Probably it has been asked already, but what is the best setup for wm? my understanding is that the gpu is not used at all in the output building.
So should I go for a xeon since it has more threads than i7? or is there a particular cpu that works better because of particular instruction sets or other devilries?
There is no universal answer to this question, but if there was one it would begin with ‘What is your budget’?
Well since remnant finally made voronoi and expander to also use muliple threads, and WM can build multiple devices at a time, id probably go for some cpu with tons of cores, perhaps an amd?
However, cores arent everything ofcourse, Intel has been king for single core performance for a long while now.
My i7 3770k @ 4.5ghz simply blasts away any of my projects.
Id probably go for an intel with 6 cores and HT especially so if WM isnt the only cpu intensive app youre going to use. I would also worry about memory usage tbh, without sufficient memory you will start to page to your virtual memory, which is not TOO bad if ya got an ssd, but still. It also depends on project size, my projects usually peak out at around 20 GB so i upgraded to 32gb memory.
Ofcourse efficient projects will use less. Like device sharing etc will really reduce this.
i dont reccomend an AMD 8-core (for heavy WM usage) although im using the standard version not the pro version so i dont think it even takes advantage of the power… personally im upgrading when the haswell-E series i7’s come out.
Pro version can and will use up to 32 cores assuming you have them, at the end its single core power * number of cores = speed 
I dont think I could live without the pro version, it would make the compiling double as slow at least :o
Anyone had experience with hpc clustering? I mean version 3.0 will support up to 256 threads: as far as I know you need clustering, you’re not going further than 40 cores with a multi-processor/multi-core machine.
I would also consider per-core clock while getting a processor for WM. Also, amd is crap for any serious professional work, heating problems and what not. I personally prefer intel I7 or xeon if you can afford it. Remember, processor is the most important investment you will ever make, so don’t go cheap on it.
But is it correct to say that GPU is not used at all (apart from the realtime 3d view maybe)? also, how can I actually use the 256 max threads of version 3.x?
Gpu may not be used in world machine, but it would make your life infinitely better if you have a mid range nvidia gpu. That way you can use your full memory in wm. And multiple displays of course!
For utilizing 256 threads, you need to have 256 logical processors. But that would mean a tremendous investment on at least 13 processors, and a distributed systems setup (don’t have much experience with that). I wouldn’t recommend investing that much for just world machine, unless you know how to make money off it.
Well, not just the ‘3D’ part is using your Videocard extensively, also the 2D factor, Layout viewer etc al make quite some use of your GPU, but i guess it kind of depends on what quality/resolution you set it to in WM options too, its like Photoshop using Opengl
I started looking at azure and amazon ec2, they offer good prices, and wm is a perfect example of an application that would benefit from a batched processing.
The only thing that is not clear is if only “hpc-native” applications can be batched or if clustering is a sort of layer that sits between the cpu and the application, thus letting you run every application in a cluster.
I plan on signing up with amazon ec2, but don’t know if it works with wm and vue/terragen. If you use it then please let us know too.
I believe you would just need to get a Windows Server EC2 instance, then use a VNC solution to connect and give you a GUI to interact with. I assume that WM will run on Windows Server, it’s pretty much the same kernel and APIs.
So it seems that Windows HPC only allows cli commands. This leads me to the frightening world of wm automation.
I guess the workflow is setting up a cluster (more or less trivial), and then scripting each instance so that it renders a batch of tiles.
I’d like to know which kind of wm license is needed for this, the site licence?
Also the model offered by Azure is quite simple, you create a batch of virtual machines, you cluster them, you run jobs on them. Amazon EC2 is still mysterious to me, in practice you borrow computational power when it’s available? so you don’t actually install wm on all nodes?
I tested wm on an EC2 ubuntu node, and it works like a charm even under wine!!
It actually uses more since the last dev version (up to 256 threads), i can confirm it works pretty good for most devices even at a high thread count (80+).
Budget is the limit, if budget is unlimited then the best thing on which to run WM today would be a server with a 8 socket motherboard and 8 of those 15 core 30 thread monsters : http://ark.intel.com/products/75255/Intel-Xeon-Processor-E7-8870-v2-30M-Cache-2_30-GHz for a total of 240 threads in parallel along with 2TB of memory. But such a huge server only makes sense at very high résolutions (128 000 +) for “normal” 1 to 8K résolutions anything will do the job, go toward intel at equal core count because of HT, make sure you get enough memory as even at 8K it gets hungry on ram on complex graph and ram is pretty cheap nowadays (don’t get less than 16 gig, preferably 32).
As i pointed in my just made post, you can get a machine that will max this. Also clustering doesn’t make any sense for a non distributed system, WM is a single program, it’s not made for clustering and there’s no way to make it automagically act like it is even if an OS virtualized ressources across a network (WM is very data intensive too and it would spend most of it’s time with CPUs waiting on data transfers). It will also cost less to build a single very heavy machine vs plenty of small ones give better performance and allow you to run massive non tiled (thus coherent even on nodes that require neightboring information like erosion) builds.
Even a 256 thread monster doesn’t cost THAT much compared to what high end servers used to cost (maybe 30-50K$ range tops) and at 10K you can get a decent middle ground, if not willing to spend that much you can rend a fairly decent server (40 threads / 256GB ram) at ovh for around 300€ before taxes / month iirc (can also rent it just a week for a worse price / time ratio but good if you need to just run a build).