Tác giả giữ toàn bộ bản quyền (All Rights Reserved) — không phải giấy phép mở. Bạn vẫn xem và học miễn phí trên OpenER; muốn tái sử dụng cần xin phép tác giả.
[MUSIC PLAYING] PROFESSOR: Hello, and welcome back.
Let's talk a bit about energy efficiency. Energy efficiency is something that relates processor power with the processor performance.
Power dissipation has become much more important over the past couple of decades. Back in the '80s, architects and designers had very different design objectives.
Back then, the only goal was to increase the performance as fast as possible. So they did everything they had.
They used all the tools that they had under their belt to do that, to increase the performance.
So they shortened the logic depth, they used the maximum possible supply voltage to crank up the frequency of processors, and, as a result, processor performance was doubling every other year.
But the price, in power, was high. The power was doubling every approximately three years.
That was not that big of an issue in the early '80s, when the processors were relatively simple and the power consumption was relatively low.
But, over time, the power consumption hit the limits. And, essentially, across all the power domains, the power became the primary limiting factor sometime in the early 2000s.
And, nowadays, power is limited in every compute domain.
Power is limited for cell phones, for desktop processors, and for warehouse-scale computers in data centers.
The nature of these power limitations is different. For example, in cell phones, the power is limited by the size of the battery that we are willing to carry on ourselves.
In desktops, the power is limited by the amount of heat that a relatively inexpensive fan can remove from a processor.
And, in data centers, it's both the heat and also the cost of energy that it takes us to run that data center.
So the goal of designs, nowadays and in the foreseeable future, is to maximize the performance under power constraints.
So it's not anymore to just get the maximum performance, but to deliver available performance under power and energy constraints.
In order to get a bit better insight into energy efficiency, let's first understand, where does the energy go in CMOS? All our logic gates are built in CMOS technologies, nowadays.
So let's examine the very simple inverter, here. Here is the logic symbol of an inverter that we have seen many times. There are three additions to it.
It got the power supply and the ground, so supply voltage and the ground, because it needs a supply and a ground in order to switch the logic levels from 0 to 1.
And then we see capacitance at its output. So there is a symbol of an inverter on the left-hand side and a schematic of an inverter on the right-hand side.
They represent exactly the same circuit, just two different levels of abstraction. The capacitances happen because the nature of transistors.
Every transistor that we use in an inverter has its input and output capacitances.
So whenever we are trying to switch the output level of an inverter from 0 to 1, we have to charge a capacitor.
That capacitor consists of capacitances that are at the output of this inverter, but also at the inputs of any of the inverters that are downstream that are being driven by this inverter.
How much energy does it take us to charge a capacitor from 0 to 1, from 0 to V? Well, the amount of energy is equal to CV squared.
If you remember from high school physics, if you haven't taken any EE classes yet, half of that energy gets stored in a capacitor and the other half gets dissipated during the charging period.
So, in the end, we end up with the half of CV squared on this capacitor, and a half of CV squared was dissipated on this transistor.
Eventually, when we switch back from 1 to 0, we dump that half CV squared into ground.
That is the dominant part of power dissipation, these days, or energy dissipation in CMOS technologies. The other one is the leakage.
Leakage stems from the fact that transistors are not ideal switches anymore. They were much more ideal back in the '80s and '90s.
You know, when we would turn off the transistor and we would set this, its voltage, input voltage, to zero, it would really turn off.
There would be very little current that would be trickling through that transistor. Nowadays, they are more like dimmers, if you prefer an equivalence to light switch.
So we are just able to dim the current that is going through the transistor. So when transistor is supposed to be off, it is still leaking some amount of current.
And one thing that you may find somewhat paradoxical, that the balance between the switching energy that goes into charging capacitance and leakage energy is somewhat surprising.
We spend about 70% on charging and discharging capacitances, which is our useful use of the energy, and 30% may go into leakage. That may be a bit of a surprise.
It comes from the basic things that are associated with technology, with the CMOS technology, and we are not going to dive into that.
That is covered in other classes like 151 or 130. But let's pop back up at the bit of a higher level.
How much energy does it take us to complete a task? And let's say that our task is to execute a program, for example, to compress a video.
So that's, again, a complex thing to evaluate and there are many parameters, so it is a good idea to break it down.
And, in this case, it's a good idea to break it down into two components.
So energy per program which may be executing our tasks can be broken down into the number of instructions that we have in the program and the energy per instruction.
So the number of instructions per program, we have already encountered when we were trying to measure the performance. And, in this case, the shorter the program, the less energy it is going to use.
I don't need to say much more about it.
That is usually outside of designer's range of things that they can influence, but it does affect the architectures and general computer design.
On the other hand, what we can much more influence is this second part, which is the energy per instruction, and the dominant part of that is CV squared.
There is some leakage that goes on, but I'm going to ignore the leakage at the moment.
So there is some energy that is associated with computing that is proportional to the capacitance times the voltage squared.
And capacitance, like we've had in a simple case of an inverter, is now a total equivalent capacitance of all the gates that are switching, say, in a processor core.
