WEBVTT

1
00:00:09.144 --> 00:00:34.433
Okay.

2
00:00:37.289 --> 00:00:49.920
Okay.

3
00:01:32.245 --> 00:02:12.085
Okay.

4
00:02:26.219 --> 00:02:32.099
Okay hey, good morning. Everyone while I'm waiting for people to arrive.

5
00:02:32.099 --> 00:02:35.430
Could you tell me if you can hear me.

6
00:02:35.430 --> 00:02:41.669
I've got the chat window open, so oh, great, Tom. Thank you.

7
00:02:41.669 --> 00:02:47.159
Hey, thanks, Amanda. Cause I, if I turn on the sound, I'll get feedback so.

8
00:02:47.159 --> 00:02:53.340
And while we're waiting for people to arrive, if anyone has any questions before the start of the lecture. Yeah.

9
00:02:55.650 --> 00:03:00.599
Okay, and.

10
00:03:00.599 --> 00:03:07.020
Okay, well it's good to set up. I've got while I'm waiting I'm actually have, um.

11
00:03:07.020 --> 00:03:13.590
2 machines in front of me, I've got an iPad, which is a host, and I also have my.

12
00:03:13.590 --> 00:03:19.139
Think pad here, which is a reasonable parallel computer and it's all right actually.

13
00:03:19.139 --> 00:03:22.919
And in the, um.

14
00:03:22.919 --> 00:03:28.979
Take pad, I got the chat window open, because for some reason, I can't get the chat window working on the iPad.

15
00:03:28.979 --> 00:03:40.199
So, I'm sharing the, I'm sharing the iPads, so I'm sharing the iPad screen so actually, I could do more of that, I guess. Okay.

16
00:03:41.250 --> 00:03:44.310
Great. Okay.

17
00:03:44.310 --> 00:03:49.740
Well, probably.

18
00:03:50.935 --> 00:04:03.805
Kind of just hard to start, so just we'll keeping stuff. I've got the chat window open and I'm in theory I'm watching it and you can also speak up. I think this is a small. This is a small class.

19
00:04:03.805 --> 00:04:07.615
So, 16 people registered last. I checked yesterday.

20
00:04:07.949 --> 00:04:12.120
So, okay.

21
00:04:14.250 --> 00:04:19.860
And I'm recording this, and if things work out, I can post a recording.

22
00:04:21.089 --> 00:04:25.949
All decide where to post it later? Um, possibly just on my own website.

23
00:04:27.718 --> 00:04:35.488
Okay, so, in any case, so this is the course of.

24
00:04:35.488 --> 00:04:41.639
Applied parallel computing for engineers I'm Franklin and.

25
00:04:41.639 --> 00:04:45.088
If anyone's not here to call me no joking.

26
00:04:45.088 --> 00:04:54.358
So that what I want to do today is go through the syllabus, give you a chance to ask any questions.

27
00:04:54.358 --> 00:05:01.889
And to, um, and then to go through an introduction to parallel computing, someone.

28
00:05:03.149 --> 00:05:07.139
And hello so.

29
00:05:07.139 --> 00:05:11.999
So, the idea is the point about this course.

30
00:05:11.999 --> 00:05:15.749
Is to get students and knowledge and experience hands on experience.

31
00:05:15.749 --> 00:05:19.319
And tools to develop parallel computing, um.

32
00:05:19.319 --> 00:05:25.019
Um, programs and applications, and I will largely in this course.

33
00:05:25.019 --> 00:05:28.139
Of our equipment, you can actually afford to buy.

34
00:05:28.139 --> 00:05:41.579
So this is, I'm comparing it to the same course over in computer science. Obviously he runs the Super computer center and we'll have you using.

35
00:05:41.579 --> 00:05:49.199
Um, very fast, but also, very expensive machines. All we teach you smaller machines. Um.

36
00:05:49.199 --> 00:06:01.168
1st, almost any application can use the smaller machines and 2nd is large machine. It's actually, at this point, an assembly, it blocks a smaller machines that I'll be teaching you about.

37
00:06:03.113 --> 00:06:14.634
And I'll be teaching you some software at some theory, but I'll be more in the applications as the title for the course. So so target for the courses generally, seniors and grads, though.

38
00:06:14.634 --> 00:06:18.923
I'm very welcoming to students from other departments and programs. So.

39
00:06:19.319 --> 00:06:25.619
Basically, if you want to take the course, you're welcome to take the course. And I don't, um.

40
00:06:25.619 --> 00:06:34.499
Again, I don't try to duplicate really Chris brother's course. So you're welcome to take both courses.

