Consolidated Business & Enterprise Computing Rant Thread

Discussion in 'Business & Enterprise Computing' started by elvis, Jul 1, 2008.

  1. Luke212

    Luke212 Member

    Joined:
    Feb 26, 2003
    Messages:
    10,224
    Location:
    Sydney
    exactly. commercialising something is a long and tedious process.

    the technical part is the easy part.
     
  2. GumbyNoTalent

    GumbyNoTalent Member

    Joined:
    Jan 8, 2003
    Messages:
    10,230
    Location:
    Briz Vegas
    Shirley your genius knows no bounds?
     
  3. OP
    OP
    elvis

    elvis Old school old fool

    Joined:
    Jun 27, 2001
    Messages:
    45,349
    Location:
    Brisbane
    Yes, you're right, they're incredibly lazy in that they're trying to process things 25GB/s+ RAM speeds instead of 1/10th that speed reading from disk, thus taking 10 times the wall time to process what already takes months across thousands of processors.

    Given that we're spending hundreds of millions per year on compute time, if you could reduce that through your expert programming super powers, we'd pay you millions for a single job alone. Want to be an overnight millionaire? Show us your code.
     
  4. caspian

    caspian Member

    Joined:
    Mar 11, 2002
    Messages:
    12,047
    Location:
    Melbourne
    [​IMG]
     
    power likes this.
  5. OP
    OP
    elvis

    elvis Old school old fool

    Joined:
    Jun 27, 2001
    Messages:
    45,349
    Location:
    Brisbane
    Hi there everyone, me again, the guy spouting on about the death of x86 in high end compute, coming soon to a theatre near you.

    Some reading this morning. Intel are hurting bad, profits and sales down:
    https://www.nextplatform.com/2021/04/23/the-intel-datacenter-party-starts-feeling-the-hangover/

    And Jim Keller, the galaxy brain who co-designed the Apple M1 (along with the AMD K7 and K8 back in the day, and co-authored x86_86 and Hyper-Transport), is now CTO of a new mob designing custom silicon for AI/ML in large scale compute:
    https://www.hpcwire.com/2021/04/22/jim-keller-cto-tenstorrent-2021-hpcwire-person-to-watch/
    https://www.tenstorrent.com/technology/

    * x86 yawn *
     
  6. Cape_Horn

    Cape_Horn Member

    Joined:
    Dec 23, 2001
    Messages:
    2,485
    Location:
    Shooting Baker
    hmm...
    I wonder if there is anyway to early invest in this company...
     
    mooboyj likes this.
  7. NSanity

    NSanity Member

    Joined:
    Mar 11, 2002
    Messages:
    18,428
    Location:
    Brisbane
    Does this coincide with the year of Linux on the Desktop? Or Arm dominating in the Datacenter?

    High-end Compute as in HPC will still be dominated by x86 in terms of new deployments for a long time, simply because its more flexible. Having niche platforms for niche solutions will totally still be a thing - but HPC clusters are general purpose things for *years*. Having a platform that requires everyone to radically refactor their code to use the platform isn't very accessible/optimal for a shared platform is unwise.

    Keller's stuff is *super* interesting, and no doubt will utterly dominant for its use case as seemingly everything he gets involved with skyrockets.


    Intel's current woes are simply executive level nonsense that has core issues stemming from business people running a semiconductor company, rather than engineers. Gelsinger will right that ship - and Intel has the pockets to ride the storm.
     
    Last edited: Apr 28, 2021
  8. OP
    OP
    elvis

    elvis Old school old fool

    Joined:
    Jun 27, 2001
    Messages:
    45,349
    Location:
    Brisbane
    Desktop is dead (I spend my life rolling out web based interactive visualization for stuff that matters - emails and spreadsheets crew are last century's news), but yes, ARM in the data centre as OS babysitter to fun AI/ML GPUs and IPUs doing the real work.

    I arbitrarily picked 2030 above as my line in the sand. Because I'd go for 2025, and I've learned to add 5 years to everything when it comes to waiting for corporate boneheads to catch up on what's really a good idea.

    "Flexible" isn't worth a damn when a chip at 2x the price can do 50x the workload. In HPC, that's just being stupid to not jump on board, especially when you're in an arms race to come up with the next vaccine/cure/whatever that the world desperately needs.

    Join me in the steak dinner bet!
     
    Last edited: Apr 28, 2021
  9. bcann

    bcann Member

    Joined:
    Feb 26, 2006
    Messages:
    6,118
    Location:
    NSW
    at 50x the workload, if they threw in some kind of x86/x64 emulation mode, even if it lost 25% performance, then they would crush it.
     
  10. NSanity

    NSanity Member

    Joined:
    Mar 11, 2002
    Messages:
    18,428
    Location:
    Brisbane
    If this was truly the case, we'd not have so much java and other rubbish frameworks running the world.

    Hardware is cheap. Talented developers are rocking horse shit.
     
