PS3

[davieddy]

Originally Posted by crash893

we didnt start on p4s we had to work are way up to them but its a good thing that we started before hand for lessiosn learned.

But "we" (e.g. George) did start out with double precision FP.
Short of emulating double precision using single precision, the task
is impossible.

[jasonp]

Originally Posted by jasong

To be brutally honest, the probable reason the proof hasn't shown up at all is probably because he knows that my crappy reputation here, though he probably doesn't agree with it, is his best protection possible.
...
I am egging you people on because the greater your fall, the more likely it is that will change your ways, in my opinion.

As davideddy has said, the proof hasn't shown up because it is impossible. You cannot wish away the need for double precision floating point. Trying to emulate quad-precision using double precision slows things down by a factor of 10 (I've written the code for this) and the Cell processor sees a similar slowdown trying to emulate double precision using single precision.

Nobody's tried a proof of concept with a graphics card because you don't have to try hitting your head with a hammer to know that it's a bad idea.

PS: do you want to prove this to yourself? That's an admirable notion. Msieve contains a reasonably compact FFT multiplication library written in C; if you teach yourself the necessary programming skills then you can move the code over yourself, and see how long a multiply at a given digit level takes. I can shortcut the process and tell you it will take forever, but your tone makes me think you won't believe me. What do I know, I've just been doing this for over a decade.

[Mini-Geek]

Originally Posted by jasonp

Nobody's tried a proof of concept with a graphics card because you don't have to try hitting your head with a hammer to know that it's a bad idea.

Can someone please explain to Jasong, me, and anyone else that doesn't really know, why it's as obvious that you need double precision as hitting your head with a hammer is bad or as a snow plowing sedan being bad?
I'm not blindly believing cranks at the drop of a hat and taking all they say as gospel truth, but really - what is double precision exactly and why is it so important?
p.s./sidenote/on a related note: if we had quad-precision CPUs (and coded to use them fully), would the searching be sped up by as much as the single-double jump?

[davieddy]

Originally Posted by Mini-Geek

Can someone please explain to Jasong, me, and anyone else that doesn't really know, why it's as obvious that you need double precision as hitting your head with a hammer is bad or as a snow plowing sedan being bad?
I'm not blindly believing cranks at the drop of a hat and taking all they say as gospel truth, but really - what is double precision exactly and why is it so important?
p.s./sidenote/on a related note: if we had quad-precision CPUs (and coded to use them fully), would the searching be sped up by as much as the single-double jump?

With double precision, each element of the FFT has <20 bits of
the number +33 guard bits.
In single precision I think you need more "guard bits" than you have
bits left over.
As for "quad precision", the answer is more than "double" speed
improvement but not as impressive as going from impossible to feasible.

[crash893]

what about this?

http://ati.amd.com/products/streamprocessor/specs.html

[cheesehead]

Originally Posted by Mini-Geek

Can someone please explain to Jasong, me, and anyone else that doesn't really know, why it's as obvious that you need double precision as hitting your head with a hammer is bad or as a snow plowing sedan being bad?

It's not that the need for double precision is obvious. It's that the need for double-precision (i.e., why single-precision is inadequate) has been discussed at length in the past, so "need for double-precision" is a handy way to refer to the fifty-thousand-or-so words in those discussions.

I'm not blindly believing cranks at the drop of a hat and taking all they say as gospel truth, but really - what is double precision exactly

Answer:

Originally Posted by Wikipedia ([URL

http://en.wikipedia.org/wiki/IEEE_754[/url]]"The IEEE Standard for Binary Floating-Point Arithmetic (IEEE 754) is the most widely-used standard for floating-point computation ... The standard defines formats for representing floating-point numbers ... together with a set of floating-point operations that operate on these values. ...

IEEE 754 specifies four formats for representing floating-point values: single-precision (32-bit), double-precision (64-bit), ...

...

Following is a description of the standard's format for floating-point numbers. ..."

... or, read section "4.2 Floating-Point Arithmetic" in Knuth's The Art of Computer Programming, Vol. 2.

Originally Posted by Mini-Geek

and why is it so important?

Read this thread, listed at the top of the Hardware Forum thread listing: "Sticky: The prime-crunching on dedicated hardware FAQ ". Follow the links to past threads, and read those. It won't be easy reading. (Eventually, maybe, someone will summarize it all in a single narrative on the wiki.)

p.s./sidenote/on a related note: if we had quad-precision CPUs (and coded to use them fully), would the searching be sped up by as much as the single-double jump?

No. (For a more complete answer, read that sticky and the threads it references.)

[cheesehead]

Originally Posted by crash893

what about this?

http://ati.amd.com/products/streamprocessor/specs.html

Note that it doesn't mention any detail about the double-precision, such as how much slower that is than single-precision, or whether it's native-mode or emulated.

[mdettweiler]

Maybe we should start lobbying IBM/Sony (I think IBM is who developed the Cell processor) to make the next version of the Cell processor (and thus, hopefully the PS4) with dual-precision floating point capability?

[fivemack]

The next version of the Cell does have dual-precision floating point capability; it uses DDR2 memory of up to 16GB per processor, 3.2GHz clock, two processors in a blade, four DP operations per SPE clock cycle. The blades fit into a 'BladeCenter H', a snip at $8000, and cost $10000 each. $10k for 217GFlops peak, so $46/Gflop-peak, versus $16/Gflop-peak from Q6600 white-boxes.

http://www-03.ibm.com/systems/info/b...s22/index.html

[cheesehead]

Originally Posted by fivemack

The next version of the Cell does have dual-precision floating point capability

Yes, it seems to have honest-to-gosh native-mode double-precision ...

so $46/Gflop-peak, versus $16/Gflop-peak from Q6600 white-boxes.

... at a price.

http://www-03.ibm.com/systems/info/b...s22/index.html

Hmmm ... There's a very interesting sentence in their PDF (http://www-03.ibm.com/technology/cel...rief_5-02b.pdf) about simulations for the petroleum industry (italics are my emphasis):

Simudyne's simulation platform is specifically designed to assist in hydrocarbon exploration by growing algorithms that help speed up seismic surveys, optimize complex logistics networks and create software that improves according to the same principles that drive adaptation in natural systems.

They've programmed evolution!

(Take that, creationists!

Though .. I suppose they'll point to the word "create" just before what I italicized, and say IBM has programmed Creation ...

... to which the comeback would be, "You mean ... that ... God ... is ... an algorithm grown on Simudyne's simulation platform?")

[cheesehead]

Mini-Geek,

I think I owe you, Jasong, and the others a better answer to your post #25 ("Can someone please explain ... what is double precision exactly and why is it so important? ... if we had quad-precision CPUs (and coded to use them fully), would the searching be sped up by as much as the single-double jump?") than I gave in post #28.

I'm working on a reasonably short way of condensing the single-versus-double discusion, and will post it when done. Feel free to nag me if you don't see it here soon.