Note that even transistor density is no longer viewed to be able to maintain its exponential growth, since physical barriers are being hit when transistor approach atomic size. Future performance improvements will depend even more on smarter and more efficient algorithms as hardware improvements start hitting physical limits
cluo1
Another reason making the clock frequency hard to increase further is that an instruction is executed in different stages and the clock tick should fit the longest step. Even if the clock frequency grows beyond that, the processor will not get much performance gain since the short steps are still delayed by long steps.
bochet
At least in this plot, transistor density is still increasing exponentially while the other curves stops increasing. What's the benefit of higher transistor density when the frequency freezes? I think lower power consumption and smaller trips are two advantages, but is there any more?
Also read the original article (The Free Lunch Is Over) here:
ykt
Smaller transistors require lesser power. So with the same power requirements, you can have more transistors. More transistors means more ability to add complex logic to your CPU. Maybe add more complex instructions.
Cake
If we are unable to take advantage of the green curve, is it still a good idea to try and pack more and more transistors in a single chip, given that there are potentially more downsides as mentioned in class?
R2D2
The green curve is transistor density, you can't pack more transistors in a single chip unless you DO actually take advantage of the green curve. Transistor density and frequency scaling go hand in hand, as transistors shrink, they can be operated at a faster clock rate. But since frequency is scaling due to power problems, the green curve offers the advantage of putting more logic in a chip of fixed area. This trend will also max out because transistors cannot shrink beyond a certain limit and too many closely packed transistors will also result in higher heat density.
Bye
I would say that there are incentives to pack more transistors out of two reasons.
Transistors are not limited to instruction execution. Many other functions within a CPU require transistors and their performances could potentially benefit from improved transistor design and architecture, CPU cache, for example.
There is multi-core design. When we hit the boundary for clock rate of a single core, we might as well improve the performance by having multiple cores inside a CPU and ask software engineers to write parallel codes. Well that's why we take this course..
nishadg
If these trends of exponentially growing clock frequency and transistor density continued, rather than leveling off, the power density (and therefore the heat dissipation necessary) would be the same as that of a nuclear reactor or a rocket nozzle.
dmerigou
The transistor density will also eventually stall because of physical limits. First, you can't make layers of silicon that are less than 1 atom wide. Second, to make a batch of silicon wafers, you use photolithography techniques in many etching phases , and the precision of the engraving depends on the wavelength used; nowadays X-Rays are used because visible light or UV have a too large wavelength that would cause diffraction phenomenons.
blah329
This slide illustrates several points, such as how Moore's Law about the increase in transistors is still upheld. As a result of better transistor technology, the clock frequency has also increased, and this has been accompanied by an increase in voltage, since more voltage is needed to drive an oscillator at a higher frequency. Since power is directly related to frequency and voltage, an increase in these two factors leads to an increase in power, as shown by the blue curve. Although higher clock frequencies--on the order of terahertz-can be achieved, it is not feasible on chips, because operating at these frequencies generates far too much heat, which causes the silicon to melt. This explains why both the power and clock frequency curves have both begun to level out. Since ILP is already maxed out for processors, one way to acquire quicker processing units is by increasing the amount of resources within them--i.e. adding more cores.
xka
The path to increasing speed over time:
Architects use ILP to run multiple independent instructions at once, but programs do not inherently lend themselves to many independent instructions
Architects continue to add more transistors per area, but the ability to diffuse heat eventually limits the number of transistors
Speedup is deferred to software developers who now have the ability to use multiple cores, "parallel programming"
dyzz
another interesting reason why transistor density is unable to further increase is a particle physics boundary. We are already starting to get close this limit with modern processors. Essentially the transistors will eventually hit a limit where Heisenberg uncertainty will begin playing a large role. There are predictions that 2036 is when technology will begin to bump up against these quantum uncertainty principles.
Further reading: http://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=4567410
ShadoWalkeR
Most desktop CPUs still use air cooling with plastic fan coolers even though there are better cooling options available. For example using liquid cooling can drop GPU temperatures from 90 Celcius to 55 Celcius under similar loads. So, with liquid cooling, we can still increase the frequency by a few percent.
cluo1
Still confused about the relationship between clock frequency and transistor density on the chip
fxffx
This diagram shows that the clock frequency can not scale up anymore, and this is possibly because of the heat problem. As the frequency becomes too high, the CPU can not diffuse all the heat it generates.
