3 ms·

The problem is that these advancements are one-offs and are increasingly difficult to both create and adopt. There was a great paper I read on compiler optimiz

The problem is that these advancements are one-offs and are increasingly difficult to both create and adopt. There was a great paper I read on compiler optimizations (I was a language guy in college) that pointed out that compiler optimizations were responsible for maybe a 2x improvement in program speeds, whereas CPU improvements were responsible for 1000x. Almost all the singularity-type techno-optimism was based on a curve drawn during the amazingly magical part of CPU development: shrink the die, double the speed. We are now past that, although many techno-optimists are casting around for a reason it will continue.

Guvante 13y ago
Honestly if we can get multi-core programming working we can probably keep it going. Although that is an enormous if.
- carsongross 13y ago
  Most problems are not trivially parallelizable and, in any event, Amdahl curves look a lot different than the exponential curves used to argue for The Singularity. We may have to make due with what we've got. Which, thankfully, is quite a lot.
  - marcosdumay 13y ago
    Well, our brain has a single thread performance that is already many orders of magnitude smaller than our current CPUs. Any claim that Amdahl's law will forbid human-like inteligence must explain why that isn't a problem to us. Or, like you said, yep, we do have a lot.
  - im3w1l 13y ago
    >Most problems are not trivially parallelizable Numeric optimization can benefit from parallel line search, and threads using different initial points. Matrix algebra benefits from more cores. Sorting benefits from more cores. Encryption benefits from more cores (De)compression benefits from more cores. A server handling simultaneous requests benefits from more cores. Cracking hashes benefits from more cores Computing merkle trees benefits from more cores Flash banners in 20 tabs "benefits" from more cores
tim333 13y ago
One of the interesting things in the Kurzweil type techno-optimism stuff is that the calculations/sec you get for $1000 seems to have been doubling regularly since 1900 or so long predating Moore's law and the phenomena seems largely independent on which particular technology is used to implement it. (see https://en.wikipedia.org/wiki/File:PPTMooresLawai.jpg https://en.wikipedia.org/wiki/File:PPTMooresLawai.jpg). I figure it must be largely an economic phenomenon in that it's worth while for the manufacturers to invest to be able to say 'you should buy next year's model - it's 50% faster than this years.' These days the technical implementation seems to be more cores - there seems no absolute limit on how many parallel processors you can have.

Guvante 13y ago

Honestly if we can get multi-core programming working we can probably keep it going. Although that is an enormous if.

carsongross 13y ago
Most problems are not trivially parallelizable and, in any event, Amdahl curves look a lot different than the exponential curves used to argue for The Singularity. We may have to make due with what we've got. Which, thankfully, is quite a lot.
- marcosdumay 13y ago
  Well, our brain has a single thread performance that is already many orders of magnitude smaller than our current CPUs. Any claim that Amdahl's law will forbid human-like inteligence must explain why that isn't a problem to us. Or, like you said, yep, we do have a lot.
- im3w1l 13y ago
  >Most problems are not trivially parallelizable Numeric optimization can benefit from parallel line search, and threads using different initial points. Matrix algebra benefits from more cores. Sorting benefits from more cores. Encryption benefits from more cores (De)compression benefits from more cores. A server handling simultaneous requests benefits from more cores. Cracking hashes benefits from more cores Computing merkle trees benefits from more cores Flash banners in 20 tabs "benefits" from more cores

tim333 13y ago

One of the interesting things in the Kurzweil type techno-optimism stuff is that the calculations/sec you get for $1000 seems to have been doubling regularly since 1900 or so long predating Moore's law and the phenomena seems largely independent on which particular technology is used to implement it. (see https://en.wikipedia.org/wiki/File:PPTMooresLawai.jpg https://en.wikipedia.org/wiki/File:PPTMooresLawai.jpg). I figure it must be largely an economic phenomenon in that it's worth while for the manufacturers to invest to be able to say 'you should buy next year's model - it's 50% faster than this years.' These days the technical implementation seems to be more cores - there seems no absolute limit on how many parallel processors you can have.