[HN Gopher] Google's Tensor Processing Units could have been dev...
       ___________________________________________________________________
        
       Google's Tensor Processing Units could have been developed using
       stolen designs
        
       Author : wslh
       Score  : 43 points
       Date   : 2024-01-12 20:49 UTC (2 hours ago)
        
 (HTM) web link (www.theregister.com)
 (TXT) w3m dump (www.theregister.com)
        
       | infotainment wrote:
       | _According to Singular 's website, the biz "develops and licenses
       | hardware and software technologies for high-performance energy-
       | efficient computing, both large scale and embedded."_
       | 
       | Sounds like a patent troll.
        
         | jacobgorm wrote:
         | Please have some respect for the little guy trying to make a
         | living inventing things. We can't all be corporate drones, even
         | if it pays well.
        
       | light_hue_1 wrote:
       | Just look at the patents that are being claimed.
       | https://patents.google.com/patent/US9218156B2/en
       | https://patents.google.com/patent/US8407273B2/en
       | 
       | This is embarrassing. It's just patenting the idea that you would
       | have low precision hardware for math operations and then extend
       | the results back out so the program doesn't know.
       | 
       | A patent troll at its worst. I've never heard of this guy (and
       | I'm an ML researcher) but the idea that you would give up a bit
       | of precision for performance during training (and particularly
       | during inference!) has been around forever.
       | 
       | The amount of prior art for this idea is.. obscene. The patent
       | examiners did not do their jobs at all.
        
         | tdullien wrote:
         | US Patent examiners have for decades been very much of the
         | position to not worry much about existing prior art when
         | granting patents, relying instead on the idea that people will
         | challenge and overturn patents when necessary (this applies to
         | the non-obviousness bar, too).
         | 
         | This, coupled with a tendency for Patent examiners to have a
         | subsequent well-paid career drafting patents, has lead to
         | thousands of sketchy patents on the books that have high odds
         | of being overturned, but only if a sufficiently resourced party
         | challenges them.
         | 
         | We shall see how these patents hold up...
        
           | pan69 wrote:
           | It's a completely broken system (as we all already know). Why
           | have "examiners" at all? The way it should probably work is;
           | if a patent is overturned in court, then the patent office
           | should bare the legal costs.
        
         | wslh wrote:
         | I remember, before GPUs, that using 32 (i386) bits instead of
         | the FPU for fast 3D operations was common.
        
         | jacobgorm wrote:
         | The guy does not claim to be an ML researcher, looks more like
         | a Computer Architecture person, so the fact that you never
         | heard of him does not make him a nobody or mean anything
         | really.
         | 
         | No ideas have been around "forever", things that may seem
         | obvious to you now often cost other people years of research to
         | come up with. The better the idea, the more obvious it
         | typically appears in hindsight.
         | 
         | If Jeff Dean was aware of the work as the Register article
         | states, it probably was significant, and hopefully Google will
         | pay the inventor for his ideas.
        
       ___________________________________________________________________
       (page generated 2024-01-12 23:01 UTC)