192xGT200 + 1068xBloomfield: 295 TFlops Supercomputer

Discussion in 'Beyond3D News' started by B3D News, Apr 23, 2008.

  1. B3D News

    B3D News Beyond3D News
    Regular

    Joined:
    May 18, 2007
    Messages:
    440
    Likes Received:
    1
    French website PC INpact broke the news of an upcoming GPU-accelerated supercomputer, ordered by France's CEA for delivery in early 2009 from Bull. The cluster's performance confirms that GT200 will be rated at 1TFlop and that Nehalem/Bloomfield will clock up to at least 3GHz.

    Read the full news item
     
  2. randomhack

    Newcomer

    Joined:
    Apr 4, 2008
    Messages:
    41
    Likes Received:
    0
    Those tflops quoted are single precision or double? Looks like single precision since the 24gflops figure quoted for Nehalem is based on SP figures. So the question still remains : how fast will the GT200 DP be?
     
  3. Arun

    Arun Unknown.
    Moderator Legend Veteran

    Joined:
    Aug 28, 2002
    Messages:
    5,023
    Likes Received:
    299
    Location:
    UK
    Yeah, it's SP. Presumably GT200 DP will be 1/4th speed for MUL and *maybe* 1/2th speed for ADD (so 1/6th for MADD); how that ties in with the 'missing MUL', I don't know. My expectation, FWIW, is that the SFU unit's MUL simply won't be exposed any more than in G80 - but I could be wrong.
     
  4. Megadrive1988

    Veteran

    Joined:
    May 30, 2002
    Messages:
    4,638
    Likes Received:
    148
    How do we compare this to G80?

    Is it like G80's estimated 1/2 TFLOP (Nvidia's programmable figure) or like its estimated more 'realistic' peak 1/3 TFLOP?

    Obviously one would hope GT200's 1 TFLOP of SP floating point performance is comparable to G80's 1/3 TFLOP, instead of the 1/2 TFLOP, thus making GT200 roughly 3x G80 instead of 2x.
     
    #4 Megadrive1988, Apr 24, 2008
    Last edited by a moderator: Apr 24, 2008
  5. CarstenS

    Veteran Subscriber

    Joined:
    May 31, 2002
    Messages:
    4,798
    Likes Received:
    2,056
    Location:
    Germany
    That'd be highly unusual in SC-space, wouldn't it?
     
  6. Arun

    Arun Unknown.
    Moderator Legend Veteran

    Joined:
    Aug 28, 2002
    Messages:
    5,023
    Likes Received:
    299
    Location:
    UK
    SP for the quoted numbers; that doesn't mean the chips don't support DP. As for your comment in the other thread, G80 doesn't support DP obviously but for GT200 it's rumoured to be 1/4th the performance of SP, presumably for MULs. Maybe ADDs will be 1/2th, which would bring MADDs to 1/6th. Also, for MULs, if you consider the hidden MUL and assume it's properly exposed in GT200 and nothing else changed (both of which I doubt), then that'd mean MUL is 1/6th and ADD is 1/2th. But once again, 1/4th would still be the most quoted figure I suspect.
     
  7. CarstenS

    Veteran Subscriber

    Joined:
    May 31, 2002
    Messages:
    4,798
    Likes Received:
    2,056
    Location:
    Germany
    I'm pretty sure, those TFLOPs mentioned are in fact DP ones - everything else would be totally unusual in the Supercomputing space. See also


    I'm still not sure what to make of all this. AFAIK AMD did promote only the 64 "fat ALUs" to be 64-bit capable - thus having one fifth of the SP-FLOPS available.

    As for GT200 I think everything from as low as 1/8th (as mentioned by Farhan) to as high as 50 percent (as described in this PDF (linked by Jawed some time ago) is possible.

    If it's indeed one eigth, and I think that's definitely possible given Nvidias track record of making a feature available first with almost no focus on speed, the DP-figure of 192 TFLOPS would nicely break down to 1 SP-TFLOPS per single GPU.

    edit:
    As for G80 - that was bad thinking on my part.
     
    #7 CarstenS, Apr 25, 2008
    Last edited by a moderator: Apr 25, 2008
  8. aaronspink

    Veteran

    Joined:
    Jun 20, 2003
    Messages:
    2,641
    Likes Received:
    64
    You are either way too optimistic OR way too trusting of PR blurbs if you think they are going to get 1 DP TFLOPs per card with GT200.

    That or you think that 400 watt graphics cards are going to be mainstream.

    Aaron Spink
    speaking for myself inc.
     
  9. randomhack

    Newcomer

    Joined:
    Apr 4, 2008
    Messages:
    41
    Likes Received:
    0
    Heres the bull press release :
    http://www.wcm.bull.com/internet/pr/rend.jsp?DocId=350329&lang=en

    They mention 1068 8-core nodes. That probably means 2 quad-core CPUs/nodes and therefore that is perfectly capable of giving 103 DP TFlops,

    Then they also mention 48 GPU nodes with 512 cores each. GPUs will provide 192 TFlops. So they are expecting 192/(48*4)=1 TFlops per 128 cores. Now the question is whether these tflops are DP or SP. Very confusing :(
     
  10. CarstenS

    Veteran Subscriber

    Joined:
    May 31, 2002
    Messages:
    4,798
    Likes Received:
    2,056
    Location:
    Germany
    No, there are 48*512 GPUs. Together they give 192 DP-TFLOPs, that'd be 128 GFLOPS (edit: DP) per GPU.
     
  11. randomhack

    Newcomer

    Joined:
    Apr 4, 2008
    Messages:
    41
    Likes Received:
    0
    :O
    Check your math :)
    Nobody puts over 24000 GPUs together either.
     
  12. CarstenS

    Veteran Subscriber

    Joined:
    May 31, 2002
    Messages:
    4,798
    Likes Received:
    2,056
    Location:
    Germany
    To the first one: Right *doh*
    But 24k processors - regardless if C- or GPUs - doesn't sound all that much in SC-space apart from my apparent inability to correctly use the windows calc. ;)
     
  13. CarstenS

    Veteran Subscriber

    Joined:
    May 31, 2002
    Messages:
    4,798
    Likes Received:
    2,056
    Location:
    Germany
    So, how about 4 TFLOPS per one of those 48 Units - being some kind of Tesla parts then? If there's really 512 'processors' in each of them, that'd be about 8 GFLOPS per 'processor' which i highly doubt in any way i can imagine right now.
     
Loading...

Share This Page

  • About Us

    Beyond3D has been around for over a decade and prides itself on being the best place on the web for in-depth, technically-driven discussion and analysis of 3D graphics hardware. If you love pixels and transistors, you've come to the right place!

    Beyond3D is proudly published by GPU Tools Ltd.
Loading...