Power problems: heat isn't effectively diffused
Note that even transistor density is no longer viewed to be able to maintain its exponential growth, since physical barriers are being hit when transistor approach atomic size. Future performance improvements will depend even more on smarter and more efficient algorithms as hardware improvements start hitting physical limits
Another reason making the clock frequency hard to increase further is that an instruction is executed in different stages and the clock tick should fit the longest step. Even if the clock frequency grows beyond that, the processor will not get much performance gain since the short steps are still delayed by long steps.
At least in this plot, transistor density is still increasing exponentially while the other curves stops increasing. What's the benefit of higher transistor density when the frequency freezes? I think lower power consumption and smaller trips are two advantages, but is there any more?
Also read the original article (The Free Lunch Is Over) here:
Smaller transistors require lesser power. So with the same power requirements, you can have more transistors. More transistors means more ability to add complex logic to your CPU. Maybe add more complex instructions.
If we are unable to take advantage of the green curve, is it still a good idea to try and pack more and more transistors in a single chip, given that there are potentially more downsides as mentioned in class?
The green curve is transistor density, you can't pack more transistors in a single chip unless you DO actually take advantage of the green curve. Transistor density and frequency scaling go hand in hand, as transistors shrink, they can be operated at a faster clock rate. But since frequency is scaling due to power problems, the green curve offers the advantage of putting more logic in a chip of fixed area. This trend will also max out because transistors cannot shrink beyond a certain limit and too many closely packed transistors will also result in higher heat density.
I would say that there are incentives to pack more transistors out of two reasons.
Transistors are not limited to instruction execution. Many other functions within a CPU require transistors and their performances could potentially benefit from improved transistor design and architecture, CPU cache, for example.
There is multi-core design. When we hit the boundary for clock rate of a single core, we might as well improve the performance by having multiple cores inside a CPU and ask software engineers to write parallel codes. Well that's why we take this course..
If these trends of exponentially growing clock frequency and transistor density continued, rather than leveling off, the power density (and therefore the heat dissipation necessary) would be the same as that of a nuclear reactor or a rocket nozzle.
The transistor density will also eventually stall because of physical limits. First, you can't make layers of silicon that are less than 1 atom wide. Second, to make a batch of silicon wafers, you use photolithography techniques in many etching phases , and the precision of the engraving depends on the wavelength used; nowadays X-Rays are used because visible light or UV have a too large wavelength that would cause diffraction phenomenons.
This slide illustrates several points, such as how Moore's Law about the increase in transistors is still upheld. As a result of better transistor technology, the clock frequency has also increased, and this has been accompanied by an increase in voltage, since more voltage is needed to drive an oscillator at a higher frequency. Since power is directly related to frequency and voltage, an increase in these two factors leads to an increase in power, as shown by the blue curve. Although higher clock frequencies--on the order of terahertz-can be achieved, it is not feasible on chips, because operating at these frequencies generates far too much heat, which causes the silicon to melt. This explains why both the power and clock frequency curves have both begun to level out. Since ILP is already maxed out for processors, one way to acquire quicker processing units is by increasing the amount of resources within them--i.e. adding more cores.
The path to increasing speed over time:
Architects use ILP to run multiple independent instructions at once, but programs do not inherently lend themselves to many independent instructions
Architects continue to add more transistors per area, but the ability to diffuse heat eventually limits the number of transistors
Speedup is deferred to software developers who now have the ability to use multiple cores, "parallel programming"
another interesting reason why transistor density is unable to further increase is a particle physics boundary. We are already starting to get close this limit with modern processors. Essentially the transistors will eventually hit a limit where Heisenberg uncertainty will begin playing a large role. There are predictions that 2036 is when technology will begin to bump up against these quantum uncertainty principles.
Further reading: http://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=4567410
Most desktop CPUs still use air cooling with plastic fan coolers even though there are better cooling options available. For example using liquid cooling can drop GPU temperatures from 90 Celcius to 55 Celcius under similar loads. So, with liquid cooling, we can still increase the frequency by a few percent.
Still confused about the relationship between clock frequency and transistor density on the chip
This diagram shows that the clock frequency can not scale up anymore, and this is possibly because of the heat problem. As the frequency becomes too high, the CPU can not diffuse all the heat it generates.