Which card is best for compute (GPGPU) right now?

Discussion in 'Video Cards & Monitors' started by toehold, Apr 18, 2012.

  1. toehold

    toehold Member

    Joined:
    Mar 10, 2002
    Messages:
    11
    Location:
    Melbourne
    We live in interesting times if i understand the current dynamics. If you go to video editing sites they recomment nvidia for compute minimum gtx 570 or 580s. These cards are EOL and there may be a small bargain window here.

    The new 600 series appear to have the compute function severely restricted, which will push compute use to professional workstation cards like the 4000 - a lot more $$$.

    On the other hand the new AMD cards from 7770 up appear to have the compute function significantly enhanced (but perhaps not mature in drivers or apps) and not restricted.

    This leaves me undicided in which way to go, should I look for a gtx580 bargain (but older tech), go for a AMD 7000 series card or be pushed into the w/station cards?
     
  2. bigheadache

    bigheadache Member

    Joined:
    Apr 23, 2005
    Messages:
    3,260
    WHat software are you exactly talking about? If your software is CUDA then buying a Radeon isn't going to be much use. I suspect the video editing sites you are looking at recommend Nvidia because they are using Adobe. If the software supports OpenCL then you can choose which to use.
     
  3. OP
    OP
    toehold

    toehold Member

    Joined:
    Mar 10, 2002
    Messages:
    11
    Location:
    Melbourne
    software

    yes I am using adobe, but also vReveal and sony vegas, the last 2 are improving in open cl and I understand that cs6 (which is out shortly-I am playing with photoshop cs6 atm) will also be much improved in open cl. I think the future is open cl but not sure how far in the future...
     
  4. m3k

    m3k Member

    Joined:
    Dec 29, 2007
    Messages:
    715
    save your money on gpu- get a 570.. and invest in cpu + ram

    Ive seen the new Murcury playback Cuda utilisation in cs6.. there are no real world enhancements from a gtx 260 to a gtx 570 unless your using 4k video ( but if your editing 4k video- there are better software than adobe )

    and with video editing i would be generally shocked/surprised if the addition of a workstation gpu will do anything amazing period.
    Im actually baffled as to why workstation gpus exist to be quite honest. perhaps its better with specialist software? (anyone wana clarify?)

    I don't know what cs.6 does different but i know from a developer standpoint- Gpu acceleration isn't easy to code for. so the benefits are very limited especially today.

    you have to realise software like these are enhanced to run smooth for mac n00bs with bad macbook pro's with budget graphic cards that don't even have driver updates. ( don't kill me applefanboys i own a mbp/ipad )

    Some CAD/3d software are using OPEN CL- And PHYS-X/Alternative physics, But they run just as fine between Nvidia and AMD- actually the alternative physics solvers are less buggy only trade off is they are slightly slower.

    Nvidia 680's has more cuda cores (1536) as opposed to 580's (512)

    I dont know what you mean by compute function restricted-
    especially for Adobe products that use cuda (for like 2 features)

    Does it have less performance for openCL/GL? could you speak about this more/show benchmarks, open CL is pretty interesting and ive been considering a 680.

    To be fair with current trends I wouldn't have troubles going AMD/ATI, their new cards are power efficient, they don't run hot- and they are well priced and do Open CL just fine.

    generally videocards dont do asmuch as they could for video editing.. I think your more limited by your ram/ hardrive and creativity haha
     
    Last edited: Apr 18, 2012
  5. richan3185

    richan3185 Member

    Joined:
    Jan 27, 2002
    Messages:
    1,831
    Location:
    Kalgoorlie
    My understanding is while more CUDA cores, each is a little less capable than the previous generation. However, the 680 is hamstrung severely in GPGPU.

    Additionally, the next release of the CUDA compiler will be hardware aware...hence, CUDA will function through a software translator on AMD hardware as well. It will take quite a while for the various companies to update their software using the newest compiler.

    I hate proprietary standards...i desperately hope OpenCL dominates.
     
  6. Luke212

    Luke212 Member

    Joined:
    Feb 26, 2003
    Messages:
    9,772
    Location:
    Sydney
    7970 is faster, but cuda has more support. so the answer is 'it depends'

    proprietary means better support thus better for business. open source can suck my dick for the amount of time wasted on poor implementations
     
  7. m3k

    m3k Member

    Joined:
    Dec 29, 2007
    Messages:
    715
    Open cl is dominating for alot of real time rendering applications, open source is great, apple is the best example why.
     
  8. Paronga

    Paronga Member

    Joined:
    Sep 27, 2006
    Messages:
    1,409
    Location:
    Morrabbin, Victoria
    there is also more to GPGPU then trying to get photoshop or video editing to work on them.

    Such as simulation work IE matlab, mathematica python and what not.
    These are things i try to take advantage of.

    while granted, GPUs don't really speed up things like a large matrix time a large matrix, some operations can be accelerated.
     
  9. saba

    saba Evil Vizier

    Joined:
    Aug 4, 2001
    Messages:
    2,710
    Location:
    Melbourne
    Rows and columns can be vectorised.
     
  10. Paronga

    Paronga Member

    Joined:
    Sep 27, 2006
    Messages:
    1,409
    Location:
    Morrabbin, Victoria
    well yes, but multiplication is the real magic on GPU.
    The thing with GPU is that it's far easier to stuff 4 GPUs in a PC then CPUs.
    So that generally why multiplication would be handed off.

    and i think that the precision of GPU operations is lower than CPUs?
    i may be wrong on that thought!
     
  11. richan3185

    richan3185 Member

    Joined:
    Jan 27, 2002
    Messages:
    1,831
    Location:
    Kalgoorlie
    I assume using the same bit precision will result in the same result. i.e. 32 v 64. However, since the CPU is fully 64bit "wired" it suffers little to no performance drop running 64bit mults, compared to the gpu which is wired 32bit and suffers extreme performance drop.

    AMD *dozer modules can run up to 128bit floats with minimal performance impact...granted there are other issues with the *dozer modules.

    Not 100% sure if this is correct...someone might be able to clarify.
     
  12. Paronga

    Paronga Member

    Joined:
    Sep 27, 2006
    Messages:
    1,409
    Location:
    Morrabbin, Victoria
    i also thought there were issues with large data sets.
    If i've got a 6gb complex matrix multiplied by another 6gb complex matrix i think the GPUs run into some main memory access issues?

    but really i'm not 100% sure on the details. I thought the CPU had better main memory access?
     
  13. richan3185

    richan3185 Member

    Joined:
    Jan 27, 2002
    Messages:
    1,831
    Location:
    Kalgoorlie
    Correct. The cpu has much quicker access to the system ram. The pcie cards must cross the bridge and northbridge for access. Hence why there is actual performance benefits from pcie3 for compute tasks.
     
  14. Paronga

    Paronga Member

    Joined:
    Sep 27, 2006
    Messages:
    1,409
    Location:
    Morrabbin, Victoria
    and you can cram more ram (12-64)gb in a computer than you can cram on a GFX card.

    so yea. GPGPU is always a balancing act!
     

Share This Page

Advertisement: