<html>
  <head>
    <meta http-equiv="Content-Type" content="text/html; charset=UTF-8">
  </head>
  <body>
    <p><br>
    </p>
    <div class="moz-cite-prefix">Marcus wrote:<br>
    </div>
    <blockquote type="cite"
      cite="mid:F6E259DD-038A-4C85-AC65-8CC4BA0BD377@snoutfarm.com">
      <meta http-equiv="Content-Type" content="text/html; charset=UTF-8">
      Cerebras says they can scale past 1 trillion parameters..  
      <div><br>
      </div>
      <div>
        <div style="display: block;" class="">
          <div style="-webkit-user-select: all; -webkit-user-drag:
            element; display: inline-block;" class="apple-rich-link"
            draggable="true" role="link"
data-url="https://www.hpcwire.com/2021/08/24/wafer-scale-to-brain-scale-cerebras-touts-linear-scaling-for-up-to-192-cs-2-systems/"><a
              style="border-radius:10px;font-family:-apple-system,
              Helvetica, Arial,
sans-serif;display:block;-webkit-user-select:none;width:300px;user-select:none;-webkit-user-modify:read-only;user-modify:read-only;overflow:hidden;text-decoration:none;"
              class="lp-rich-link" rel="nofollow"
href="https://www.hpcwire.com/2021/08/24/wafer-scale-to-brain-scale-cerebras-touts-linear-scaling-for-up-to-192-cs-2-systems/"
              dir="ltr" role="button" draggable="false" width="300"
              moz-do-not-send="true">
              <table
style="table-layout:fixed;border-collapse:collapse;width:300px;background-color:#E9E9EB;font-family:-apple-system,
                Helvetica, Arial, sans-serif;"
                class="lp-rich-link-emailBaseTable" width="300"
                cellspacing="0" cellpadding="0" border="0">
                <tbody>
                  <tr>
                    <td vertical-align="center" align="center"><img
                        style="width:300px;filter:brightness(0.97);height:139px;"
                        draggable="false"
                        class="lp-rich-link-mediaImage"
                        alt="Cerebras-MemoryX-SwarmX-CS-2_banner.png"
                        src="cid:part1.ZWqt9IRo.Aw0DrPxl@swcp.com"
                        width="300" height="139"></td>
                  </tr>
                  <tr>
                    <td vertical-align="center">
                      <table style="font-family:-apple-system,
                        Helvetica, Arial,
                        sans-serif;table-layout:fixed;background-color:rgba(233,
                        233, 235, 1);" class="lp-rich-link-captionBar"
                        width="300" cellspacing="0" cellpadding="0"
                        bgcolor="#E9E9EB">
                        <tbody>
                          <tr>
                            <td style="padding:8px 0px 8px 0px;"
                              class="lp-rich-link-captionBar-textStackItem">
                              <div style="max-width:100%;margin:0px 16px
                                0px 16px;overflow:hidden;"
                                class="lp-rich-link-captionBar-textStack">
                                <div
style="word-wrap:break-word;font-weight:500;font-size:12px;overflow:hidden;text-overflow:ellipsis;text-align:left;"
class="lp-rich-link-captionBar-textStack-topCaption-leading">
                                  <a rel="nofollow"
href="https://www.hpcwire.com/2021/08/24/wafer-scale-to-brain-scale-cerebras-touts-linear-scaling-for-up-to-192-cs-2-systems/"
                                    style="text-decoration: none"
                                    draggable="false"
                                    moz-do-not-send="true"><font
                                      style="color: rgba(0, 0, 0, 1);"
                                      color="#000000">Wafer Scale to
                                      ‘Brain-Scale’ – Cerebras Touts
                                      Linear Scaling up to 192 CS-2
                                      Systems</font></a></div>
                                <div
style="word-wrap:break-word;font-weight:400;font-size:11px;overflow:hidden;text-overflow:ellipsis;text-align:left;"
class="lp-rich-link-captionBar-textStack-bottomCaption-leading">
                                  <a rel="nofollow"
href="https://www.hpcwire.com/2021/08/24/wafer-scale-to-brain-scale-cerebras-touts-linear-scaling-for-up-to-192-cs-2-systems/"
                                    style="text-decoration: none"
                                    draggable="false"
                                    moz-do-not-send="true"><font
                                      style="color: rgba(60, 60, 67,
                                      0.6);" color="#A2A2A9">hpcwire.com</font></a></div>
                              </div>
                            </td>
                          </tr>
                        </tbody>
                      </table>
                    </td>
                  </tr>
                </tbody>
              </table>
            </a></div>
        </div>
        <div><br>
        </div>
        That would be a power budget of a HPC center, but not out of the
        ordinary.  Less than 10 MW.   AWS, Azure, Google and national
        labs have facilities like that.</div>
    </blockquote>
