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.
And welcome back. Now let's think about how to improve the performance of a system using pipelining. So again. Let's talk about this maximum clock frequency.
How small can we make our period? How high can we make our frequency? So we think about the delay through a circuit. Here's a standard circuit we've now tilted on its side.
We've got the idea that first thing that happens like what happens at dawn? What happens at dawn is clock goes high. That clock is going to go high.
And that rising edge of the clock starts the whole day in motion. The first delay we see is t clock to Q. How long till I see that output on the inputs?
Turned around, grabbed it's stuff and then it's available — Q is stable with whatever the previous state was. I've got a delay. We're going to assume that this wire has no delay.
So maybe in an exam we might say well, assume there's a wire delay. But for now, and by default, there's no delay in the wire.
So the next delay you see is t sub combinational logic. Whatever the commercial logic delay is. There might also be a delay because this input may not be synchronous.
We talked about that before as well. Next, we've got this line which is t setup.
We've got to have that stable t clock to Q plus T sub cl stable before t setup, before the next rising edge.
So in fact, we think about our max delay as being clock to Q. Again, clock to Q here. That's the first delay you see.
The cl delay here and our setup time, which is the last thing that happens all that has to fit into one period.
So we saw this idea that we had a register and then you have here is an adder and a shifter. So we're going to have now two elements.
And the way I think about this is imagine an assembly line. And Lucy is working on an assembly line and she's got to do something.
She's got to wrap the chocolates and she's got to write a little poem, a little haiku on it at the same time. And she has to complete successfully before she can pass it on.
Well, that means the assembly line is going to have to wait till she can do both of those things in the worst case. OK, next one. Do the wrapping and then the haiku.
That's a lot. That's like the adder and the shifter. Couldn't you have another person who then does the haiku? Then you're dividing her work into two pieces and now you divide it up.
But now, it takes a little longer.
Now rather than going from, let's say brand new chocolate to one Lucy and out, I'm now going to have to figure out, well can I divide Lucy's chocolate into two pieces and then maybe have somebody to help divide the load, help share the load with that.
And maybe it then goes to the next stage. So it might be longer before the chocolate gets out because it's there. But the frequency is faster.
The time between I can get to the next chunk, the time between the period is shorter. So here's the idea. I got to register. The clock goes high.
Remember we saw this before. The clock goes high. I wait t clock to Q before I see that output on my inputs. Then I have this really big combinational logic delay.
This is an adder or shifter delay in this case. And until that's stable. And then all that has to be stable t setup before the next rising edge of the clock.
So my clock period is limited by that huge, it's that those three terms add to the minimal — middle clock period is those three terms.
Well, one of those terms is really big and then that means that minimum is really big. Could I do anything about that? And that's the idea of pipelining.
What if I inserted another register here in the middle? Oh, my gosh. Now I'm taking Lucy's job of wrapping and doing the haiku and doing a separate thing.
Here again, an adder and a shifter. So now what's our timing? Rising edge of the clock. Now my inputs are available.
I have clock to Q sitting here on the adder. Now I've got my adder delay, which is much smaller.
This outer delay better be done before t setup, before the next rising edge of the clock. And then, it goes to the next stage.
And the next stage I've got t clock to Q, t clock to cube before, that number is available to s of i minus 1. So here's t clock to Q. There's this.
Now I've got a delay between when s of i minus 1 passes to the next level and that's t shifter delay. So that's right in here is my t shifter delay.
And again, that's smaller than shifter plus t adder.
And again, that has to be all stable before — oops, now it's drawing a straight line there — but stable before t setup for the next guy. So we love this.
I have more outputs per second, but you're paying a price.
So it's like you're paying a price that for one particular chocolate, for one data to go through it takes a little bit maybe longer for one thing to go through.
But overall pipelining means that more things per second, more outputs per second. We love that. So this is the last slide. Let's recap some of our terms here.
We've got a clock. That's our heartbeat of the system.
The setup time is the time before the rising edge of the clock where the data has to be stable on every flip-flop or every register.
Hold time is where that data has to be stable after the rising edge of the clock.
Clock t Q is delay between when rising edge happens and that data is quiescently available at the output ports.
Flip-flop is how you build registers and registers we're going to use all throughout building our data path for our working RISC V machine.
Registers are going to hold these data and only change the values on the rising edge of a clock or on that kind of load level or when there's a reset it'll ignore the input and then use and reset to 0.
That's it. we'll see at the next video.
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.