A Tale of Teslas

by - 2012/03/31 28 Comments Production, Random

Yesterday the courier arrived and we received the two – sponsored by Nvidia – Tesla computing units we’ve been waiting for days! It’s been a nice afternoon at the institute, with Sergey configuring the correct drivers in order to unleash the power of two cards and Ian making fruit smoothies to relieve the tension. More interesting information comes after the documentary “A Tale of Teslas”.

If you skipped the video, here the render time for the BMW benchmark file:

  • 1 Tesla: 1 min 4 secs
  • 2 Tesla: 34 secs

Although this result is not really comparable to the one obtainable with a GTX 580 (23 secs),  there is a main difference: Tesla cards have 6GB of memory, each. This means that in the dual configuration we could render a test scene (the bridge scene) with several GB worth of textures and data (otherwise impossible on a GTX).
The time for rendering such scene was:

  • Dual Tesla config: 5 mins
  • Dual Xeon rendernode: 15 mins

We are going to get Brecht over here very soon to see if he can improve the performances of Cycles on this cards. Will we eventually be able to replace the whole renderfarm with a GPU-based solution? We don’t know yet!

Currently we’ve put the cards next to more humble GTX 260’s in two workstations, this will help designing a lot. Special thanks to Nvidia for giving us these cards to test!

  1. DingTo says:

    Yes, yes, yes! We definitely need more Cycles speed! :)
    I hope Brecht can speed it up, thanks for the updates!

  2. Tycho says:

    whaa great joke :)
    I have been trolled … when you say would it render at least as faster I thought u are really speacking the truth : well done guys …

    • Tycho says:

      my answer mean nothing :)
      I would say you have trolled us saying two nvidia will render slower as the current render farm :)
      but it’s all the better :)

      • ton says:

        Current estimate: two Teslas have about the same OpenCL Cycles render power as 4-5 i7 nodes. I think we can optimize things to replace 8 or more even. If it would replace 16 nodes we’d have a very unhappy Justcluster builder… :)

  3. Idur says:

    What exactly Tesla cards did you get?

  4. David Black says:

    How about some proper optimization, bidirectional path tracing, photon map and irradiance caching for final gather? :-)

  5. Chris says:

    Can I borrow one? :D

  6. Ben says:

    “Let’s first read the manual…” Yeah right.

  7. Tycho says:

    Just for the men who are too lazy : I sum up the card capacity.
    So
    1 GPU (C2070/C2075) is :
    * 575 MHz clocked for each core
    * shaders : 448 at 1150 Mhz
    * memory : DDR5 (144Gb/s max) 6Go but ECC takes 12.5 % of this if activated (is it here ?) clocked at 3000 MHz
    * processing power : 1288 GFlops
    So, comparing to one set of xeon (a dual CPU) :
    * processing power : 25-30 GFlops
    So a single card will replace 4*2 xeon (at least since memory bus really higher in NVIDIA cards).
    I bet that at the end of mango, you will render on many 2 teslas + 2 xeon racks connecting each other by InfiniBand :)

  8. Phil2.0 says:

    LoL, you just added more pressure on Brecht than he already has.

    poor Brecht, hope he won’t Break! (sorry :) )

  9. Joanna says:

    Exciting! Are you guys using the Nouveau driver?

    • T_X says:

      The (very awesome FLOSS) nouveau graphics driver does not have opencl support upstream yet, does it? (I hope Nvidia will get their guys on the nouveau driver soon instead of their own proprietary, closed-source one…)

  10. Ace Dragon says:

    It seems like those would be ideal cards for the movie project itself, but I sincerely hope it doesn’t mean that Cycles optimizations for CPU rendering will go to the backburner.

    Personally, there are many optimizations techniques for the CPU that the Cycles engine currently does not do, if Brecht develops Cycles optimizations designed mainly to work on the Tesla cards, then it would seem to be similar to charging a fee of thousands of dollars for the ability to get stunning results out of it (as the vast majority of Blender users do not have Tesla’s nonetheless two of them).

    Personally I would think that the Cycles engine would not be very ‘open and free’ like the rest of Blender if it was designed to require the absolute best hardware to use.

    In all that that, I would definitely want to see a focus on getting Cycles to render faster on the CPU, in all respects most of that should also work on the GPU’s.

    • ton says:

      If the optimizations are on OpenCL level, everyone will benefit. Intel is spending a lot of time on getting this work perfect for cpu.
      Apart from that: building something today for such high-end hardware means we have something very cool finished & tested for the masses in a few years :)

      • Gill says:

        Like Ace said, the road to optimizing cycles against the most expensable video cards out there nobody in the blender community owns is scaring me. Even more considering how much bottleneck is a gpu for cycles development and how much time it needs for a feature to be implemented in gpu too instead of for cpu only.

        I’d really like to see optimizations for cpu before going on that hard and uncertaint gpu road.

  11. Jan the Freeze says:

    Nerd summer camp!

  12. ctdabomb says:

    maybe it should be called A Tale of Two (cities) Teslas?

  13. Nik says:

    Emo, it’s a crap!

  14. Florisn says:

    The video shows something else that’s very interesting: contrary to all former evidence, it’s actually physically possible for girls to enter the blender institute! I’m pretty sure I spotted one in the background! Who’d have thought…

    • T_X says:

      Hehe, true :). But there was also Angela Guenette before @Sintel – who mysteriously disappeared (I was a big fan of her and her work)… Maybe the question is not whether they can enter the Blender institute, but whether they can also exit it alive again? ;)

  15. T_X says:

    Blender needs grid computing support :). BOINC integration or something like that. I bet there are a lot of people who would love to contribute CPU/GPU cycles. It’d match such a unique, awesome, community oriented project perfectly!

  16. Malcolm says:

    Why Nvidia gave only 2 teslas? Only 2?!?!
    Come on NVIDIA!!

  17. Kyle says:

    I would like to know what font was used for the various title screens. Could someone help me out/point me to where I might download it?

  18. YAFU says:

    “this result is not really comparable to the one obtainable with a GTX 580 (23 secs)”
    In the BMW benchmark spreadsheet nobody is even close to this time with single GPU. ¿Had you use two GTX 580 cards?

  19. Vince says:

    I have been searching endlessly thru many forums for getting my Tesla k20c to work with blender. Sorry about the large date difference between your posts and mine but I could sure use the help.

    I can get my GTX card to appear in Blender 2.71 but I can’t get my tesla card to even appear in blender. Do you have any suggestions on how I can get my Tesla to work?