin it in his normal way

At this point, I see little else happening.
IF Fermi proved to be a completely crap bit of work with bugs, bad architecture choices, etc., seeing as its their current new product basis, then they might really be in a more serious loss making position.

 

[Friendo]

they cry DUH

 

[PlasmaBomb]

Hard to find/not at Newegg is different than impossible to find. Going to Google shopping I found a few places that have them in stock. If you think the 5970 is hard to find, I'm willing to bet Fermi will make it look plentiful in comparrison. Just my guess... we'll see soon enough.

Newegg currently have them in stock :hmm::hmm:

 

[1h4x4s3x]

Already gone.

 

[PlasmaBomb]

Already gone.

ETA:03/23/2010

Dang