Page 5 of 12 FirstFirst ... 34567 ... LastLast
Results 81 to 100 of 222
  1. #81
    I tried a little overclocking with my 750Ti.

    EDIT : be careful if you want to overclock, driver 337.50 beta limits core clock to +135MHz

    Using driver 335.23 the card reaches 1289MHz (boost 1367MHz) GPU and 1525MHz memory without voltage increase.
    It gives 50.54s for Mike Pan's BMW with 256x256 tiles (instead of 57-58s with Blender 2.70a at stock clock)

    I got 15:46.65 for the Cavalier.

    @Rolf: how much do you overclock to reach 14:45.41 with one card ?
    Last edited by Crus-T; 02-May-14 at 18:21.



  2. #82
    are the downloads from Graphicall.org compiled for MAXWELL ?



  3. #83
    Member zajooo's Avatar
    Join Date
    Mar 2012
    Location
    Slovakia
    Posts
    174
    Originally Posted by Rolf View Post
    Here my build, i use primary. Blender 2.69.11 eb4f2b4 http://goo.gl/c2cjtF on newer build, the utilization of 2 or more 750ti´s look like this =>
    Attachment 305943

    It looks very suspicious, Rolf do you think it can have something to do with this 2.71 optimalization, that does ...

    • CUDA handling is more asynchronous now. This results in a lower CPU usage while CUDA rendering with multiple GPUs


    devs can be found here ... https://developer.blender.org/rB1d016758330b
    Last edited by zajooo; 03-May-14 at 07:54.
    Part of a small team at visarts.sk

    my CV | Tutorials (SK) | Tutorials (ENG)



  4. #84
    Member zajooo's Avatar
    Join Date
    Mar 2012
    Location
    Slovakia
    Posts
    174
    Originally Posted by Crus-T View Post
    Those cards are CUDA 6 capable, what about unified memory ?
    it has nothing to do with HW, it is a programming stuff, to adress memory same way on RAM and VRAM
    Part of a small team at visarts.sk

    my CV | Tutorials (SK) | Tutorials (ENG)



  5. #85
    Member
    Join Date
    Feb 2014
    Location
    Germany
    Posts
    24
    Originally Posted by zajooo View Post
    It looks very suspicious, Rolf do you think it can have something to do with this 2.71 optimalization, that does ...

    • CUDA handling is more asynchronous now. This results in a lower CPU usage while CUDA rendering with multiple GPUs


    devs can be found here ... https://developer.blender.org/rB1d016758330b
    Yes, I think it has to do with this.
    So, i made some tests.
    I use mib2berlin´s Cornellbox 2.7x benchmark.

    Blender 2.69.11
    1x 750ti = 2m43s 2x 750ti = 1m24s Speedup with 2 cards ~94%

    Blender 2.70.5 fd80ac4c3e84
    1x 750ti = 3m52s 2x 750ti = 2m37s Speedup with 2 cards ~47%

    Blender 2.70.5 fd80ac4c3e84 without the commit you post.
    1x 750ti = 3m52s 2x 750ti = 1m58s Speedup with 2 cards ~96%



  6. #86
    Member zajooo's Avatar
    Join Date
    Mar 2012
    Location
    Slovakia
    Posts
    174
    Great, thanks for feedback. Go ahead and post it like an issue on developers page !
    Part of a small team at visarts.sk

    my CV | Tutorials (SK) | Tutorials (ENG)



  7. #87
    Member
    Join Date
    Oct 2013
    Location
    Alsace
    Posts
    38
    Also posted on the bug tracker here https://developer.blender.org/T40027

    Hi,


    I’ve tested with two 750Ti cards not connected to any display, blender-2.70-30361a7-win64-vc12 and I get the following times for the cornell_bench_27:
    One card 4’19"
    Two cards 2’11"


    Cheers!



  8. #88
    Originally Posted by Tibi View Post
    Also posted on the bug tracker here https://developer.blender.org/T40027

    Hi,


    I’ve tested with two 750Ti cards not connected to any display, blender-2.70-30361a7-win64-vc12 and I get the following times for the cornell_bench_27:
    One card 4’19"
    Two cards 2’11"


    Cheers!
    Thx for posting this. I will look into finding a good solution asap. But else we will have to revert the async fixes again and just take the cpu burning as that seems less bad then having the least of both sync and async cuda device handling



  9. #89
    Originally Posted by Rolf View Post
    Mike Pan BMW Benchmark

    Based on my own build, Revision 61305
    Build with VS2008, Scons, Nvidia Cuda Toolkit 6.0 RC


    System:
    CPU: Intel [email protected],5ghz
    GPU1: Nvidia GTX 570 @732/950/1464
    GPU2: Nvidia GTX 750ti @(1085)~1228/1375
    OS: Windows 7 64bit | Blender 2.69.11

