In summary, it says that while programmers used to get "free" performance gains on single-threaded applications from rapidly increasing clock speeds and instruction-level parallelism, these gains have since leveled. Kayvon noted in class that clock rates stopped increasing in ~2005 because of limits on power consumption. The power consumed to flip a transistor is the square of the voltage required for that action, meaning that there is a non-linear increase in power associated with an increase in clock rate. Now, performance gains are achieved mainly through hyperthreading, or running multiple threads in parallel on a single CPU, and multicore architecture, or running multiple CPUs on the same chip.
I do have a question related to the first question posted on this slide. How is it that the number of transistors in a chip continues to increase so rapidly given the power consumption problem? (I.e., why is that the power problem slows clock rate increase so much more than transistor increase? Is it just because although more transistors are being added to chips, they're not being flipped as rapidly as they would need to be to get clock rate increases?)
rohitban
From what I understand, although the number of transistors is increasing exponentially, the voltage that each transistor needs to be activated decreases thus suppressing the power consumption as the power consumption is proportional to the square of the voltage across the transistor.
In summary I don't believe it is so much the power consumption as it is the heat generated per unit area that is the main problem. More transistors packed in would give a higher heat density which would need more cooling in one part of the processor than another.
yuel1
@HLAHat I believe you are correct. As we approach ever smaller transistor sizes (think atoms) we will encounter more and more challenges with the current MOSFET technologies (think quantum tunneling and gate leakage).
ekr
This graph seems to indicate that ILP has maxed out at about 6-7, even though the previous slide suggests that increasing operations per cycle above 4 provides very little performance speedup. I'm wondering why chip producers ultimately decided on that many operations per cycle? If it's because it was still worth it despite the minor speedup, it seems like that means that at some point, when transistors become so small that adding more becomes difficult, it will actually be more beneficial to start increasing operations per cycle again.
ak47
So I've never taken an architecture class but I'm wondering: what causes the voltage needed to flip a transistor to change? We say that as that voltage increases, power consumption increases quadratically...but why does voltage increase at all?
sam
I think higher voltage result in faster devices but also in higher power dissipation.
haodongl
From what I understand, the power consumption is related with processor clock rate and the number of transistors on chip. Since the increment of clock rate and power consumption have almost stopped, we will see a slowing down or stop of increasing in the number of transistors in a few years, right?
yuel1
@haodongl I don't think that will be the case. Certain heterogeneous microprocessors don't power on all components at once, so they can manage to have more transistors without boosting the power consumption.
So, it seems the transistor count is still following exponential growth, but this has to eventually taper off too, right?
Here is the article from which the graph above was taken: https://www.cs.utexas.edu/~lin/cs380p/Free_Lunch.pdf
In summary, it says that while programmers used to get "free" performance gains on single-threaded applications from rapidly increasing clock speeds and instruction-level parallelism, these gains have since leveled. Kayvon noted in class that clock rates stopped increasing in ~2005 because of limits on power consumption. The power consumed to flip a transistor is the square of the voltage required for that action, meaning that there is a non-linear increase in power associated with an increase in clock rate. Now, performance gains are achieved mainly through hyperthreading, or running multiple threads in parallel on a single CPU, and multicore architecture, or running multiple CPUs on the same chip.
I do have a question related to the first question posted on this slide. How is it that the number of transistors in a chip continues to increase so rapidly given the power consumption problem? (I.e., why is that the power problem slows clock rate increase so much more than transistor increase? Is it just because although more transistors are being added to chips, they're not being flipped as rapidly as they would need to be to get clock rate increases?)
From what I understand, although the number of transistors is increasing exponentially, the voltage that each transistor needs to be activated decreases thus suppressing the power consumption as the power consumption is proportional to the square of the voltage across the transistor.
http://www.extremetech.com/computing/190946-stop-obsessing-over-transistor-counts-theyre-a-terrible-way-of-comparing-chips
In summary I don't believe it is so much the power consumption as it is the heat generated per unit area that is the main problem. More transistors packed in would give a higher heat density which would need more cooling in one part of the processor than another.
@HLAHat I believe you are correct. As we approach ever smaller transistor sizes (think atoms) we will encounter more and more challenges with the current MOSFET technologies (think quantum tunneling and gate leakage).
This graph seems to indicate that ILP has maxed out at about 6-7, even though the previous slide suggests that increasing operations per cycle above 4 provides very little performance speedup. I'm wondering why chip producers ultimately decided on that many operations per cycle? If it's because it was still worth it despite the minor speedup, it seems like that means that at some point, when transistors become so small that adding more becomes difficult, it will actually be more beneficial to start increasing operations per cycle again.
So I've never taken an architecture class but I'm wondering: what causes the voltage needed to flip a transistor to change? We say that as that voltage increases, power consumption increases quadratically...but why does voltage increase at all?
I think higher voltage result in faster devices but also in higher power dissipation.
From what I understand, the power consumption is related with processor clock rate and the number of transistors on chip. Since the increment of clock rate and power consumption have almost stopped, we will see a slowing down or stop of increasing in the number of transistors in a few years, right?
@haodongl I don't think that will be the case. Certain heterogeneous microprocessors don't power on all components at once, so they can manage to have more transistors without boosting the power consumption.