back to article Undergrads build 12.6-TFLOPS cluster out of four nodes, 112 cores

The ISC (International Supercomputing Conference) Student Cluster Competition once again cements its reputation as the place where LINPACK records go to fall. Last week we saw not one, but two teams top the current 12.03 TFLOPS record that was established at the ASC'16 spring competition. Team HUST, from the Huazhong …

  1. Tom 7

    Not moores law?

    I couldn't find any rules on the competition sight (didnt look too hard) but I would have expected a somewhat different graph.

    Not trying to belittle anyone's work - just curious.

  2. zvonr

    hmmm

    From what I can tell Team Hamburg's system is better than No1 and No3 system by only using 2 nodes... so i would think they win @ TFLOP/$ by a significant margin...(which is what should matter)

    1. Gotno iShit Wantno iShit

      Re: hmmm

      The competition is power limited at 2x 1560W. I guess it uses that limit rather than $ as one of the things they seek to encourage is vendor collaboration. Trouble with using pure $ as a limit would be scale, some really cheap implementations could return huge bang/buck but not be capable of scaling to be useful at real world problems.

      I agree though, the cheaper approach ought to win, the completion needs weighting to account for both Watts and $.

  3. jms222

    Moore's law is about process geometry not how many cores you can gaffer tape together which is purely about money, And gaffer tape.

    1. This post has been deleted by its author

  4. druck Silver badge

    Mostly GPU

    A naive look at the top two results suggests that most of the work was being done by the 8 K80 GPUs, and the extra 56 CPUs that the 1st place team used, didn't bring much to the party.

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Other stories you might like