1) it slows down the most when sending geometry out of core, and frankly
even that ain't all that bad. 'Significantly' is relative here when it's
already that fast.
2) There is currently a hard-coded 4GB cap on on-board geo cache, to try
and balance performance in these early stages. So even if your card has
6+, it will send geo out of core if hits 4GB. The devs have talked about
upping this though, especially with bigger cards coming along.
3) I can't recall the exact numbers, but they're able to fit roughly
110-120M unique triangles in 4GB, that geo being simple geo with a
single UV mesh.
4) Biased.... man I'm not gonna split hairs on this. If that's a big
deal for you it's a big deal for you. Seems like the sort of thing
Archviz people would care about more than VFX or Animation folks.
Redshift does have a hearty Monte-Carlo-only mode that I am partial too
though....
5) I lit and rendered all 75 of my shots on Yellowday
<http://www.timcrowson.com/yellowday/> in about 30 days on 12 multi-GPU
boxes. The toughest shots' frame times averaged 45min/f, with most in
the 5-15m range, and several far faster than that. The densest shots
here used about 8-9GB total, IIRC.
6) As always, everything is relative to your business and what the needs
of your production are. No news there.
-Tim
On 10/3/2014 1:34 PM, Simon van de Lagemaat wrote:
At the cost of a smaller fast memory pool? From what I understand
Redshift slows down significantly when it starts using off board RAM?
I know this is one of the main reasons many other renderers haven't
gone to the GPU yet since cards with decent amounts of memory are
still priced far too high.
How is Redshift with overall memory usage? i.e. what is the memory
footprint per polygon etc. Also it's biased correct? I
become nauseous when someone mentions that word, is their approach
better than other IC approaches? Similar pitfalls?
I may try it this weekend on my 780gtx (only 3GB onboard) at home
since it's been getting good PR here. With the type of memory heavy
rendering we do for film and the efficient platform agnostic pipeline
friendly workflow we have with Arnold I don't see it being useful
outside some smaller commercial jobs but I'd like to get a taste of
the speed and interactivity. I suppose if we were a 5 person shop
again something like RS would be a godsend.
On Fri, Oct 3, 2014 at 11:06 AM, Steven Caron <[email protected]
<mailto:[email protected]>> wrote:
that extra GPU power is a dramatic increase in speed though.
--
Signature
