Cray MTA

By: Veedrac (ignore.delete@this.this.com), March 21, 2021 9:54 am
Room: Moderated Discussions
noone (blah.delete@this.blah.com) on March 21, 2021 9:15 am wrote:
> Veedrac (ignore.delete@this.this.com) on March 21, 2021 1:33 am wrote:
> > Chester (lamchester.delete@this.gmail.com) on March 20, 2021 7:54 pm wrote:
> > > anon (anon.delete@this.gmail.com) on March 20, 2021 6:04 pm wrote:
> > > > Veedrac (ignore.delete@this.this.com) on March 20, 2021 11:27 am wrote:
> > > > > But all of those are second order. You need to fix speculation. You cannot hope
> > > > > to have 30+ IPC unless you can speculate unboundedly, in a way that's immune to
> > > > > branch prediction, false memory hazards, loops, function calls, and so on.
> > > > >
> > > > > As far as I know, this means microthreads. So if I was to rewrite the world, I'd start by figuring
> > > > > out to build a core to handle a hundred plus microthreads with zero overhead, and work from
> > > > > there. I don't think there's anything fundamentally in the way of a core like this.
> > > >
> > > > https://en.wikipedia.org/wiki/Cray_MTA
> > > >
> > > > for a historical example of a many threaded core.
> > > >
> > > > Similarly, IBM Power chips often have 8-way SMT (for the same reason of trying to always have
> > > > work that you can make forward progress on). How would what you are suggesting differ from these?
> > >
> > > I think Veedrac wants to build a GPU. There you definitely have "a hundred plus microthreads" per
> > > core (where core is an Nvidia SM, AMD CU, or Intel subslice), with zero overhead. Each thread is
> > > generally independent of other threads, which means you can reorder execution unboundedly.
> >
> > You're both thinking ‘solve the execution problem’. I don't care about the execution problem
> > right now. The 128 core Ampere Altra Max is months away, and while GPUs presumably have room
> > to improve, they're already (as you say) way faster than anything I'm proposing.
> >
> > What I'm trying to solve is the *speculation* problem; the issue that two independently executable
> > instructions in your program might not be a mere 600 instructions apart, such that Apple's
> > ROB might be able to handle it, but 10,000 instructions, or 10,000,000, or maybe just a couple
> > of thousand but with multiple loops of unpredictable length between them.
> >
> > This is distinct from muticore and GPU approaches because
> > you might still just have a single pool of execution
> > units that each cycle some subset of the microthreads are dispatched to, though you have more flexibility
> > to not do that since sufficiently distant microthreads need
> > not be able to talk to each other with a zero-cycle
> > latency. And it's distinct from SMT because thread-to-thread communication is just register wakeup, albeit
> > shaped a tad differently, and you can spawn a thread on practically every branch.
> >
>
> The fundamental concept with this idea is memory dependencies
> between the these microthreads and how to deal with them.
>
> You should take a look at MIT's Swarm architecture, because I think it's exactly what you're talking about.
> The entire architecture constructs an enormous, distributed reorder buffer, built out of small hardware
> ROBs, with spawning and coordination controlled by software. And most importantly, a way of dealing with
> the memory dependencies between the spawned hardware threads. It's definitely not simple though.

Having just skimmed some of their stuff, I wouldn't say *exactly* what I'm talking about, but yeah, same general idea. There's nothing particularly new about the observation that microthreads—or tasks, or whatever—are a good idea. My key argument is just that, since they're the only approach I know of that could possibly work, that's where I'd start.

Memory is tricky, but I think it gets easier to figure out if you're assuming we're rewriting the world anyway, and also that (like any sane world rewrite) we're using capabilities. At least, easy enough that you should be able to make it no more burdensome than Rust from a software developer's perspective.
