You are viewing a single comment's thread from:

RE: Gridcoin - Computational performance chart (PFLOPS)

in #gridcoin • 6 years ago

Good read.

  • might be worth to separate GPU and CPU projects
  • would be interested to see team Gridcoin performance vs total project or BOINC performance
  • top three gpu-projects performance might give some false positives; what I mean is that FLOPS surprisingly is quite a muddy measurement unit and WU optimisation might have a great impact on recorded performance - if batches of WUs are less parallelized, they will be processed slower and network FLOPS will go down even in absence of hardware or participation changes
  • as per above, drawing conclusions about participation or network power changes just from FLOPS changes might lead to erroneous interpretations
6 years ago in #gridcoin by
$0.05
  • Past Payouts $0.05, 0.00 TRX
  • - Author $0.04, 0.00 TRX
  • - Curators $0.01, 0.00 TRX
3 votes
Reply 7
Sort:  
  • Trending
    • [-]
     6 years ago 

    Isn't it also true that Collatz only does integer operations? If so, using FLOPS for this project could be misleading.

    $0.07
    • Past Payouts $0.07, 0.00 TRX
    • - Author $0.05, 0.00 TRX
    • - Curators $0.01, 0.00 TRX
    4 votes
    Reply
    [-]
     6 years ago (edited)

    Good point, but...
    IOPS for Collatz, single precision FLOPS for Amicable, double precision FLOPS for MilkyWay... and we loose simple ONE number to several numbers, what is bad for general public. What is more, most programs use a mix of floats, integers etc. Also, at some point in Collatz computations integers might get promoted to floats, I don't know. Deeper you go, more complicated it becomes.

    $0.03
    • Past Payouts $0.03, 0.00 TRX
    • - Author $0.03, 0.00 TRX
    • - Curators $0.00, 0.00 TRX
    2 votes
    Reply
    [-]
     6 years ago (edited)

    I think both @h202 and yourself have valid points.

    It may not be perfect but at least we have a number we can use as a baseline to compare ourselves with eg. the TOP500 most powerful systems.

    [-]
     6 years ago 

    @hotbit @parejan

    I do think most "real-world" projects (i.e. modeling physical systems) use floats. (Unless you're talking about book-keeping operations like iteration variables and counters, which I suspect only make a small contribution.) Collatz, on the other hand, exclusively runs using (integer) mod and division operations. I don't see how floats could enter those computations. More generally, I suspect the projects which substantially mix integers + floats are in the minority.

    Maybe we could do an approximate conversion from IOPS to FLOPS, but not bother with single vs double precision? I do agree several numbers are bad for presentation, so definitely good to avoid that.

    $0.00
      Reply
      [-]
       6 years ago (edited)

      Thanks for your feedback @hotbit.

      • I will consider this and see how to integrate it in the State-of-the-Network updates.
      • In my weekly project updates you can monitor this in the chart "Team RAC vs Overall RAC %". I think this indicator is "fairly" accurate in terms of team Gridcoin performance vs total BOINC performance.
      • I had to think about your comments but if WU optimisations take place or if WUs can be less parallelized, doesn't this mean that the FLOPS will go up or down and the Gridcoin computational performance chart accordingly? I need to read you articles again. :-)
      • Personally see this chart as an estimation of computational performance and understand the limitations. On the other hand it does provide a trend if you look at it over a longer period.
      $0.03
      • Past Payouts $0.03, 0.00 TRX
      • - Author $0.03, 0.00 TRX
      • - Curators $0.01, 0.00 TRX
      2 votes
      Reply
      [-]
       6 years ago 

      Personally see this chart as an estimation of computational performance and understand the limitations. On the other hand it does provide a trend if you look at it over a longer period.

      Exactly. FLOPS might be the the best measuring tape we have on hand at the moment.
      There seems to be two forces influencing changes in the graph:

      • total hardware capabilities
      • parallelization / optimization of WUs in different batches (especially for GPU)

      Thanks for periodically putting together all these statistical data! Upvoting is the least I can do as a thank you.

      $0.04
      • Past Payouts $0.04, 0.00 TRX
      • - Author $0.03, 0.00 TRX
      • - Curators $0.01, 0.00 TRX
      2 votes
      Reply
      [-]
       6 years ago 

      Thanks @hotbit, much appreciated!

      $0.00
        Reply

        Coin Marketplace

        STEEM 0.28
        TRX 0.12
        JST 0.034
        BTC 63236.79
        ETH 3280.48
        USDT 1.00
        SBD 3.85