Observations on the parallel between GPU and Movies.
I just realized something - the recent article on Weta Digital's upgrades for ROTK and the discussion about how the GPU and CPU might flip relationships are parallel: Weta Digital is adding 27 Terabytes of memory for all the data in ROTK, and while I was searching for data on the IBM Power 5 I noticed that the IBM supercomputer Asci Purple, which will use Power5 processors, will also have 50 TB of storage. It will also have something like 116,000 processors.
So the two ideas meshed in my mind somehow - the data requirements of digital movies seems to be outpacing that of the massive simulations of the world. Processing power is nowhere nearly as in demand, but clearly storage is.
I can only wonder, given the ever-increasing integration between computer technology and movie production, when the first attempt at distributed processing of a movie will occur. At the very, VERY end of the credits on the extended DVD of the Fellowship of the Ring, it gives thanks to LOTR Fanclub members, and the list goes on for probably 10-15 minutes, literally, so one could also conceivably get credit for processing a chunk of a movie.
But I wonder just how realistic movie technology is going to get, and if that will help to bump up the demand curve for processing power to the point where it, too, may one day outpace demand for large physics or chemical simulations.
Comments
the "Cell" chip in PS-3.....
is more a GPU than CPU. It has 4 parallel cores, each with 8 pipelines which then has 4 SIMD "pixel engines" each - so 256 of these "pixel engines".
Has ~1B transistors on 65nm process. Design cost is ~$380M split across IBM/Sony/Toshiba teams - almost up to the production costs of LOTR or Matrix trilogies. Fab cost however is over $2B. It's to tape out by YE2003, with pre-production silicon by YE2004.
It's architected/blessed by Jim Kahle the guy behind the Power4.
Comments.
I wonder, seeing as IBM is designing the cell, if it will use the AltiVec unit for its SIMD? It could use a re-design for higher performance, and I hope we see that. Also, I wonder if we will ever see a dedicated, large cache attached as well?
And what does "tape out" mean? I never did find a definition for it.
that's when the design team....
generates a data tape containing the laid-out layers of an IC and hands it to the fab guys. Photomasks are produced from the tapes, a multimillion$$ (in this case) exercise which you wanna be sure is worth it. Word is that Cell is months ahead of schedule, a phenomenon practically unheard of in the ASIC world.
Don't think it's in S0ny's political bent to use a chip that runs alTIV3c. They're very NIH. They have no choice but to work with IBM since they're prob the only house which can fab this IC. They also have the most experience on multiproc ICs with Power4/5. +05Hi8@ is in this cuz they did the PS-2 E-Engine and they're probably handling back-compatibility. The PS-3 has a E-Engine inside its I/O proc, just as the PS-2 has a PS-1 in its I/O proc.
except for the motion of
the fishies in Finding Nemo, I've been consistently disappointed by this summer's CGI work. They're obviously not using nearly enough cpu to be convincing, yet the CG budget on Matrix 2&3 is apparently ~$110M. Something is out of whack methinks, since $100M is what IBM is spending to build Blue Gene - a bleeding-edge peta-FLOPS platform requiring research and hundreds of engineers designing novel things over 3-4 years.
Big-budget Hollywood movie production practices seem terribly inefficient. Most of the money ends up in various pockets, ie. some people's margins are huge. Perhaps a good way to launder money.