41
00:06:34.499 --> 00:06:45.418
And, for example, well, he's teaching some, but I will not teach him. Well, partly, um, for those of you who know what npi is, which is past the interface.

42
00:06:45.418 --> 00:06:58.228
Is I think, in most cases, it's useless, because in most cases, your job is not large enough that you need to distribute your, um.

43
00:06:58.228 --> 00:07:04.288
Program over the distributed machines, and also the overhead is considerable.

44
00:07:04.288 --> 00:07:17.488
So, in any case, you may take all the courses at the end of the course I make it. Oh, yeah. I'll put in a component of quantum computing at the end of the course, because it's fine.

45
00:07:17.694 --> 00:07:22.074
And and it's new and so on,

46
00:07:22.074 --> 00:07:26.903
and on a few pieces from the quantum computing course that I've taught in the fall,

47
00:07:27.384 --> 00:07:27.564
um,

48
00:07:27.564 --> 00:07:29.483
the hardware that I'll teach it,

49
00:07:29.514 --> 00:07:29.994
um,

50
00:07:30.593 --> 00:07:30.954
um,

51
00:07:30.954 --> 00:07:31.403
1st,

52
00:07:31.434 --> 00:07:33.233
multicolored Kelsey on.

53
00:07:33.833 --> 00:07:38.634
So, you know, everything is multi core. My iPhone is multi core and.

54
00:07:38.999 --> 00:07:43.949
The machine in my lab will tell you more about later, has some.

55
00:07:43.949 --> 00:07:50.608
Dual 14 cores. Each core is a hyper threaded, so that's like 56 threads on the machine.

56
00:07:50.608 --> 00:07:59.428
And so that's the, if you have to, or you wish to do an application in parallel to your 1st target.

57
00:07:59.428 --> 00:08:10.259
Oh, should be just a multi currency on for various reasons that its shared memory and is the learning curve is is, uh, much smaller.

58
00:08:10.259 --> 00:08:25.048
And so these are, um, this is the easiest firewall computer to start with the next step with that, which will also teaching the course of NVIDIA accelerator and.

59
00:08:25.048 --> 00:08:33.298
Because there you can have thousands of course, now invidious terminology changes some time to time.

60
00:08:33.298 --> 00:08:43.318
So, they've actually recently started deprecating the term CUDA f*** up. It's a not compute unit unified device architecture and.

61
00:08:43.318 --> 00:08:50.969
It so you can have thousands of threads running in parallel, but each core is much less powerful.

62
00:08:50.969 --> 00:08:54.688
And there's also as memory available, but.

63
00:08:54.688 --> 00:08:59.158
These have, um, you know, this is the way you'd scale up to.

64
00:08:59.158 --> 00:09:04.019
And also the, the fastest machines in the world.

65
00:09:04.019 --> 00:09:09.269
Tend to have thousands or tens of thousands of and.

66
00:09:09.269 --> 00:09:13.918
Well, such as Chris machine and.

67
00:09:13.918 --> 00:09:25.254
Progress and videos. Interesting. Of course, he started out as making Co processes for gaming on PCs and they, um, NVIDIA people are very smart, hardworking people that like to make money.

68
00:09:25.703 --> 00:09:29.663
They their customers observed that the 1st and video.

69
00:09:31.073 --> 00:09:45.474
We're very fast for doing some graphics Co processing things involving screen buffers and so on. So, people were contenting their applications to look like a graphics application. So they could run it on in video.

70
00:09:46.528 --> 00:09:53.999
And, um, this is had a field of research called graphics program.

71
00:09:53.999 --> 00:09:59.099
General programming and and the immediate people, they observed what.

72
00:09:59.099 --> 00:10:05.999
Their customers were trying to do, and they modified their.

73
00:10:05.999 --> 00:10:12.839
To be more generally useful. For example, they had I triple the standard double position floating point. For example.

74
00:10:12.839 --> 00:10:19.979
So, they're an interesting example of how to be business success, which is look at what your customers want and give it to them.

75
00:10:19.979 --> 00:10:23.219
You know, it's an interesting idea. It's a.

76
00:10:23.219 --> 00:10:30.629
Surprising that all companies don't do it, but some more more recently they've observed it. Don't.

77
00:10:30.629 --> 00:10:38.849
People want to do the machine learning and so on, and they've got it hardware to do a particular. They've identified a particular thing in machine learning.

78
00:10:38.849 --> 00:10:47.099
That was done a lot and it was half precision vector multiply an ad.