    GTX570:
    Default settings: 83.14 s
    128x128 tiles: 63.77 s
    256x256 tiles: 48.84 s
    512x512 tiles: 49.92 s

    GTX750ti:
    Default settings: 68.87 s
    128x128 tiles: 62.34 s
    256x256 tiles: 55.25 s
    512x512 tiles: 64.19 s


    GTX570+GTX750ti
    Default settings: 39.27 s
    128x128 tiles: 32.28 s
    256x256 tiles: 27.29 s
    512x512 tiles: 31.12 s
    There is just no way currently that I'm getting those same results as you.
    Currently my cpus out do the 750ti.
    The problem with all these benchmarks is I'm not completely sure everyone is using the same render specs. The BMW scence has a note that GPU yielded 53secs, I never got that either with my 580 (old machine).

    I have a GTX580 lying around, but currently pairing the 750ti with a quadro FX580 (display).
    Playing games (steam on linux - Metro Light, Half life 2) seems almost on par with the 580.

    Currently:

    BMW1M-MikePan.blend: default (@200 samples)
    CPU: 01:30.90
    GPU: 02:19.68

    pavillon_barcelone_v1.2.blend (@100 samples)
    CPU Benchmark (scene): 04:49.02
    GPU Benchmark (scene): 07:31.40


    System Specs:
    OS: Linux (debian)
    Blender version: 2.70 (not 2.70a but downloaded that and got simular results)
    nvcc version: release 6.0, V6.0.1
    CPUs: 2 × xeon E5540 (quad @ 2.53GHz)
    GPU: 1 × GeForce GTX 750 Ti (2048 MB, 640 Cuda Cores, Graphics Clock 1320Mhz - 1450Mhz MAX, Mem Clock 5400Mhz)
    Ram: 72Gb

    Am I doing something wrong? I looked into building blender myself, but didn't see any cuda dependencies (quick look - I could've missed something). So I thought I'd ask here.

    Some suggestions would be appreciated!
    Thanks ☺



  10. #90
    Member mib2berlin's Avatar
    Join Date
    May 2008
    Location
    Germany
    Posts
    3,507
    Hi, the official 2.70/70a doesn´t support the GTX 750!
    The older benchmark files has bad setup for GPU so many user change the tile setting.
    You can try my bench special for GPU:

    http://www.blenderartists.org/forum/...2-70-Benchmark

    To compare:

    http://www.blenderartists.org/forum/...=1#post2627177

    I think all user use default setting because all other setups are slower .

    The builds from http://builder.blender.org/download/ support GTX 750.

    Cheers, mib.
    OpenSUSE Leap 42.1/64 i5-3570K 16 GB Blender 2.7 Octane 3.03
    GTX 760 4 GB, GTX 670 2 GB Driver 375.26 | Blender for Octane



  11. #91
    Hi,
    Interesting. It appears that my runtime built cuda kernel was hopeless. (cycles_kernel_sm50_CBEC943857C11B0474F9F57C8809F0 EF.cubin).
    I found a link to a user generated kernel linked online (kernel_sm_50.cubin - sorry i forget but thanks!)

    So I put it in the scripts/addons/cycles/lib and my benchmarks are as expected. So much appreciated to the person that supplied the kernels.

    @mib2berlin, thanks for the links, I'll check them out.

    update: It appears the supplied cuda (sm_50) kernel that comes with the latest build is not bad but slower than the one I got online. Interesting that not all kernels are the same then...I'll compile my own from the build...

    cycles_cornell_bench_27.blend on todays build 2.70 be980b9
    Generated kernel: 04:08.81
    Build kernel: 04:08.01
    Downloaded kernel: FAIL refused to render with kernel. (interesting! rendering on current release - see below)

    cycles_cornell_bench_27.blend on stable build 2.70a
    Downloaded kernel: 2:52.53

    So the supplied kernel with the current builds..pretty much suck for linux (that's my summary). Both build supplied and generated (cuda 6.0) are not up to task of this one. (binaries differ)
    This downloaded kernel is a keeper! (thanks to ever supplied it - forgot sorry).

    Does anyone want to comment? I have no idea. I suspect compiler issues (gcc?).

    ALl the best and thanks. If anyone want's this kernel (linux) let me know.
    Last edited by joules; 16-May-14 at 10:27.



  12. #92
    @Joules : 2:52 for Cornell Box with 2.70a + kernel looks ok to me.
    I got 2:58 with my 750Ti (MSI OC) and below 2:30 with a little overclock.

    You should try the BMW benchmark at 256x256, you should range in the 50-60s.



  13. #93
    Yes 2:52 is fine I'm sure it would be less if not also used as the primary display. The issue is that this is a different kernel that I found online and not one that came with stock blender. I'm curious as to why the daily builds refuse to render with it. you can basically drop in the kernel_sm_50.cubin file to the current release (that doesn't come with sm_50 kernels as default). For excellent results. I'm going to track down the kernel.cu file that yielded this one and compile it myself.

    I'm unsure where it's capped by the cuda compiler (nvcc) or blender...saving the tin foil hat till last. Because these results for a under $200AUD card are pretty damn good.



  14. #94
    Hi there!

    I upgraded my old Rig (2x gtx 480) with a single gtx 750ti. Everything works fine as single GPU. But when I use the combination of all three GPUs, cycles begins to build the BVH etc, and the renders nothing, juts a blank image. I use the daily Build from the 16th of May. Any Ideas?


    PS.: small projekts (Mem at 65.xxM or so) render fine with the three GPUs, big Projects take more Memory than before and get stuck!



  15. #95
    Originally Posted by zuggamasta View Post
    Hi there!

    I upgraded my old Rig (2x gtx 480) with a single gtx 750ti. Everything works fine as single GPU. But when I use the combination of all three GPUs, cycles begins to build the BVH etc, and the renders nothing, juts a blank image. I use the daily Build from the 16th of May. Any Ideas?


    PS.: small projekts (Mem at 65.xxM or so) render fine with the three GPUs, big Projects take more Memory than before and get stuck!
    I think you can gather from the last few pages of this thread is something aint right with the cuda rendering with daily builds, especially with the 750ti. I suggest you go back to the release build (2.70a which I tested as working with a drop in kernel_sm_50.cubin) and I'll edit this post with a link to a good kernel_sm_50.cubin kernel (../scripts/addons/cycles/lib).

    Later when I have time I'm going to track down at what caused it to go south (in render times).

    Also please test your GPUs with the cornell box benchmark.

    Kernel Downloads (martjin berger):
    Windows kernel_sm_50.cubin
    Linux kernel_sm_50.cubin

    Note these kernels do not work with current builds (as tested). Please use with releases <= 2.70a.

    Joules
    Last edited by joules; 16-May-14 at 20:42.



  16. #96
    CUDA handling is more asynchronous now.
    Affects it only MAXWELL and multiple GPUs or any GPU???

    I have around 28% longer Render Time with GPU!!!

    While rendering the GPU Usage drops for several seconds to 56%

    WIN 7 64Bit
    GTX 650
    BMW1M-MikePan-Scene
    [email protected]

    blender-2.70-f93bc76-win64-vc12 (10.04.2014)
    2min 7sec

    blender-2.70-be980b9-win64-vc12.zip (Build from Yesterday 16.05.2014)
    2min 57sec



  17. #97
    Theres a few things going on.
    I manage to build my own kernel_sm_50.cubin for builds 2.69, 2.70a (releases) and master (daily builds).
    1st - and the most obvious, don't run daily builds theres no real benefit - there are changes to the cycles renderer that are probably on going.

    2nd - I'm getting some success building my own kernel based on the master branch that is faster than the default. The same compiler flags for release 2.70a yield a render that is a little slower on the master branch - 3:13s vs 2:51s (cornell_bench_27) - but my build is slightly better than the default on both versions. However 20s can be quite significant on larger scenes.
    However kernels are incompatible with each other anyway (2.70a kernel will not run in master branch builds (your dailys) vice versa). I'm unable to test with mutiple GPUs.

    @Dito, go back to 2.70a unless theres some specific reason.

    I'm tired now
    Last edited by joules; 17-May-14 at 19:21.



  18. #98
    Member
    Join Date
    Feb 2014
    Location
    Germany
    Posts
    24
    Here is a Blender 2.70.5 Build with nearly the speed like 2.69.11.
    Feel free to test.
    My time with this Build [email protected]/1502 @BMW tilesize 256x256 ~00:54.60.

    Blender 2.70.5 f574b1c
    Only with sm_50 kernel.



  19. #99
    My test for v2.70.5 on Mike Pan's BMW at 256x256
    (with 337.88 WHQL driver, i5-4670, win 8.1 64-bit)

    [email protected]/1350 (stock) 00:58:02
    [email protected]/1525 00:50:02

    Quite the same as with 2.70a... (respectively 00:58:10 and 00:50:62)



  20. #100
    Member cekuhnen's Avatar
    Join Date
    Mar 2003
    Location
    Grosse Pointe Woods, Michigan, USA
    Posts
    11,983
    Hi guys,

    I am not very familiar with the card differences.
    Is there much difference between the GTX 780 TI and the 750? The 780 is 5 times more expensive.
    3D Design Generalist - Faculty Industrial + Interior Design
    Chair Interior Design - Wayne State University



Page 5 of 12 FirstFirst ... 34567 ... LastLast

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •