View Single Post
P: 17
 Quote by Coin Hi Cold Winter, you may want to read the ATLAS group's somewhat tortured narrative of exactly how they went about their calculation. The limiting factor turns out to actually be not CPU power, but RAM. Performing the calculation turns out to involve constructing some tables that grow to hundreds of gigs in size, and they found that they essentially had to store the entire table in RAM, as the calculation requires so many more-or-less-random accesses to this table that access times would have been prohibitive had they allowed swapping to disk.
I've looked at the Atlas code. Yes it seems to be all "C", but with a few modules that seem more GUI related. I am perhaps a bit more jaded when it comes to iron than others. For e.g. I have a supplier pawning DDR2 ram for $9.99/GB... that puts 128GB in the$1278 range. Same source has a 16 way SATA2 controller at around $600 and of course, 250GB SATA2 drives are now in the$120 range ( 16X means $1920 ) Motherboards are now in the$1500 range for 4 socket Opterons... that leaves 4 cpus ( AMD has announced 4 core units for 2009 ? ) typically in the $1200/piece range. I figure I can build one H...!! of a monster for around$10,698 in todays terms.

In 18 months that could be under \$4000... which I think I can swing at that time.

BTW, thats 2Terabytes of disk mirrored and striped, so getting around the other difficulties ( capacity and I/O speed ) noted in that article isn't a biggy. That SATA2 controller at 300MB/sec will be hitting all disks at about 192Mb/sec... which should translate into a run time guesstimate of 11,000 seconds ( <200 minutes? ).

I'm inclined to go with FP math on this so that conversion will double the bandwidth requirement at a bit higher speed. Although I have to check the SSE capability with the 32 bit integers that the Atlas programmers originally used. That could certainly impact the run times in both directions.

In effect, given some time ( and a budget to fit my limited means ) I should be able to hammer E8 quite nicely. Certainly for much less than what the LieGroup are talking about.