79
00:10:47.099 --> 00:11:00.568
The, um, there's not a lot of significant digits with a lot of machine learning data. So they used actually a 2 by 16 foot floating point and headway and added hardware to take a vector. And.

80
00:11:00.568 --> 00:11:09.808
Multiply matrix and that effect are very fast. Those are the Tensor units you have for and they've also getting back to the graphics roots have observed. Um.

81
00:11:09.808 --> 00:11:22.678
People like gray tracing, so they've added hardware to do things like, test array against the sphere or box and those would be their retracing units. And then what Nvidia did is they.

82
00:11:22.678 --> 00:11:31.469
Combined features, they got genuine synergy so they, they have something that's called real time. Retracing. And what? Um.

83
00:11:31.469 --> 00:11:38.759
And video does is they don't rate trace in detail, every pixel, these machine learning techniques.

84
00:11:38.759 --> 00:11:44.788
To determine which pixels are more important to compute more accurately and that's how they get effectively real time. Very tracing.

85
00:11:44.788 --> 00:11:58.318
So, it's a video and a case, and the most recent in video don't even have video out because they're high end. Customers are not doing graphics. Excellent general, Super computing. So that and video.

86
00:11:58.318 --> 00:12:09.958
Oh, so for a while they were calling them compute accelerators and they gave up on that back to GP. Okay. So that's the 2nd, architecture.

87
00:12:10.884 --> 00:12:24.323
And I'll have a little on then finding a middle module copy from my phone course on quantum computing. So those are the main types of architecture and I may have most and most of this will only be on a machine in my lab.

88
00:12:25.649 --> 00:12:29.729
Well, not real quantum computing, and I may do a little on, um.

89
00:12:29.729 --> 00:12:34.859
On, um, Amazon, perhaps I haven't decided yet.

90
00:12:36.928 --> 00:12:40.649
so that's hardware well for people that arrived 

91
00:12:40.649 --> 00:12:46.708
More recently to this class, I have a chat window open and you're welcome to type.

92
00:12:46.708 --> 00:12:57.749
um chat questions if you wish and i'll i'll try to go ahead so let's if i see i'm not always watching the window flew off to the side of my site but i do have that open 

93
00:12:57.749 --> 00:13:04.078
And so far as I know, I've not muted people also to speak a small class. You could even speak up.

94
00:13:04.078 --> 00:13:07.318
That was the hardware now the software.

95
00:13:07.318 --> 00:13:12.384
Um, um, there's various tools.

96
00:13:12.384 --> 00:13:25.854
Well, there's low level stuff we'll talk a little about and comment and put on the syllabus here about some models, um, parallel computation and these would be if you have a, a C plus plus program. Um.

97
00:13:26.844 --> 00:13:40.974
What does the standard say, for example, previous C plus plus the standard did not say the expression within a statement did not specify anything about the order they would get evaluated in so that if it's circle arguments to a function called argument was a.

98
00:13:42.808 --> 00:13:51.599
Was an expression said nothing about what order they'd be evaluated. Correct? So, we've got, um, standards about what the abstract machine model must say.

99
00:13:51.599 --> 00:13:55.078
The latest C plus plus, um.

100
00:13:55.078 --> 00:14:00.599
It's a little stricter on that actually, but we'll talk about some of the underlying parallel models.

101
00:14:00.599 --> 00:14:06.058
Now, the reason that he's underlying model to not specified things like, um.

102
00:14:06.058 --> 00:14:09.869
Strict ordering is that it's very expensive to implement that.

103
00:14:09.869 --> 00:14:14.548
Another way for you to look at that is if we have say, the multi core Intel.

104
00:14:14.548 --> 00:14:20.908
Let's see on and so they're all accessing a common memory.

105
00:14:21.894 --> 00:14:36.594
But if you force every reference to be written back to the local course cash into the common memory so that the memory view is consistent. That's very slow.

106
00:14:36.869 --> 00:14:40.769
Question from Tom that if we haven't taken chaos.

107
00:14:40.769 --> 00:14:44.129
Is there anything we should read about?

108
00:14:45.958 --> 00:14:52.139
Let me think about that 1. um, well, the stuff we're using from chaos will, um.

109
00:14:52.139 --> 00:14:58.109
Well, the big thing there is, you got this machine model of, um.

110
00:14:58.109 --> 00:15:04.438
An instruction and memory. Well, I'll talk about that in more detail later when, um.

111
00:15:04.438 --> 00:15:14.458
Actually later in this class, so okay, so, um, that's 1 thing I have in chaos also in Coco. Um.

112
00:15:14.458 --> 00:15:20.339
And it's also another reason I put chaos in a.