That capacitance is the capacitance that we need to switch, in every instruction, to move certain amount of wires from 0 to 1.
It depends on the technology that we are using and also on the architecture or processor and its features.
On the other hand, the supply voltage is a very strong parameter, here, in trading off power for performance or energy for performance.
Supply voltage nominally is around 1 volt, more like 0.8 volts or 0.7 volts in the recent single-digit nanometer technologies, but the energy depends quadratically on the supply voltage.
So if we lower the supply voltage by 10%, we save 20% of the energy, 21% to be more — 19% to be more precise.
All right, so the basic goal, here, of a designer is to try to reduce the capacitance and the voltage to reduce the energy per task, and voltage is a more powerful knob.
One thing to keep in mind, lowering the voltage also affects the performance because transistors are slower when they run at a lower supply voltage.
So trading off — changing supply voltage trades off energy savings for loss in performance.
And we are going to see, in different compute domains, we're going to be seeing processors operating at a bit of a higher voltage versus a little bit less of a voltage to achieve different power performance trade-off point.
Okay, talking a little bit more about energy trade-off examples, one thing that happens in every — when we scale the technology node and implement a new processor, we get some natural energy savings.
So back in the '80s and '90s, every time — and up to probably about 2010 — every time we scaled the technology, we go into the next node in Moore's law, the capacitance would shrink by about 30%.
We would save 30% of our capacitance. Now, those capacitance savings are more like 15% from node to node.
They're still — you know, don't throw away 15%. 15% discount is huge.
Supply has recently been also scaling by about 15%, and that has been traditionally going on since like mid-'90s or so.
Early on, the supply reduction — again, supply is supposed to be reduced by 30% to match what we get from the capacitance scalings, and that is something that is associated with Dennard's law — but it has been more like 15% because designers were greedy and they wanted to maximize the performance
and they were just trying to keep that supply as high as possible.
Now, supply reductions are limited by this leakage because it affects — if we really want to drop the supply voltage, the leakage is going to go up.
So as a result, when we multiply C and V squared now from a node to node, we get almost a 40% discount in energy. That allows us to do things with that.
I mean, when we build a new product, if you keep exactly the same architecture and same number of cores, it naturally becomes 40% more energy efficient or has 40% less energy.
Or we can pack 40% more stuff in there if we have the fixed energy budget. So that's kind of convenient. That's exactly what has been happening with the scaling lately.
So in a brief summary, the energy efficiency has been improving because of Moore's law and reduced supply voltage, and architects have been using that to pack more features and improve performance from that.
A little bit about power and performance trends over time. This is a plot that we have shown in the first lecture.
It is showing years of introduction of a whole lot of processors over the past 48 years since Intel's 4040 in the early '70s.
And then, here, we see the number of transistors, which is Moore's law, steadily increasing with a little bit of a slowdown up here.
Then we are seeing frequency and single-thread performance, power, and then number of cores.
So what we have seen is we — frequency was the primary means for gaining the performance in the '80s and '90s, and architects and designers worked together to crank it up.
But it hit a limit in early 2000s and has been flat since. The reason for that is because the power had to stay flat. So architects worked hard.
They had to continue working on trying to improve the single-thread performance, which is what we have seen here.
But, also, we're utilizing parallelism by adding more cores to add the performance to add to end applications.
How things are going to change, over time, with the end of Moore's law, it's not easy to predict it exactly, but we got some sense.
So, first, these capacitance savings are not going to come that easy. Voltage savings are not going to happen easily either.
Simply, as I mentioned before, we can't reduce the voltage, otherwise things are going to leak.
In order to — so capacitance savings are also not going to happen for free because we are slowing down this shrinkage of transistors themselves.
Lots of savings in the capacitance are likely coming from packing things in a third dimension.
So when things are spread apart on a board, in two dimensions, we get a lot of system capacitance basically over there.
But if you can pack things on top of each other, we are going to save quite a bit of energy there because we'll have less capacitance, and that is going to try to help us work around this power wall.
The important takeaway, here, is a basic relationship between the performance and power. And it is shown here as the Energy Iron Law.
Energy efficiency, which is managed — which is expressed as the number of tasks that we can perform per joule, is what relates the performance and the power.
So performance is the number of tasks that we can execute per second. Power is joules per second. Energy efficiency is tasks per joule.
So energy efficiency, now, becomes really the key metric of our designs, and it is something that tells us actually how much performance we can get out of a power-limited design.
If the power-limited design is a cell phone, the power on the average is 600 milliwatts. I mean, that's what we can drain out of the battery on the average, what is specked out.
The max power of cell phones is limited to a few watts, but for a very short burst of time.
On the other hand, if you are working with a Google-scaled data center, that may have a power envelope of 20 megawatts.
That tells us, the energy efficiency of that entire data center, tells us what kind of a performance we can get out of a data center. And that's basically it.
We are going to look into pipelining as a way to improve performance after the break. See you then.
Đăng nhập để đánh giá học liệu này.
Đăng nhập để bình luận — chưa ai bình luận về học liệu này.
Không có khóa học liên quan.