Bitcoin miner for GIMPS...

[science_man_88]

to be honest cost is the main issue based on what's in the thread already however I do think it is possible to some extent because of Laurv's posting about shifts and adds technically with a bittest those could be turned into a way to do TF and LL at very least just not efficiently. squaring can be done using bit shifts and subtraction is just like adding a negative number to something. it's really the bit testing needed that would make it fail I think that and it's inefficiency. because although it works without bit testing to square mersenne numbers not so much outside of them:

(08:58) gp > 5<<2+5<<0
%1 = 25
(08:59) gp > 7<<2+7<<1+7<<0
%2 = 49

they both can be squared using bit shifts the problem is testing to which amounts they need to be shifted by otherwise it becomes 5*7 if you shift 5 by all 3 of [0,1,2]. And technically subtraction without a counter is like saying these numbers are congruent mod a number. but these methods take too long to perform efficiently.

[axn]

Originally Posted by airsquirrels

There may be more broad market value to an ASIC that could do VERY fast FFTs, which we could use....

Yes. They're called DSPs.

[Madpoo]

Originally Posted by xilman

I'm trying to teach myself Verilog and FPGA design. Perhaps a TF implmentation may be forthcoming. Don't bet on it happening any time soon, unless anyone else wants to work with me.

Even so, it's unlikely (IMAO) that a sufficiently fast implementation will happen any time soon.

I did FPGA design back in college... pretty basic stuff and the end goal of the class was the design of a simple 4-bit limited-op CPU. Xilinx chips and I forget the name of the software we used.

It was pretty interesting to work out the design and then hook up a programmer and, for testing, using a push-button clock (so we could step through cycle-by-cycle and check inputs/outputs).

It did open my eyes to the possibilities of such things for custom applications, and I'm pretty sure the custom ASICs for bitcoin mining are along those lines. It made me wonder if the ASICs used in them aren't really FPGAs that could be reprogrammed with a different function. Or do they really such a huge market that they custom fabricate crap like that?

[Mark Rose]

Originally Posted by Madpoo

I did FPGA design back in college... pretty basic stuff and the end goal of the class was the design of a simple 4-bit limited-op CPU. Xilinx chips and I forget the name of the software we used.

It was pretty interesting to work out the design and then hook up a programmer and, for testing, using a push-button clock (so we could step through cycle-by-cycle and check inputs/outputs).

It did open my eyes to the possibilities of such things for custom applications, and I'm pretty sure the custom ASICs for bitcoin mining are along those lines. It made me wonder if the ASICs used in them aren't really FPGAs that could be reprogrammed with a different function. Or do they really such a huge market that they custom fabricate crap like that?

There used to be a market for FPGA-based mining hardware, but the ASICs left them in the dust.

[Madpoo]

Originally Posted by axn

Yes. They're called DSPs.

Except most DSP chips for commercial applications have what I think you might call "interesting" precision. I read an article the other day about a new DSP with some kind of FFT algorithm that could do compression/decompression really fast and efficient, but it depends entirely on the fact that audio and video has a lot of empty frequencies. My takeaway was that it would suck at doing anything with more random data like what happens during LL.

(my apologies for probably mangling the underlying concept...I skimmed the article)

[ewmayer]

Originally Posted by science_man_88

subtraction is just like adding a negative number to something.

Holy crap! I'm gonna need to rewrite half my code to take advantage of that.

Re. the rest of your clever scheme: so to multiply a pair of GIMPS-wavefront-sized inputs, we simply need to do 70-million-or-so 70-million-bit-wide shift-and-adds ... got any similarly genius ideas for the ensuing modular reduction step? And that doesn't even take advantage of the fact that for LL testing we are *squaring* the input, rather than general-multiplying ... I bet we could get a further speedup by taking advantage of that. See, this is the reason I spend time around here...

[science_man_88]

Originally Posted by ewmayer

Holy crap! I'm gonna need to rewrite half my code to take advantage of that.

Re. the rest of your clever scheme: so to multiply a pair of GIMPS-wavefront-sized inputs, we simply need to do 70-million-or-so 70-million-bit-wide shift-and-adds ... got any similarly genius ideas for the ensuing modular reduction step? And that doesn't even take advantage of the fact that for LL testing we are *squaring* the input, rather than general-multiplying ... I bet we could get a further speedup by taking advantage of that. See, this is the reason I spend time around here...

I also said:
"just not efficiently." modular reduction could be as simple as the subtraction step done in the thread I made in the lounge http://mersenneforum.org/showthread.php?t=20803 as all a modular result tells you is what he remainder on division is and you can get there using subtraction. oh and the shifts can potentially be decreased as they are being made using subtraction of the exponent because the powers of two that are a multiple of the next power of two will all be 1 for remainder. anyways enough with this self defeating post.

[ewmayer]

Originally Posted by science_man_88

...anyways enough with this self defeating post.

Ever hear of the saying to the effect of "when you find yourself in a hole..."? [Hint: put down the shovel.]

[science_man_88]

Originally Posted by ewmayer

Ever hear of the saying to the effect of "when you find yourself in a hole..."? [Hint: put down the shovel.]

I was pointing out that technically it's not the lack of instructions that would be the downfall it would be the way it's wired and the order it's forced into.

[ewmayer]

Still digging, I see.

[Hint-which-will-surely-be-ignored: Fundamentally quadratic complexity versus subquadratic.]

[science_man_88]

Originally Posted by ewmayer

Still digging, I see.

[Hint-which-will-surely-be-ignored: Fundamentally quadratic complexity versus subquadratic.]

you can make a test version that uses 2*x^2-1 which is nearly the 2*x^2 used in the TF when the bit is 1. and as I've said through PM I've made code that can do both before I'm just not sure where I posted it. Also like I've said through PM I've also tried to use what the basic idea shows us to try and make a test within the mersennes themselves. I have also tried to do this on the forum and I can't remember where it is. found the code :http://mersenneforum.org/showpost.ph...&postcount=396