I wouldn't even say it's generally true. The memory footprint of a scene to be rendered is hardly ever that directly related to the footprint of the scene in the DCC client. Acceleration structures paired with what global features you use alone can reduce or multiply the footprint at any given bucket fetch by an order of magnitude. How the engine deals with the memory/recompute trade-off also makes a considerable difference, as well as how it passes and chooses to fetch and dump data from the client.
To top the lot off, even simple things like textures hardly have the same impact, especially when you look at how certain threading models can multiply those like rabbits or not (e.g. PRMan for the longest time would absolutely brutalize your RAM for practically any threading unit it spawned), and how the same thing is usually reduced to proxies first in the DCC app. The DCC client also doesn't have to keep pre-allocated considerable amounts of space for multiple raster-like buffers, something that engines that keep them live for AOV injection deal with, some of them popping a handful of gigs right away just to pre-allocate those if memory is available to do it, but it might keep around other large stashes of caches that a renderer won't even remotely care about. So all in all I still have to experience a relationship that direct between "scene size", which is hazy to begin with, and memory footprint at render time, that's long gone in these days of PPS uber shaders and PBL rendering. On Wed, May 21, 2014 at 3:49 AM, Steven Caron <car...@gmail.com> wrote: > while generally true, there are many other memory consuming parts of the > renderer... various cache types (irradiance, photon, sss), textures, > aov/image buffer, and ray reserve memory. > > > > On Tue, May 20, 2014 at 5:17 AM, Stefan Kubicek <s...@tidbit-images.com>wrote: > >> >> Translated scene data is usually a lot more efficient to store than the >> actual, raw scene, which has a lot of metadata, construction history, etc, >> that is not needed for rendering. Hence, a scene consuming 10gb in RAM is >> usually easily renderable on a 6gb card, depending on how much additional >> unique and procedural geometry you create at render time (e.g. subd, >> displacement, hair). >> > -- Our users will know fear and cower before our software! Ship it! Ship it and let them flee like the dogs they are!
