1. This site uses cookies. By continuing to use this site, you are agreeing to our use of cookies. Learn More.

News Google unveils Tensor Processing Unit hardware

Discussion in 'Article Discussion' started by Gareth Halfacree, 19 May 2016.

  1. Gareth Halfacree

    Gareth Halfacree WIIGII! Staff Administrator Super Moderator Moderator

    Joined:
    4 Dec 2007
    Posts:
    13,054
    Likes Received:
    2,152
  2. schmidtbag

    schmidtbag New Member

    Joined:
    30 Jul 2010
    Posts:
    1,082
    Likes Received:
    10
    Hmm this is actually really interesting. I never thought about making a processor intentionally "less accurate" but when you're, in a way, trying to replicate the human brain, constant accuracy isn't a necessity. It does get me to wonder, though, what would happen if it makes a mistake in its learning, and ends up remembering that mistake for good. Again - not so different from humans, but depending on exactly what they intend to use it for, this extra speed might not be worth it.
     
  3. Cheapskate

    Cheapskate Insane? or just stupid?

    Joined:
    13 May 2007
    Posts:
    10,626
    Likes Received:
    760
    "Less accurate"
    "Kill the humans"
     
  4. adam_bagpuss

    adam_bagpuss Have you tried turning it off/on ?

    Joined:
    24 Apr 2009
    Posts:
    4,228
    Likes Received:
    148
    isnt this the chip from terminator !

    arghhh SKYNET - we are all doomed
     
  5. dunx

    dunx ITX is where it's at !

    Joined:
    1 Sep 2010
    Posts:
    463
    Likes Received:
    13
    Less accurate = RISC ! A very old concept, and programming an ASIC isn't impossible to do in three weeks.

    Especially given their resources and funding....

    dunx
     
  6. schmidtbag

    schmidtbag New Member

    Joined:
    30 Jul 2010
    Posts:
    1,082
    Likes Received:
    10
    I think there's a bit of confusion here. RISC is known for being less precise, but just about every commercial processor in existence will give answers of an equal level of accuracy (which is 100%).

    Think of it like this: The answer to 1.1+2.2=3.3. Whether you have an x86, a MIPS, or an ASIC, all 3 of those will tell you that answer every single time. But Google's processor might make a mistake and say the answer 3.2 or 3.4. That's a lack of accuracy, because not only is it wrong, but the answer isn't consistent. If it were a lack of precision but accuracy was ok, then it would say 1.1+2.2=3 every single time.

    Does that make sense?
     
  7. Wwhat

    Wwhat Member

    Joined:
    2 Oct 2005
    Posts:
    263
    Likes Received:
    1
    Meanwhile if you try to translate a western language in google translate, say English to German (to make it easy), it's all a pitiful disaster, and not improving for years and years and years. And let's not even mention asian languages and such, I think we all gave up on that one and accepted that that isn't a thing.

    And if you are a user that does not something google thinks you want you are forced to suffer it, and nothing can make all their AI and algorithms clear that maybe amongst the 7+ billion people there might be some who do not want what they have decided everybody wants and needs.

    And I could go on and on.
     

Share This Page