  11. OP
    OP
    elvis

    elvis Old school old fool

    Joined:
    Jun 27, 2001
    Messages:
    45,349
    Location:
    Brisbane
    Accurate for enterprise shitware. Less accurate for people potentially making trillions from the next super drug.

    Very different worlds. In my world, very few people are "talented developers", but they can all use Tensorflow on GPUs because those "rocking horse shit" developers made it a trivial library/API call.

    Open source, royalty free, complex tools like these are rapidly becoming the norm in science.
     
    Last edited: Apr 28, 2021
  12. NSanity

    NSanity Member

    Joined:
    Mar 11, 2002
    Messages:
    18,428
    Location:
    Brisbane
    Yeah Tensorflow etc have made a huge difference to adoption of GPGPU compute - agreed.

    still, not everything fits in a GPU bucket. Hence our boy Keller's gig.
     
  13. OP
    OP
    elvis

    elvis Old school old fool

    Joined:
    Jun 27, 2001
    Messages:
    45,349
    Location:
    Brisbane
    And that's the point. Keller is working on Gen 2 of something that's currently Gen 1 real silicon that you can buy today. You can bet your sweet arse with his leadership it'll do something pretty amazing from here on, and if they can wrap a Tensorflow-esque bit of software around it so that you don't need genius programmers to us it (hell, even if that means you can only "use" 90% of it), then that's another nail in the "x86 generalist" coffin.

    GPU never was the end goal. Silicon diversity was. General purpose everything at best gives you average results everywhere. Specialisation has pros and cons, but in the high end compute game, the literal arms race of getting to the answer before your competitors is everything.

    I maintain a very tiny (by comparison to the big boys) cluster right now where GPU is in high demand, but it's not installed on every node. We're looking at other things too - FPGAs and IPUs and all sorts of gizmos, to scatter around the cluster. Year on year the x86 CPUs on board do less and less, and year on year people write bits of code to take advantage of specialised silicon with convenient APIs and libraries wrapped around them to do the specific things they need done. And when it takes the wall time from months to weeks, then it's worth every bit of investment.

    Time is the biggest killer of ideas for these folk. Anything that takes too long is forgotten, funding removed, and projects die. You just can't afford to write general purpose code and wait months to see an answer when the people funding it want an answer in half that time, and your competition is already working on it.

    Again, pros and cons. x86 makes life easy when you've got all the time in the world. But not everyone has all the time in the world, and the "H" in "HPC" is the key point in that discussion, and precisely where my predictions around the death of x86 start and finish. General purpose enterprise shitware? Sure, that'll be x86 for another several human lifetimes, but I also give very few shits about that level of banality. Digital paper pushing is mooks' work, and not at all interesting to me.
     
  14. millsy

    millsy Member

    Joined:
    Mar 31, 2007
    Messages:
    13,588
    Location:
    Brisbane
    Just let me know when I don't have to buy a damn mac to get an m1 style processor that's not shit plz
     
  15. OP
    OP
    elvis

    elvis Old school old fool

    Joined:
    Jun 27, 2001
    Messages:
    45,349
    Location:
    Brisbane
    On the desktop? Sorry man, that's the least interesting place to play with cool tech.

    HPC and edge compute is where all the fun hardware toys are for at least the next decade. What email and Excel slingers do isn't pushing any new boundaries.
     
  16. millsy

    millsy Member

    Joined:
    Mar 31, 2007
    Messages:
    13,588
    Location:
    Brisbane
    > using excel
    > not pushing boundaries

    Buddy they had to make a 64bit one for a reason, many boundaries pushed. You're just mad you don't need a supercomputer to do actual work in excel like the rest of the world

    /s
     
    elvis likes this.
  17. EvilGenius

    EvilGenius Member

    Joined:
    Apr 26, 2005
    Messages:
    10,866
    Location:
    elsewhere
    Clearly not embedding enough word docs. Amateur.
     
  18. looktall

    looktall Working Class Doughnut

    Joined:
    Sep 17, 2001
    Messages:
    26,668
    I save my excels into an email and then add that as an attachment while replying to people by leaving the body of my email blank and editing the email trail with my answers to their questions.
     
  19. GumbyNoTalent

    GumbyNoTalent Member

    Joined:
    Jan 8, 2003
    Messages:
    10,230
    Location:
    Briz Vegas
    oldguy_qld likes this.
  20. Luke212

    Luke212 Member

    Joined:
    Feb 26, 2003
    Messages:
    10,224
    Location:
    Sydney
    another topic i called a decade ago.... nbn flat pricing for RSPs. heretical thing to say for many years.

    Slowly people are coming around to it... lets watch and see when the penny finally drops. (another 5 years?)

    there is 0 reason not to have flat cost for availability. nbn is an energy distributer not energy producer. and on top of that it has fibre, which makes it like an energy distribution company that has room-temperature super conductors. its trivial to supply any capacity.

    even IT people dont seem get this. if you still dont get it, thats on you.
     
    Last edited: May 3, 2021

Share This Page

Advertisement: