On Wed, Nov 30, 2005 at 12:53:47PM +1000, Rob Studdert wrote: > On 29 Nov 2005 at 20:22, John Francis wrote: > > > Some research groups at Universities did put together Mac-based > > rendering systems, especially once the cost per compute cycle on a > > Mac became more or less competitive with the Intel processors. > > I expect too that a lot of the potential the Macs had in CPU cycles was > severely degraded by their relatively poor net-working performance (until > late). That's pretty much why Syquest media was so popular in Mac > environments > for such a long time, it was quicker to walk data across a room on a Syquest > disk than wait for the network.
Actually, sneakernet (as it is colloquially referred to) is still a good way to move 'large' amounts of data (for appropriate values of large). The bandwith of a hand-carried case full of DVDs is still more than most people are prepared to pay for electronic transmission. Anyway, ost of the render farms don't require much in the way of I/O. Each of the systems gets a copy of the model database initally, then gets told which frame (or frames) to render. There's no communication between systems except for the propagation of the database at the start, and collection of the completed frames at the end. Those applications which do require massive inter-system I/O (such as a complex radiosity solution to pre-calculate the ambient lighting in a large scene) will still be run on a tightly-coupled multi-processor system (which is where Sun, SGI, etc. still have real advantages; the design of a good crossbar backplane far outweighs raw CPU power).
