I've never really understood why Moore's law is considered to be so important. While it definitely is an interesting bit of trivia, I don't see how the number of transistors per chip is an indication of technical progress. Just because a chip has a bunch of transistors on it doesn't mean that the chip is designed in an intelligent efficient way.
ykt
With more transistors, you could have more logic circuits and do more interesting and complex stuff than before. It just gives the option of having more capability.
tcm
Good answer, ykt! Indeed, having more transistors was incredibly important in the early days of microprocessors, since it allowed us to go from 8 bit to 16 bit to 32 bit processors. When we reached the point where we could integrate a 32-bit processor plus a 1st-level cache onto a single chip, performance improved dramatically. With more transistors, it became practical to do superscalar processing, out-of-order execution, etc. Now we can integrate multiple levels of cache onto the chip. The boundary between on-chip vs. off-chip is important, because it is very costly (both in terms of latency and power) to go off-chip.
bAAn
I would like to add that while this has been great for the high performance computing market, of late, this has been haunting the low power/low end microcontroller market. This is partly because of the inherent price pressure of the high volumes and the fact that MCU technology is nowhere close to the "cutting edge" of semiconductor technology and still lags processor technology by a couple of generations (at least). However, increasing transistor counts are making already satisfactory MCUs cheaper and straining the profit margins of these companies. Take a look at this report (http://www.eetimes.com/document.asp?doc_id=1327470) that says that even though the MCU volume will rise to 49B units from about half that today, the total revenues remain flat since the average selling price also drops proportionally to the growth. So MCU vendor sales teams need to keep more people/market harder and push more products to make the same same money.
caiqifang
Just curious, what are other popular approaches, besides splitting programs to parallel computing and enhancing single core calculation, that help in speeding up overall performance?
HIMYM
https://en.wikipedia.org/wiki/Instruction-level_parallelism This page explains the two approaches to ILP and gives an example of how to measure ILP.
hanzhoul
I wonder with the increase of transistor density, will power density increases as well? Or it will stay the same.
ferozenaina
@hanzhoul - power density is a measure of heat dissipation - in W/cm2. I think it would increase as the conductor tracks would become thinner, increasing the resistance and generating more heat.
JJ
For those of you who are doing exam review and might be unclear about superscalar, a superscalar processor is a CPU that implements ILP within a single processor by executing more than one instruction during a clock cycle on different execution units on the processor. I think there is quiz question related to this when I got it wrong.
I've never really understood why Moore's law is considered to be so important. While it definitely is an interesting bit of trivia, I don't see how the number of transistors per chip is an indication of technical progress. Just because a chip has a bunch of transistors on it doesn't mean that the chip is designed in an intelligent efficient way.
With more transistors, you could have more logic circuits and do more interesting and complex stuff than before. It just gives the option of having more capability.
Good answer, ykt! Indeed, having more transistors was incredibly important in the early days of microprocessors, since it allowed us to go from 8 bit to 16 bit to 32 bit processors. When we reached the point where we could integrate a 32-bit processor plus a 1st-level cache onto a single chip, performance improved dramatically. With more transistors, it became practical to do superscalar processing, out-of-order execution, etc. Now we can integrate multiple levels of cache onto the chip. The boundary between on-chip vs. off-chip is important, because it is very costly (both in terms of latency and power) to go off-chip.
I would like to add that while this has been great for the high performance computing market, of late, this has been haunting the low power/low end microcontroller market. This is partly because of the inherent price pressure of the high volumes and the fact that MCU technology is nowhere close to the "cutting edge" of semiconductor technology and still lags processor technology by a couple of generations (at least). However, increasing transistor counts are making already satisfactory MCUs cheaper and straining the profit margins of these companies. Take a look at this report (http://www.eetimes.com/document.asp?doc_id=1327470) that says that even though the MCU volume will rise to 49B units from about half that today, the total revenues remain flat since the average selling price also drops proportionally to the growth. So MCU vendor sales teams need to keep more people/market harder and push more products to make the same same money.
Just curious, what are other popular approaches, besides splitting programs to parallel computing and enhancing single core calculation, that help in speeding up overall performance?
https://en.wikipedia.org/wiki/Instruction-level_parallelism This page explains the two approaches to ILP and gives an example of how to measure ILP.
I wonder with the increase of transistor density, will power density increases as well? Or it will stay the same.
@hanzhoul - power density is a measure of heat dissipation - in W/cm2. I think it would increase as the conductor tracks would become thinner, increasing the resistance and generating more heat.
For those of you who are doing exam review and might be unclear about superscalar, a superscalar processor is a CPU that implements ILP within a single processor by executing more than one instruction during a clock cycle on different execution units on the processor. I think there is quiz question related to this when I got it wrong.