< Previous Post in ThreadNext Post in Thread >
TopicPosted ByDate
What are your ideas for a radically different CPU ISA + physical Arch?Moritz2021/03/20 04:21 AM
  What are your ideas for a radically different CPU ISA + physical Arch?Stanislav Shwartsman2021/03/20 05:22 AM
    I like the analysis of current arch presentedMoritz2021/03/20 09:13 AM
    Did you read this old article?Michael S2021/03/21 01:12 AM
  Deliver programs in IRHugo Décharnes2021/03/20 06:34 AM
    Java bytecode and Wasm exist, why invent something else? (NT)Foo_2021/03/20 07:01 AM
      Java bytecode and Wasm exist, why invent something else?Hugo Décharnes2021/03/20 07:55 AM
        Java bytecode and Wasm exist, why invent something else?Foo_2021/03/20 09:50 AM
          Java bytecode and Wasm exist, why invent something else?Hugo Décharnes2021/03/20 11:40 AM
            Java bytecode and Wasm exist, why invent something else?Foo_2021/03/20 03:54 PM
              It's called source code, no?anonymou52021/03/20 11:43 PM
                It's called source code, no?Foo_2021/03/21 04:07 AM
                Thoughts on software distribution formatsPaul A. Clayton2021/03/22 12:45 PM
    Deliver programs in IRJames2021/03/20 10:24 AM
      Deliver programs in IRHugo Décharnes2021/03/20 11:28 AM
        Deliver programs in IRHugo Décharnes2021/03/20 11:36 AM
    Deliver programs in IRLinus Torvalds2021/03/20 12:20 PM
      Deliver programs in IRHugo Décharnes2021/03/20 12:51 PM
      I'd like to be able to NOT specify order for some things ...Mark Roulo2021/03/20 04:49 PM
        I'd like to be able to NOT specify order for some things ...Jukka Larja2021/03/20 11:26 PM
          NOT (unintentionally) specify orderMoritz2021/03/21 05:00 AM
            NOT (unintentionally) specify orderJukka Larja2021/03/22 06:11 AM
              NOT (unintentionally) specify orderMoritz2021/03/22 11:40 AM
                NOT (unintentionally) specify orderJukka Larja2021/03/23 05:26 AM
          I'd like to be able to NOT specify order for some things ...Mark Roulo2021/03/21 08:47 AM
            I'd like to be able to NOT specify order for some things ...Victor Alander2021/03/21 04:14 PM
      Next architecture will start with MLwumpus2021/03/21 11:24 AM
        Next architecture will start with MLLinus Torvalds2021/03/21 01:38 PM
          Maybe SQL was the better example for general purpose machineswumpus2021/03/22 07:33 AM
            Maybe SQL was the better example for general purpose machinesanon2021/03/22 08:10 AM
        Next architecture will start with MLML will move to PIM2021/03/22 02:51 AM
    Deliver programs in IRanon2021/03/21 02:22 AM
      Deliver programs in IRanon22021/03/21 03:52 AM
        Deliver programs in IRrwessel2021/03/21 04:05 AM
          Deliver programs in IRanon22021/03/21 06:08 PM
            Deliver programs in IRrwessel2021/03/21 09:47 PM
              Deliver programs in IRdmcq2021/03/22 03:33 AM
                Deliver programs in IRrwessel2021/03/22 05:27 AM
  What are your ideas for a radically different CPU ISA + physical Arch?Veedrac2021/03/20 10:27 AM
    Cray MTAanon2021/03/20 05:04 PM
      Cray MTAChester2021/03/20 06:54 PM
        Cray MTAVeedrac2021/03/21 12:33 AM
          Cray MTAnoone2021/03/21 08:15 AM
            Cray MTAVeedrac2021/03/21 09:54 AM
    monolithic 3Dwumpus2021/03/21 11:50 AM
  What are your ideas for a radically different CPU ISA + physical Arch?Anon2021/03/20 11:06 PM
  What are your ideas for a radically different CPU ISA + physical Arch?rwessel2021/03/21 04:02 AM
  What are your ideas for a radically different CPU ISA + physical Arch?juanrga2021/03/21 04:46 AM
  Summery so farMoritz2021/03/21 08:45 AM
    Summery so farrwessel2021/03/21 10:23 AM
      not staticMoritz2021/03/26 09:12 AM
