Re: HPC vs. server workloads (was: Is a RISC chip more expensive?)
- From: nmm1@xxxxxxxxxxxxx (Nick Maclaren)
- Date: 27 Oct 2007 20:11:39 GMT
In article <47238933$0$26482$88260bb3@xxxxxxxxxxxxxxxxx>,
"Stephen Sprunk" <stephen@xxxxxxxxxx> writes:
|> "mike3" <mike4ty4@xxxxxxxxx> wrote in message
|> news:1193266031.358472.287060@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
|> > On Sep 12, 2:55 am, n...@xxxxxxxxxxxxx (Nick Maclaren) wrote:
|> >> In article <slrnfefa0f.9fo.fl-usene...@xxxxxxxxxxxxxxx>,Florian Laws
|> >> <fl-usene...@xxxxxxxxxxxxxxx> writes:
|> >> |>
|> >> |> I don't think one could see a difference between a Mandelbrot
|> >> |> fractal rendered on a single processor compared to one
|> >> |> rendered on (in the extreme) one processor per pixel. (On still
|> >> |> images, at least)
|> >>
|> >> I was referring to more important film animations, not Mandelbrot.
|> >
|> > So then for film animations does one need a complicated, custom-
|> > built interconnect?
|>
|> For a movie, no. Each frame is rendered independently; if one cares, it's
|> possible to break up frames into separate rendering jobs as well. If it's
|> done right, there is no "jumpy" or "glitchy" material. The only cross-frame
|> computation comes when you (a) set up the rendering and (b) need to compress
|> the fully-rendered work, and both can be done in better than real time on a
|> single CPU these days. (vs rendering, which can take minutes to hours of
|> CPU time per 1/24th to 1/30th sec frame)
That assumes that the resolution of the picture is better than visual
acuity, in position, intensity and colour. There is, even theoretically,
no way to avoid ALL temporal glitching when that is not the case and each
frame is rendered entirely separately.
However, 'smoothing' such effects can be done post-hoc, and is even
cheaper than compression. Indeed, it can be done (and, I believe, is
done) as part of compression.
But getting back to my point. I was trying to point out that the
extreme parallelisation of "one processor per pixel" is feasible for
things like Mandelbrot, but not real pictures. If you split up the
rendering of a single frame that far, your communication requirements
go through the roof. Or you start glitching appallingly.
Regards,
Nick Maclaren.
.
- Prev by Date: Re: HPC vs. server workloads (was: Is a RISC chip more expensive?)
- Next by Date: Write-Allocation on modern CPUs
- Previous by thread: Re: HPC vs. server workloads (was: Is a RISC chip more expensive?)
- Next by thread: Not a homework question
- Index(es):
Relevant Pages
|
