Slashdot Mirror
IBM Scientists Find New Way To Shrink Transistors
MarcAuslander writes that IBM scientists have discovered a way to replace silicon semiconductors with carbon nanotube transistors, an innovation the company hopes will dramatically improve chip performance and get the industry past the limits of Moore's law. According to the Times: In the semiconductor business, it is called the 'red brick wall' — the limit of the industry's ability to shrink transistors beyond a certain size. On Thursday, however, IBM scientists reported that they now believe they see a path around the wall. Writing in the journal Science, a team at the company's Thomas J. Watson Research Center said it has found a new way to make transistors from parallel rows of carbon nanotubes.
There is a good reason why that buys you nothing, and costs you much: Basically, the thermal dissipation of a device of a given geometry is what gives it state, by thermalising the state change, it prevents quantum annealing from returning it to the previous or indeterminate state. To increase the number of states you have to increase the voltage proportionally to the number of states, so to add one extra bit (remember a bit is just log2(possible states)), you have to double the number of states, and therefore the voltage; all well and good? No, because the dissipation of a device is proportional to the square of voltage, so you have doubled the bit density by quadrupling the energy consumption of the device.
And this is not even taking into account the added complexity (more gates) required to at some point discriminate these levels to implement logic. This also roughly scales to the square of the number of states. So you take the square of the square of the number of states, or raise the activation energy to the fourth power of what it was, at least making the device 8x more energy dense or less efficient for a 2x gain (this is the hard limit of information theory, real numbers are worse).
Now there of course is still room at the bottom to make these quantum annealing devices we call switches more efficient, but the way you are proposing is working in the opposite direction.
If you want a computer to be reliable, that is compute things much more often than uncompute things, you have to have a thermal bias, so that P(compute) >> P(uncompute), which we do by setting up an entropy gradient and periodically saving the result of some combinatorial equation to a register where it's value is constantly reinforced by that entropy gradient (current flowing through the latch), or else held in that state by lifting it over an activation barrier (as in memristors and Flash). Either way energy is consumed in the process, as you lift it into a indeterminate state and allow it to relax into the desired determinable state. The clock and Vdd provide together provide this energy to allow this to function.
-puddingpimp