Re: Does RISC still offer a significant numbercrunching advantage?

"mike3" <mike4ty4@xxxxxxxxx> wrote in message news:101c7d96-aecb-4162-b124-987c0d8319e0@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
On Dec 8, 1:45 pm, "Stephen Sprunk" <step...@xxxxxxxxxx> wrote:
Again, it depends on how tightly coupled the workload is. For
some tasks, one 32-CPU SMP machine will perform better than
a cluster of 256 single-CPU machines, even if the CPUs are
identical between systems. On other tasks, the cluster will blow
the SMP out of the water. It all depends on how much the CPUs
need to talk to each other.

So, what about that movie FX rendering I discussed? What does a
*high*-end system like they use at ILM have in it -- the systems
they use nowadays, anyway? Do they use a single gigantic SMP
machine or a cluster of many small (either uniprocessor or 2/4-
processor SMP) machines? And what CPU type do they use? Is
it x86?

My understanding is that ILM, Pixar, et al use massive clusters (thousands of nodes) with Opteron chips. Or at least they did a few years ago; they might be using Intel chips now, or some mixture.

Rendering movie frames is embarrassingly parallel (i.e. no coupling at all) once everything is laid out, so there's no need for massive SMP; a cluser is cheaper.

Also, what about the CPU type for workstation systems used to
do the 3D modeling and stuff? Is a RISC-based system like the
POWER6 blade that was mentioned here be the best choice for
that? Or would good ol' x86 (especially with the dual/quad core
chips they've got now) do the job? Especially with a good GPU
in there too?

Having watched a few of the "making of" extras included on DVDs, it appears that they use high end x86 desktop machines for layout, character and scenery design, etc. and small low-res, low-fps tests. They use the cluster for full-resolution, full-frame-rate rendering of larger tests and the final movie.

For instance, the guy who's designing the landscape or some character's clothing doesn't need a full rendering to know how something looks (a few fps at low res is fine, and a good GPU can do that), and the director is only looking for staging, camera position, lighting, voices, etc. with just enough of a clue on the individual parts to know whether they fit together and how the story works. The vast majority of that work will be done with simple wire frames so that changes are cheap and the first full render will be as close to the final product as possible. Rendering a full movie takes _months_, even with thousands of nodes working on it, and even slightly screwing up a short scene means they have to add in another few weeks of rendering time to try again.

S

--
Stephen Sprunk "God does not play dice." --Albert Einstein
CCIE #3723 "God is an inveterate gambler, and He throws the
K5SSS dice at every possible opportunity." --Stephen Hawking

.

Relevant Pages

  • Virtual Shared Memory Cluster
    ... I am currently running commerical CFD software that only has support ... I have a cluster of machines that I ... into a single virtual SMP system with a virtual shared memory. ...
    (comp.parallel)
  • Re: 16-Node Parallel System
    ... > We need a cost-effective 16-Node compute cluster for Academic ... Then you use any O/S kernel that is SMP aware, ... additional CPUs are added to the node). ... the interconnect of a SMP's 16 CPU interconnect will not ...
    (comp.parallel.mpi)
  • Re: e3k upgrade - hardware spec
    ... > Xeon 2.8Ghz processor with 2Gb of RAM. ... > The front end one is a slower machines which is currently setup to send all ... We also have NetApp IP SAN so the new cluster will have its ... With Exchange 12 on the horizon, you'll probably want to take advantage ...
    (microsoft.public.exchange.setup)
  • Re: e3k upgrade - hardware spec
    ... >Xeon 2.8Ghz processor with 2Gb of RAM. ... >The front end one is a slower machines which is currently setup to send all ... We also have NetApp IP SAN so the new cluster will have its ... Rather than use two SG's that are packed with stores, ...
    (microsoft.public.exchange.setup)
  • Re: Network Load Balancing Problem
    ... Are you trying to test NLB from one of the NLB machines? ... MVP - Windows Server - Clustering ... With NLB manager in the Server1 i create new cluster ... the cluster response only with the second host, ...
    (microsoft.public.windows.server.clustering)