undefined | Better HN

0 pointsexistencebox12y ago0 comments

Probably? (jadedly) It's a cheap one off that earns them a lot of free publicity.

Folding and SETI have been at it for decades, and will be for decades more. That'd be a lot of power/hardware costs for little net gain modulo time spent.

(Note: 30 cpu years really isn't a lot of time when you start getting truly massive parallelism, and google is certainly at that scale. I'm honestly surprised to see this is pthreads based, I would have expected it to lean heavily on CUDA (or maybe I'm just showing my naivete to the field of parallel programming))

0 comments

birken12y ago

Just to put some numbers behind this:

35 CPU years = 35 * 365 = 12,775 CPU days

Let's assume that your standard Google machine has 2 8-core processors, that is 16 cores per machine

12,775 / 16 = 798 machines

800 machines is a lot for your average person, but for a company like Google it is a very small drop in the bucket. And that is assuming you need to do all of the calculations in one day. If you are willing to do the calculations over a month, you only need 26 machines.

Even if "CPU year" applies to one 8-core processor, that still means it is just 26 * 8 = 416 machines, which is a very small number for Google.

And remember, the machines are still doing other things, this is just using idle CPU (generally a plentiful resource for your standard internet giant workload).

largote12y ago

It would seem to me that this kind of program would be heavily reliant on branching for which Vector processing (which is what GPUs excel at) is not the best choice.

nathancahill12y ago

CUDA is for GPUs (specifically NVIDIA GPUs). This is CPU processing. Pthreads is the way to go.

j / k navigate · click thread line to collapse