1
00:00:03,620 --> 00:00:09,645
I'll just skip way prediction cuz I don't
think it's that important.

2
00:00:09,880 --> 00:00:15,906
I did want to say one or two words about
critical word first and early restart.

3
00:00:15,906 --> 00:00:21,932
So, critical word first, basic idea here
is that you can express to the memory

4
00:00:21,932 --> 00:00:25,362
system which word is the most important
word.

5
00:00:25,362 --> 00:00:31,430
And usually RAM arrays the buses are
oriented in a way that they'll actually

6
00:00:31,430 --> 00:00:38,029
allow it takes multiple pumps or multiple
cycles to transmit the data across the

7
00:00:38,029 --> 00:00:41,519
bus.
And usually, these arrays don't care at

8
00:00:41,519 --> 00:00:47,966
least, like main memory, like, like, D-RAM
arrays, big, big thing on your, on your

9
00:00:47,966 --> 00:00:51,380
chassis or on your system server
motherboard.

10
00:00:51,416 --> 00:00:56,526
Cough They don't really care if they
return the data in this order, zero, one,

11
00:00:56,526 --> 00:01:01,340
two, three, four, five, six, seven and
eight, Probably seven, Sorry.

12
00:01:01,340 --> 00:01:07,550
Or, if you return it in a different order
because they just wanna use the bus as

13
00:01:07,550 --> 00:01:11,620
much as possible.
So, what you can do is you can say, COUGH

14
00:01:12,220 --> 00:01:16,027
address three is the most important one,
please return that first.

15
00:01:16,027 --> 00:01:20,773
So, what can happen is it will return the
data and right here, address three comes

16
00:01:20,773 --> 00:01:25,049
back, and it comes back first always,
being cuz that's the address that we're

17
00:01:25,049 --> 00:01:29,435
trying to access for this loadness.
And can start the CPU executing again

18
00:01:29,435 --> 00:01:33,062
while the rest of the cache line is
filling into the cache.

19
00:01:33,062 --> 00:01:37,979
So you've basically done a little bit of
overlapping but this doesn't help that

20
00:01:37,979 --> 00:01:40,560
much.
Because usually, you know, there's not

21
00:01:40,560 --> 00:01:45,085
that many a cache line is not that long so
it might help you a little bit.

22
00:01:45,085 --> 00:01:49,311
So, you know, if it takes eight cycles to
fill the cache, you might save seven

23
00:01:49,311 --> 00:01:53,815
cycles if you were going to the last, if
you were trying to load the last word in

24
00:01:53,815 --> 00:01:56,929
the cache line.
But if you were trying to, let's say, load

25
00:01:56,929 --> 00:02:00,377
the second line in the cache line, it only
saves you one cycle.

26
00:02:00,377 --> 00:02:05,940
And there is some complexity in actually
building critical word first systems.

27
00:02:06,055 --> 00:02:11,178
Cough Early restart is another way to do
this which little less hardware.

28
00:02:11,178 --> 00:02:16,512
Instead of having the memory system rotate
the memory that returns back to you, it

29
00:02:16,512 --> 00:02:19,700
still returns the memory in the canonical
order.

30
00:02:19,960 --> 00:02:24,617
But, we just wait until let's say, we want
to dress three here, it comes back and we

31
00:02:24,617 --> 00:02:27,628
restart there.
So you can see there's a little bit of

32
00:02:27,628 --> 00:02:32,424
time here which we could overlay it block
the fill in with executing in the CPU

33
00:02:32,424 --> 00:02:34,754
again.
In the canonical case, if we don't have

34
00:02:34,754 --> 00:02:39,303
early restart, what's going to happen is
we have to wait for the entire cache line

35
00:02:39,303 --> 00:02:43,685
to come in and then we can restart, so we
lose, we lose these, these cycles there.

36
00:02:43,685 --> 00:02:47,458
We don't get that overlap.
Okay. Let's stop here for today and I'm

37
00:02:47,458 --> 00:02:52,006
going to remove the, the weight prediction
stuff from the material we talked about

38
00:02:52,006 --> 00:02:53,560
cuz it's not that important.