        Dynamic meta instruction encoding for instruction window compressionMoritz2021/03/28 02:28 AM
          redistributing the work between static compiler, dynamic compiler, CPUMoritz2021/04/05 02:21 AM
            redistributing the work between static compiler, dynamic compiler, CPUdmcq2021/04/05 08:27 AM
    Summery so farAnon2021/03/21 07:53 PM
  What are your ideas for a radically different CPU ISA + physical Arch?blaine2021/03/21 09:10 AM
    What are your ideas for a radically different CPU ISA + physical Arch?rwessel2021/03/21 10:26 AM
      What are your ideas for a radically different CPU ISA + physical Arch?rwessel2021/03/21 10:34 AM
        What are your ideas for a radically different CPU ISA + physical Arch?blaine2021/03/21 11:55 AM
          What are your ideas for a radically different CPU ISA + physical Arch?rwessel2021/03/21 12:31 PM
      What are your ideas for a radically different CPU ISA + physical Arch?gallier22021/03/21 11:49 PM
  What are your ideas for a radically different CPU ISA + physical Arch?dmcq2021/03/21 02:50 PM
  Microthread/low IPCEtienne Lorrain2021/03/22 02:22 AM
    Microthread/low IPCdmcq2021/03/22 03:24 AM
      Microthread/low IPCEtienne Lorrain2021/03/22 05:10 AM
        Microthread/low IPCdmcq2021/03/22 07:24 AM
    Microthread/low IPCdmcq2021/03/22 03:53 AM
      Microthread/low IPCEtienne Lorrain2021/03/22 04:46 AM
      Microthread/low IPCAnon2021/03/22 04:47 AM
    Microthread/low IPCHeikki Kultala2021/03/22 04:47 PM
      Microthread/low IPCEtienne Lorrain2021/03/23 02:36 AM
        Microthread/low IPCNyan2021/03/24 02:00 AM
          Microthread/low IPCEtienne Lorrain2021/03/24 03:23 AM
      Microthread/low IPCAnon2021/03/23 07:16 AM
        Microthread/low IPCgai2021/03/23 08:37 AM
          Microthread/low IPCAnon2021/03/23 09:17 AM
            Microthread/low IPCdmcq2021/03/23 11:42 AM
  Have you looked at "The Mill CPU" project? (nt)Anon C2021/03/22 05:21 AM
    Have you looked at "The Mill CPU" project? (nt)Moritz2021/03/22 11:13 AM
      Have you looked at "The Mill CPU" project? (nt)Andrew Clough2021/03/22 03:27 PM
        The Mill = vaporwareRichardC2021/03/23 11:47 AM
          The Mill = vaporwareMichael S2021/03/23 12:58 PM
          The Mill = vaporwareCarson2021/03/23 05:17 PM
          The Mill = doomed but interestingAndrew Clough2021/03/24 07:06 AM
            Solution in search of a problemwumpus2021/03/24 07:52 AM
              Solution in search of a problemdmcq2021/03/24 09:22 AM
          never-ware != vaporware (at least in connotation)Paul A. Clayton2021/03/24 09:37 AM
  What are your ideas for a radically different CPU ISA + physical Arch?anonini2021/03/22 07:28 AM
    microcode that can combine instructionMoritz2021/03/22 11:26 AM
  What are your ideas for a radically different CPU ISA + physical Arch?anony2021/03/22 09:16 AM
    Totally clueless.Heikki Kultala2021/03/22 04:53 PM
  Hierarchical instruction setHeikki Kultala2021/03/22 05:52 PM
    Hierarchical instruction setVeedrac2021/03/23 02:49 AM
      Hierarchical instruction setHeikki Kultala2021/03/23 05:46 AM
        Hierarchical instruction setEtienne Lorrain2021/03/23 06:16 AM
          microthreads on OS call/exceptionHeikki Kultala2021/03/23 06:34 AM
        Hierarchical instruction setVeedrac2021/03/23 08:31 AM
          Hierarchical instruction setEtienne Lorrain2021/03/24 12:13 AM
            Hierarchical instruction setVeedrac2021/03/24 06:11 AM
    Hierarchical instruction setAnon2021/03/23 07:39 AM
  What are your ideas for a radically different CPU ISA + physical Arch?Paul A. Clayton2021/03/26 07:21 AM
    What are your ideas for a radically different CPU ISA + physical Arch?wumpus2021/03/26 08:45 AM
Reply to this Topic
Name:
Email:
Topic:
Body: No Text
How do you spell avocado?