GIMPS on PS3

[CRGreathouse]

I still don't understand your TF vs LL comment.

Originally Posted by flouran

Although it may not be REALLY fast, it would be interesting to do it nonetheless.

I think it would be very interesting. I'm expect it to run at about 4% the speed of a similarly clocked-and-cored Core 2: lose a factor of 12-14 on the double emulation and another factor of 2 because there's no machine code or other optimization specifically for the Cell. Better than 8% would surprise me; better than 20% would floor me. On the other end, performance less than 1% would probably indicate that something is set up wrong.

Edit: I mean 4% the performance of Prime95 on a Core 2. It could perhaps achieve 6-8% the performance of Glucas on a Core 2 (per core-clock).

[retina]

Originally Posted by flouran

Plus it is a gaming console so it should be fast (as compared to say the current market of the x86 family of processors).

Apples to oranges again. Faster at what? You still haven't answered that.

[flouran]

Originally Posted by CRGreathouse

I still don't understand your TF vs LL comment.



I think it would be very interesting. I'm expect it to run at about 4% the speed of a similarly clocked-and-cored Core 2: lose a factor of 12-14 on the double emulation and another factor of 2 because there's no machine code or other optimization specifically for the Cell. Better than 8% would surprise me; better than 20% would floor me. On the other end, performance less than 1% would probably indicate that something is set up wrong.

Edit: I mean 4% the performance of Prime95 on a Core 2. It could perhaps achieve 6-8% the performance of Glucas on a Core 2 (per core-clock).

Maybe I can skip double emulation and make it single emulation if I can compile Mlucas NATIVELY on Yellow Dog. We'll have to see.

[flouran]

Originally Posted by retina

Apples to oranges again. Faster at what? You still haven't answered that.

Would anyone like to prove either me or retina correct?

[retina]

Originally Posted by flouran

Would anyone like to prove either me or retina correct?

Hehe, you fail to realise that I have not made any claim to be proved/disproved. Only you have made the vague claim about something being fast. I am asking you fast at what. And your comparisons are all apples to oranges anyway so I'm not sure why I bother.

[mdettweiler]

Originally Posted by flouran

Would anyone like to prove either me or retina correct?

Well, sorry to bust your chops, but I think I can prove retina correct in that you are indeed comparing apples to oranges. TF is indeed much slower than LL *IF* you use TF *in place of* LL--which nobody in their right mind would do for GIMPS-size numbers since it would take a million years to prove the primality or nonprimality of a number that doesn't have any relatively small factors. However, if you use TF as a *supplement* to LL, which is what GIMPS is doing, then you can use TF to greatly improve overall throughput by eliminating candidates with relatively small factors (which can be found with relatively little TF work) so that you don't have the spend all that time doing an LL test for every single number.

Long story short, as retina said, a case of apples and oranges--TF and LL are both important parts of the overall GIMPS testing process, though they are for different stages of the process on each exponent. Thus, if a PS3 could do TF, it would help out with *that stage* of the process, but not with the LL stage of the process. However, since there is plenty of both LL and TF work that needs to be done yet for GIMPS, a PS3 doing TF would still be quite beneficial to the overall throughput of GIMPS.

[flouran]

Originally Posted by mdettweiler

Well, sorry to bust your chops, but I think I can prove retina correct in that you are indeed comparing apples to oranges. TF is indeed much slower than LL *IF* you use TF *in place of* LL--which nobody in their right mind would do for GIMPS-size numbers since it would take a million years to prove the primality or nonprimality of a number that doesn't have any relatively small factors. However, if you use TF as a *supplement* to LL, which is what GIMPS is doing, then you can use TF to greatly improve overall throughput by eliminating candidates with relatively small factors (which can be found with relatively little TF work) so that you don't have the spend all that time doing an LL test for every single number.

Long story short, as retina said, a case of apples and oranges--TF and LL are both important parts of the overall GIMPS testing process, though they are for different stages of the process on each exponent. Thus, if a PS3 could do TF, it would help out with *that stage* of the process, but not with the LL stage of the process. However, since there is plenty of both LL and TF work that needs to be done yet for GIMPS, a PS3 doing TF would still be quite beneficial to the overall throughput of GIMPS.

TF it is on the PS3. Thanks for clearing it up.

[CRGreathouse]

Originally Posted by flouran

Maybe I can skip double emulation and make it single emulation if I can compile Mlucas NATIVELY on Yellow Dog. We'll have to see.

Single precision will require something like 3 times as many pieces and ~20 times as much time overall, I hear. I haven't done even rough calculations on that yet; perhaps I will tomorrow for kicks. But surely you're not serious in considering *rewriting* code to use single precision, are you? That might take me over a year to do myself (familiarizing myself with the code and the math + the actual coding).

[CRGreathouse]

Originally Posted by flouran

TF it is on the PS3. Thanks for clearing it up.

I don't know how fast TF would be on the processor; I don't know of anyone who has tried it. But it does seem more likely to be fast than LL on a Cell.

[flouran]

Originally Posted by retina

Hehe, you fail to realise that I have not made any claim to be proved/disproved. Only you have made the vague claim about something being fast. I am asking you fast at what. And your comparisons are all apples to oranges anyway so I'm not sure why I bother.

Do some research then if you are unsure... Besides, your comment, "And your comparisons are all apples to oranges anyway so I'm not sure why I bother." was unnecessary and if you have more things to say like that please don't post (on this thread at least) unless you wanna contribute to running MPrime on a PS3. By the way, I wasn't really trying to prove anything; I simply stated my own opinion, it was YOU who wanted my statement to be either proved or disproved.

[flouran]

Originally Posted by CRGreathouse

Single precision will require something like 3 times as many pieces and ~20 times as much time overall, I hear. I haven't done even rough calculations on that yet; perhaps I will tomorrow for kicks. But surely you're not serious in considering *rewriting* code to use single precision, are you? That might take me over a year to do myself (familiarizing myself with the code and the math + the actual coding).

Of course not, I won't write the code. Besides, if it does work, that's great; and if it doesn't, well then it doesn't.