networking multiple 2008 8-cores

Discussion in 'Mac Pro' started by jpine, Dec 5, 2013.

  1. jpine macrumors 6502

    Joined:
    Jun 15, 2007
    #1
    This is more of a "what if" question than anything else since I have very little experience with networking or "render farms", so here it goes.

    I can pick up 3 2008 8-core MP for under $2000.00 USD and network them to my current 2008 MP. Assuming I add a really good MacVidCard or two, how would this rig stack up against what we little we know about the new 6-core or faster MP which is yet to be released. My apps are Photoshop, Lightwave, and Vue (and occasionally Maya). In PS, my images are generally no smaller than 2048x1536 when building iOS apps. I do shoot RAW images as well. Lighwave and Vue sequential images as standard HD (1920x1080P). Vue tends to have lot of polygons and high res texture maps and is a big resource hog.

    Thanks!

    JP
     
  2. MattDSLR macrumors 6502

    Joined:
    Jan 23, 2011
    Location:
    Canada
    #2
    Interesting question
    looking forward to responses
     
  3. flat five macrumors 601

    flat five

    Joined:
    Feb 6, 2007
    Location:
    newyorkcity
    #3
    network rendering happens via render nodes which are written/supplied by the specific software developers.. (devs will usually throw in at least a couple of free nodes upon package purchase)

    ie- you don't just plug two computers together then push the render button and expect the slave to do anything.
    the slave needs to have specific software installed on it (the node) and the separate computers need to link up via that software..

    vue and lightwave are both capable of network rendering.. photoshop? i really don't think so but could be? i quit photoshop a few years back so i'm not sure what type of rendering abilities it may/may not have these days

    that said, with your gripe about vue-- "Vue tends to have lot of polygons and high res texture maps and is a big resource hog. "..
    no amount of additional cpus is going to improve that type of scenario since it's all (mainly) happening on one single core..

    network rendering is for rendering improvements and polycount etc doesn't have much to do with the actual render process once the scene is loaded..

    for instance, i could have one single polygon of complex glass which may take 10 hours to complete whereas i can have a million polys of clay which could finish in an hour..
    and networking will help decrease those times.
     
  4. jpine thread starter macrumors 6502

    Joined:
    Jun 15, 2007
    #4
    I forgot to say that I would add at least 2 Nvidia 570's or, preferably, 580s, probably in a used expansion chassis.

    All that being said, I can't see my wife letting me have 3 or 4 MPs taking up space anywhere in our house....not even a closet. :eek:
     
  5. seveej macrumors 6502a

    seveej

    Joined:
    Dec 14, 2009
    Location:
    Helsinki, Finland
    #5
    Hi,

    Nice, me also wants three more octos..

    As already noted, most softwares which allow for such render-farming or essentially: "automatic subcontracting of work" need to support the process "out of the box".

    I'd have to say, as a general rule: If you have to ask, you do not need a render farm.

    But wait, there are other ways...
    Assuming you have a lot of CPU/GPU intensive work, which is not getting done, or is currently precluding you from doing other things (such as browsing MR) you can set up dedicated headless machines and (as long as the process is manageable under VNC/ARD) control all those activities from your main desktop.

    I have an ex-colleague who (after going independent) does exactly that. His setup is based on (well, what do you know) four 2008 Octos - One acting as general purpose desktop and controller, one acting as file server and two machines doing the work. His (controller) desktop setup features four monitors - two monitors are used for "local" work, whereas the other two each show (using ARD) the desktop setup on his "render machines" (One working mainly with rendering, the other mainly working compression). I can not remember whether he merely uses GLAN for intra-machine comms, but I assume so, because this was his "on the cheap" solution.

    The potential problem with such a setup is, that you (the user) still need to co-ordinate the tasks for the machines quite manually, but as long as the duration of your interaction with the computer (in order to set up a task) is less than the time the computer spends working independently on the task, adding another computer to the "farm" becomes an economically viable option.

    Needless to say, unless you actually need to put a lot of drives in your machines or need exchangeable GPU's, this setup can be made using Mini's...

    RGDS,

    ----------

    I know the feeling, but that is mainly because I do not work from home and computers are only a hobby. If my computers were actual breadwinners and my home was my office, there's no chance I'd let my wife stand in the way of a good setup...

