Money > Sense: Building a $15k "image crunching" computer
The latest "Ask Dan" column at Dan's Data finds our eponymous expert fielding a query from a man looking to spend $15,000 (Canadian) on an machine capable of "crunching as much image data as computationally possible per hour." Dan gives him the straight answer, but I have to ask: Surely a couple of different machines with some high-speed interconnect could do the job more efficiently than one machine kitted out with the most expensive server-class hardware available?
It's hard to make a judgement without knowing the details of the questioner's software architecture, but Dan points out the details:
For most applications, a high-end quad-core LGA775 system (including more sensibly priced Core 2 Quads that've been overclocked to three-point-something gigahertz, which is easy to do) will be indistinguishable from a dual-quad-core Skulltrail system. If you're doing something that's highly multithreaded, though, the system with twice as many cores may perform something approaching twice as well. There are all sorts of serious-server and scientific computation tasks that certainly can take advantage of those extra cores. Some other tasks, like hosting a bunch of game servers (on a LAN or a very fat Internet pipe), or doing some kinds of video effects generation or compression, or just charging up the Folding@Home or distributed.net charts, will also benefit greatly from more than four cores.
By and large, though, four cores really are more than enough for almost everybody these days. And even if you do need eight cores, it's still hard to justify making that eight cores of overpriced Q9775, when almost-as-fast Xeons are so much cheaper.
Ask Dan: The deadly temptation of Skulltrail [DansData.com]
the latest
latest episodes
The movie and television production industry is always looking for faster/cheaper rendering farms to process 3D modeling, animation and effects.
This type of number crunching is perfect for distributed computing because while each frame may be visually dependent on the previous one, the computations are not. Complex frames are often broken up into 2^x chunks along the natural borders defined by various MPEG standards. Again visually they connect, but for rendering each piece does not need any other.
These individual little puzzles can be sent a 1,024 or more different computers all over the planet wherever processing time is available.
Once all of the pieces are ready a central computer can glue the frames together and create an MPEG file. The MPEG part does depend on previous and future frames and is best done on muti-core or closely spaced computers.
BTW: Usually all of this is done with no or lossless compression.
Then at the end, someone always yells:
Here's your movie Boss!
We never get tired of hearing it. OK I do but I keep my mouth shut.