mersenneforum.org Hardware accelerated LL Testing
 Register FAQ Search Today's Posts Mark Forums Read

 2013-04-24, 21:44 #1 TObject     Feb 2012 1100101012 Posts Hardware accelerated LL Testing I believe this subject has not been discussed for at least a couple of years. Since the technology stepped up, I was wondering what everybody’s fresh thoughts on this are. First, we have high-performance computing accelerators, such as TI’s Quad-DSPC 8681 PCIe card. How fast do you think could that be programmed for the purposes of finding Mersenne primes? Second, let us take a look at the progression of Bitcoin miners. First Bitcoins were mined on CPUs; then 100-times-faster GPU mining came along. After which field-programmable gate array (FPGA) hardware allowed to bring the power consumption down. And now application-specific integrated circuits (ASIC) are coming on the market, blowing everything else out of the water performance wise. Could the similar progression continue for Lucas-Lehmer testing? We have LL testing on CPUs, and LL testing on GPUs came along with CUDALucas. Are FPGAs and ASICs next?
 2013-04-24, 23:12 #2 Aramis Wyler     "Bill Staffen" Jan 2013 Pittsburgh, PA, USA 19B16 Posts ASIC are obviously going to be the fastest as their switches are minimal and they have memory in place. But there's no financial gain in making ASICs for LL tests. People should just put that out of their heads right now. (secretly begins designing the 100m digit asic)
2013-04-26, 06:30   #3
jasong

"Jason Goatcher"
Mar 2005

3·7·167 Posts

Quote:
 Originally Posted by Aramis Wyler ASIC are obviously going to be the fastest as their switches are minimal and they have memory in place. But there's no financial gain in making ASICs for LL tests. People should just put that out of their heads right now
Well, how do you calculate the value of a product when the reason for it's use is mainly "for the pure hell of it."?

There are so many things which are totally useless, but sell briskly because of the way the human mind works. I, personally, want to buy myself a binary watch. I am the only one in my family who would gain joy from that, because it's a totally useless product. I want it because it's "cool."

Same reasoning applies to anything that could help with LLR.

 2013-04-26, 06:49 #4 ixfd64 Bemusing Prompter     "Danny" Dec 2002 California 2·3·397 Posts One can only hope there will be a super-generous billionaire who happens to be interested in number theory.
2013-04-26, 07:45   #5
jasong

"Jason Goatcher"
Mar 2005

3·7·167 Posts

Quote:
 Originally Posted by ixfd64 One can only hope there will be a super-generous billionaire who happens to be interested in number theory.
Well, take Moore's Law and look at it another way.

My definition of Moore's Law, setting aside the historical definition, is twice the awesomeness for the same price. But that could also mean the same awesomeness for half the price.

By that definition, lowering production costs of chips could be thought of as a way to extend Moore's Law. I think there will come a point where people will be able to obtain custom-made chips. So we'll be able to take out all the unneeded crap and build hardware specifically for LLR. It'll probably still be cheaper to make them in bulk, but the definition of bulk will probably have a lot lower starting number.

2013-04-26, 14:59   #6
Aramis Wyler

"Bill Staffen"
Jan 2013
Pittsburgh, PA, USA

3·137 Posts

Quote:
 Originally Posted by jasong Well, how do you calculate the value of a product when the reason for it's use is mainly "for the pure hell of it."?
Heh, I might have been too subtle there. There is a cash prize for the 1st 100M digit number, and printing specalized integrated circuits on say, a pci (not pci-e) platform would probably be able to do them quickly enough that whoever got the hardware built would have a great chance of winning the prize. That's why I told everyone else to get that out of their heads, and then started building them myself in the post. Reduce the competition.

Last fiddled with by Aramis Wyler on 2013-04-26 at 15:00

 2013-04-26, 18:12 #7 jasonp Tribal Bullet     Oct 2004 2×29×61 Posts That DSP board looks quite nice, but it's not going to be faster than a PC; that isn't the point of such things. The more interesting question is if, say $20000 of design effort with strange parts can produce an LL test faster then$20000 worth of PCs and graphics cards. I'm just about sure the answer is no. After all, Moore's Law floats all boats.
 2013-04-26, 18:43 #8 Aramis Wyler     "Bill Staffen" Jan 2013 Pittsburgh, PA, USA 41110 Posts Eh, I had to learn a thing or two about circutry design when I was going the electrical engineering (microcircuitry spec) side of my applied computer science degree. The real difficulty in designing cpus is their general purpose nature - designing a dedicated application specific circuit is not terribly difficult, it is terribly fast (no switches, and gates almost all hardwired), and the microfabs these days can produce batches of them for not a whole lot of money. Unforunately the more dedicated they are the less versatile they are. In my example I (jokingly) said I'd secretly design an asic for 100M digit LL tests. That's pretty specific, and it would probably not funtcion very well for 50 or 150M digit tests. Its purpose would be fleeting, and that's why people don't make them, as opposed to oppresive difficulties or cost. Making CPUs is hard, because they're made to do anything. ASICs are simply converting well written library functions into silicon. That's not very hard, and the ability to ignore the clock means it can run at the speed of your electrical current even into 10s of gigaherts. I think I have 4 pci slots on my motherboard, and they're about useless. I could put a modem in one maybe, or a sound card. They're too small for video cards but would be great mount points for an ASIC. In bitcoin mining, someone found a purpose that would last long enough doing the same tired series of commands than an ASIC made sense. It's not like that is truly new technology, it's simply an interactive driver. AI had a movement a decade or so ago regarding neural networks pushed into hardware in a similar manner, but the concept wasn't ready - neural networks simply aren't well known enough to press into hardware reliably. The calculation for bitcoin mining is set.
2013-04-26, 19:55   #9
ewmayer
2ω=0

Sep 2002
República de California

2×5×1,163 Posts

Quote:
 Originally Posted by Aramis Wyler Eh, I had to learn a thing or two about circutry design when I was going the electrical engineering (microcircuitry spec) side of my applied computer science degree. The real difficulty in designing cpus is their general purpose nature - designing a dedicated application specific circuit is not terribly difficult
Ever actually design one?

 2013-04-26, 20:37 #10 jasonp Tribal Bullet     Oct 2004 DD216 Posts Um, Bill, you're lecturing a forum with several people in it who work in the hardware business. Just because you can get rid of all the CPU bells and whistles does not, by itself, mean you can solve your problem any faster. Think of a PC as the result of a multi-year, multibillion dollar ASIC design effort; that's your competition.
 2013-04-26, 20:40 #11 Aramis Wyler     "Bill Staffen" Jan 2013 Pittsburgh, PA, USA 6338 Posts I had to design a 286 series cpu with an 8 bit bus and write a word processor in assembly that would run on said design as part of the last course in that tree, but no I've not tried to design something like that since. In my particular case, I wouldn't know the math for how to code a lucas lehmer test in software, and I'd have to learn that before I pressed it to hardware. But I did have fun tutoring the EET students who had a lot more experience actually working with the circuits physically but not quite so much theory on designing them.

 Similar Threads Thread Thread Starter Forum Replies Last Post cseizert Programming 8 2016-10-27 05:55 kladner Soap Box 3 2016-10-14 18:43 rx7350 Hardware 1 2007-08-02 14:47 eepiccolo Math 6 2006-03-28 20:53 Jasmin Hardware 10 2005-02-14 01:58

All times are UTC. The time now is 20:32.

Fri May 14 20:32:49 UTC 2021 up 36 days, 15:13, 0 users, load averages: 1.70, 1.77, 1.91