    RGDS,
     
  6. Tutor, Dec 16, 2013
    Last edited: Dec 16, 2013

    Tutor macrumors 65816

    Tutor

    Joined:
    Jun 25, 2009
    Location:
    Home of the Birmingham Civil Rights Institute
    #6
    Depending on the GPU cards and render support applications you chose to use, Lightwave and Maya could perform many, many times faster on your four 2008 Mac Pro renderfarm than on a single top of the line nMP, unless you similarly outfit the nMP. Similarly outfitting the nMP will probably cost you a lot more money. Not many 3d applications use OpenCL for rendering - LuxRender is the only popular one that I'm aware of. CUDA running on GTX GPUs provides the fastest rendering capability when combined with OctaneRender for almost all popular 3d rendering applications [ http://www.otoy.com ]. With one top of the line CUDA GPU combined with OctaneRender, it will render a scene 10 to 50 times faster than a computer using only a single CPU for rendering. GPUs of the same type/spec scale linearly in OctaneRender: two render twice as fast as one; four render twice as fast as two; eight render twice as fast as four, etc. This applies when the GPUs are all in the same systems. OctaneRender seats are sold on a per system basis for about $500 (US), so one seat of OctaneRender costs no more to purchase for one Tyan Server that can house eight GTX Titans or GTX 780 Ti SC ACXs than it costs for a seat of OctaneRender to house one or two Titans or 780 Ti SC ACXs in an old Mac Pro. The seats can be moved from system to system in about an hour.

    A 2008 Mac Pro can house 2 top of the line CUDA GTX GPUs ( I have two of some vintage CUDA cards like the GTX 480C and 580C in my three 2007 Mac Pros; but I put my more powerful CUDA cards in systems that hold from three to eight on them). Thus, three Mac Pros could hold six such GPUs (2 a piece) and four Mac Pros could hold eight such GPUs. But you'd need an auxiliary power supply like the FSP Booster to power them for each system - that's the auxiliary PSU that I use.

    Octane supports, among many other 3d applications, Cinema4d, Lightwave 3d and Maya. In Cinema4d and Maya which have their own net render applications, one can set up a renderfarm that is GPU based, as well as one which is CPU based. But, I'm not certain about whether Lightwave has that ability - check with Newtek. Also, you can download the Lightwave OctaneRender demo and give it a spin if you have a qualified GTX GPU.

    CUDA also speeds up certain tasks in Photoshop (check Adobe's site for more info), but not most of its functions. However, CUDA cards have Open CL abilities also and as Adobe brings more and more CUDA, as well as OCL, assist to Photoshop, GTX cards will give you both options. I'm not aware of any net render application for Photoshop.

    PS - As an alternative, have you considered an external chassis [ http://www.bhphotovideo.com/c/search?ipp=100&Ns=p_PRICE_2|1&ci=13355&N=4041996716&srtclk=itemspp ] like the NetStore [ http://www.bhphotovideo.com/c/product/980261-REG/dynapower_usa_na255a_xgpu_netstor_6_slot_pcie.html ] that will house up to 4 double wide GPUs. That way you could put 2 top of the line GTX GPUs in your 2008 Mac Pro (along with an auxiliary PSU) and put four of them in the chassis (connected to your Mac Pro through the topmost PCIe slot). You'd then need only one OctaneRender License to have the rendering capability of up to 60 to 300 single CPU systems. You need to ensure that the chassis has at least a 1000 Watt PSU. The total cost of the chassis would be around $2.4K with host adapter and shipping, excluding the costs of the GPUs. It would also be a lot easier to set up and would not require your getting into network rendering. Moreover, it would likely be transferable to your next system.
     
  7. sirio76, Dec 16, 2013
    Last edited: Dec 16, 2013

    sirio76 macrumors regular

    Joined:
    Mar 28, 2013
    #7
    The 2008 MP are "slow" machines and also very inefficient for today standards, a single 6core nMP have about twice the render power and will be much faster in loading scene for rendering or manipulating big PS/3d files. GPU render will be interesting in the future, but today for me it's a non starter, I really doubt I can render a 50.000.000 polygons scene(I mean real geometry, not proxy) with the small amount of memory of today's cards. Also the quality and the flexibility of CPU based solution are still unmatched, that's why you wont find any big production firms using GPU rendering for their final shots. Probably if you are just doing small archiviz or products shot, then GPU solution will serve you quite well.
    If you want to increase you render power it's better to look into a small PC farm, do not invest into old tecnology(a 2007 MP is already considered vintage by Apple!). For about 2000$ you can start to build two render nodes i7-4930k based and they will be significantly faster than four 2008 MacPro when running in parallel. I've already built for myself two render nodes inside a G5 enclosure(yes, I don't want ugly PC boxes around), heach node it's about as fast as a 12core 2.66 oMP, and all I have to do is to turn the power on and launch Vray distributed rendering from my old MP(soon nMP).
    Photoshop will not take advantage of multiple machines, and since many of its operations are single threated you may want to look for fast single threat performance on a single machine, in that regard a 4/6 core nMP will do probably better than a 12 core one.
     

Share This Page