    This post/response has a dual-purpose... one is to respond/question
    specifics about this modern (special purpose) cluster-architecture
    and the other is to *allude to* Glen's point about
    topology/conformation of the "networks" (in this case the
    computational fabric) vs simple brute-force scaling (e.g. count of
    processing elements vs neurons/axons vs microbiome flux, etc.)<br>
    <p>Thanks for this reference... I haven't been (even a poser as) an
      HPC wonk for about 2 decades but was surprised when I <a
        moz-do-not-send="true"
href="https://f.hubspotusercontent30.net/hubfs/8968533/Virtual%20Booth%20Docs/CS%20Weight%20Streaming%20White%20Paper%20111521.pdf">drilled
        down on the "SwarmX Fabric"</a> Cerebras
      quotes/advertises/invokes and was unable to find many details on
      what they mean (might just be my rustyness in
      cluster-interconnect-fabric terminology, or my limited
      focus/attention-span?)  by "weight streaming" and quite what a
      "bidirectional tree topology" is... I can parse the words, but am
      having a problem finding (understanding?) the specifics.</p>
    <p>I am taking terminology about "weight sparsity" to be somewhat
      specific to Neural Net training/models but also maybe not that far
      from the more general problems encountered in sparse matrix
      solvers?    In 2001-2003 I worked with the <a
        moz-do-not-send="true"
href="https://www.lanl.gov/conferences/salishan/salishan2003/morrison.pdf">ASCI
        Q-machine</a> project on a *simulator* for the machine before it
      was delivered/constructed/debugged/testable, and became modestly
      familiar with it's topology and the problems presented that *seem*
      (superficially) parallel with those presented in this Cerebras
      system.   The Quadrics(tm) 4x4 high-speed interconnects used to
      build a hierarchical(ish) (<a moz-do-not-send="true"
href="https://www.researchgate.net/publication/260585525_Expandable_and_Cost-Effective_Network_Structures_for_Data_Centers_Using_Dual-Port_Servers/figures?lo=1">Fat
        H Tree</a>) switching fabric (overlapping quad-tree) seems to be
      similar in concept if not in spec/implementation to what Cerebras
      is building.   The Q-machine was perhaps designed for more general
      purpose problems, but the flagship (as I understood it) problem of
      the day WAS huge sparse matrix solvers, albeit to address
      Computation Fluid Dynamics and Radiation Transport (rather than
      LL) Models.  GPU computational fabric as (becoming) a new thing at
      the time,  it is amazing what the special purpose *tensor
      processors* of today seem to be capable of<br>
    </p>
    <p>There are probably some HPC wonks here who *do* know this off the
      top of their heads, but I was fascinated (as I drilled down) to
      discover how (relatively) similar the architectures and the
      problems seem (to me).  It is likely that I am just the hammer
      seeing everything as a nail, of course.</p>
    <p>Our "garish" collection of visual representations of the Q
      machine switching fabric (with simulated traffic on a simulated
      problem), much better experienced en-virtu, of course!   Ed will
      probably remember seeing this "back in the day"...<br>
    </p>
    <div align="center"><img moz-do-not-send="true"
src="https://www.researchgate.net/profile/Kei-Davis/publication/220586586/figure/fig1/AS:276976649687057@1443047788013/Representation-QS-Quaternary-fat-tree-network-with-64-computational-nodes-small-circles_W640.jpg"
        alt="" width="801" height="267"></div>
    <div align="center"><a
href="https://www.researchgate.net/publication/220586586_Graph_Visualization_for_the_Analysis_of_the_Structure_and_Dynamics_of_Extreme-Scale_Supercomputers/figures?lo=1"
        class="moz-txt-link-freetext">https://www.researchgate.net/publication/220586586_Graph_Visualization_for_the_Analysis_of_the_Structure_and_Dynamics_of_Extreme-Scale_Supercomputers/figures?lo=1</a></div>
    <div align="center"><br>
    </div>
    <p><a moz-do-not-send="true"
href="https://www.lanl.gov/conferences/salishan/salishan2003/morrison.pdf"
        class="moz-txt-link-freetext">https://www.lanl.gov/conferences/salishan/salishan2003/morrison.pdf</a></p>
    <p><a
href="https://f.hubspotusercontent30.net/hubfs/8968533/Virtual%20Booth%20Docs/CS%20Weight%20Streaming%20White%20Paper%20111521.pdf">https://f.hubspotusercontent30.net/hubfs/8968533/Virtual%20Booth%20Docs/CS%20Weight%20Streaming%20White%20Paper%20111521.pdf  
      </a></p>
    <p>nod to Glen... I might probably be able to install/load/try the
      openAI model if I wasn't wasting so much time careening down
      memory lane and trying to register what I see in my rear view
      mirrors with what I see screaming down the Autobhan-of-the-mind
      through my windscreen!</p>
    <p>Or maybe it is chatGPT/GPT3 actually *writing* this post for
      me?   Or have *I* become "one" with my AI-Overlord "dual" who is
      collaborating with me on this?   What is the distinction between
      parasitic and symbiotic?  <br>
    </p>
    <p>Who will ever know...  <br>
    </p>
    <p><br>
    </p>
  </body>
</html>