Saturday, June 11, 2005
Massively Parallel
Well, the recent news that Apple is going to switch to Intel processors is certainly interesting. After doubting the wisdom of the move at first, I am now quite excited about it. In general it's good to keep relationships nice and fluid as it helps to drive competition and progress.
In particular however, there are all sorts of very exciting technologies being developed, which is renewing my hope that we will have cheap 100 Teraflop computers by 2020 - perhaps a good bit sooner! 100 Teraflops is an exciting number: Carngie Mellon computer vision researcher Hans Moravec estimates that the human brain processes information at an effective rate of about 100 Tflops. Hans arrives at the figure by estimating that the neurons in the retina make one billion calculations per second, and then scales this up by a factor of 100,000 for the volume of the whole brain. It took approximately one billion years for 100 trillion atom cells to evolve, and another billion for them to evolve into 100 trillion cell humans with our 100 teraflop brains. We are now going to reproduce this feat in about 100 years - some 10 million times faster. In general the first step is now complete - single processor speeds probably can't be pushed orders of magnitude faster in GHz. The next step is to arrange large numbers of these processors in order to make massively parallel machines.
We are already heading down this path. Dual core computers are available now, and AMD's quad core chips are on the way in 2007. By 2015 we will have computers with 100's of processor cores, if Intel's researchers have their way. There are other fascinating advances as well. In order to feed 100's of fast processors you need very high bandwidth, something that recent advances in getting silicon lasers to work will likely help with. Furthermore the processors will need a lot of data to chew through, which new forms of solid state memory should provide. For instance Ovonyx is working on a promising new type of optical memory that switches between crystalline and amorphous states when prodded electrically.
*Oops, violating causality a little bit here: I wrote most of this post in early June a little while after the Apple announcement, but then got absorbed by the Arnowitt Deser and Misner 3+1 splitting of spacetime, and am now finishing this post on July 1st.
In particular however, there are all sorts of very exciting technologies being developed, which is renewing my hope that we will have cheap 100 Teraflop computers by 2020 - perhaps a good bit sooner! 100 Teraflops is an exciting number: Carngie Mellon computer vision researcher Hans Moravec estimates that the human brain processes information at an effective rate of about 100 Tflops. Hans arrives at the figure by estimating that the neurons in the retina make one billion calculations per second, and then scales this up by a factor of 100,000 for the volume of the whole brain. It took approximately one billion years for 100 trillion atom cells to evolve, and another billion for them to evolve into 100 trillion cell humans with our 100 teraflop brains. We are now going to reproduce this feat in about 100 years - some 10 million times faster. In general the first step is now complete - single processor speeds probably can't be pushed orders of magnitude faster in GHz. The next step is to arrange large numbers of these processors in order to make massively parallel machines.
We are already heading down this path. Dual core computers are available now, and AMD's quad core chips are on the way in 2007. By 2015 we will have computers with 100's of processor cores, if Intel's researchers have their way. There are other fascinating advances as well. In order to feed 100's of fast processors you need very high bandwidth, something that recent advances in getting silicon lasers to work will likely help with. Furthermore the processors will need a lot of data to chew through, which new forms of solid state memory should provide. For instance Ovonyx is working on a promising new type of optical memory that switches between crystalline and amorphous states when prodded electrically.
*Oops, violating causality a little bit here: I wrote most of this post in early June a little while after the Apple announcement, but then got absorbed by the Arnowitt Deser and Misner 3+1 splitting of spacetime, and am now finishing this post on July 1st.