113
00:15:20.339 --> 00:15:24.328
There's a general cultural thing that you get more out of the.

114
00:15:24.328 --> 00:15:33.719
Higher level courses like this 1 if you have a broader background. So, if you know more general things about computer engineering, it's.

115
00:15:33.719 --> 00:15:37.229
They motivate, I guess you might say.

116
00:15:37.229 --> 00:15:41.969
The things otherwise you might take some of the things here. You don't understand the point of it.

117
00:15:41.969 --> 00:15:46.708
So, I, I guess that's the concept. So.

118
00:15:46.708 --> 00:15:52.828
Well, I think of a more specific antics here, but it's more the general. Um.

119
00:15:52.828 --> 00:15:57.418
Okay. Oh, good. General knowledge. Okay.

120
00:15:58.528 --> 00:16:09.208
And it's also fun it's fun to see how computer engineering developed and it's a very complex topic. And when you're working with a complex field of knowledge.

121
00:16:09.208 --> 00:16:12.869
You have to organize it and.

122
00:16:12.869 --> 00:16:23.399
So 1 thing that those courses, like, you know, features to teach you how this field is organized and they also teach you how.

123
00:16:23.399 --> 00:16:26.908
Difficult problems are solved.

124
00:16:26.908 --> 00:16:31.019
So, now there's a model for computing. I'll show you later.

125
00:16:31.019 --> 00:16:45.958
Um, and it's all BS logic unit you've got memory and so on and so on, you have a program calendar points to the next address and memory. You're gonna pull the next instruction out of you got some fast registers.

126
00:16:45.958 --> 00:16:57.178
That may look sort of obvious now, but it's, it wasn't obvious originally and some of those things were developed incrementally by trying lots of different ideas.

127
00:16:57.178 --> 00:17:00.568
And some.

128
00:17:00.568 --> 00:17:06.509
And then seeing what worked out, and you cannot always predict in advance.

129
00:17:06.509 --> 00:17:13.528
What will work out? You have to try it and see okay now, software, um.

130
00:17:13.528 --> 00:17:21.239
May do this thing random order? Um, the C plus plus competitor like fast is 1 from NVIDIA.

131
00:17:21.239 --> 00:17:27.269
Um, they bought it from a company PGI, I think, and.

132
00:17:27.269 --> 00:17:30.868
Well, 1st, it does parallel computing with videos.

133
00:17:30.868 --> 00:17:40.888
Hardware, and, um, it also is, I think, fairly up to date C. plus plus extension talking about the C. plus plus language I like this is the language I large to use.

134
00:17:40.888 --> 00:17:47.969
Let me tell you why C. plus plus has a conceptual unity to it.

135
00:17:47.969 --> 00:17:52.588
And so it's done by any stretch drop and it, um.

136
00:17:54.298 --> 00:17:58.469
So, things fit together, which makes it in a sense. Beautiful.

137
00:18:18.298 --> 00:18:23.398
Ecosystem around it of of toolkits, useful tools.

138
00:18:23.398 --> 00:18:30.838
The 3rd thing is, it produces code that runs very past and I like efficiency.

139
00:18:30.838 --> 00:18:35.249
And when you have, and in a sense, when you have larger.

140
00:18:35.249 --> 00:18:45.808
Problems efficiency can become more important because you're if you look at the acid product curve, you know, the X axis the size of problem with the Y, axis is time and.

141
00:18:45.808 --> 00:18:56.278
When your larger problems and better acid product performance becomes more important. So C plus plus it's really nice. So you can use good program.

142
00:18:56.278 --> 00:19:03.449
Techniques, for example, you could have little functions that encapsulate little ideas.

143
00:19:03.449 --> 00:19:11.338
And you can build them into the class structure so suppose you're working with both English and metric units to play.

144
00:19:11.338 --> 00:19:18.298
And some variables, like the English, and some variables might be metric that sounds like a totally crazy, logical system.

145
00:19:18.298 --> 00:19:22.229
Hello NASA, NASA lost the probe to Mars.

146
00:19:22.703 --> 00:19:31.374
Because, like, 1 contract are used English 1 used metric, and they didn't talk to each other properly and didn't convert and they lost the Mars. Bro. Okay. Well, and C.

147
00:19:31.374 --> 00:19:43.074
plus plus, you could have little functions that convert and you might be called billions of times. Well, they get inserted in line and optimized out so the function call ends up costing you nothing.

148
00:19:43.378 --> 00:19:49.078
That sort of thing you said, plus I don't think you can do it in Python, for example. So, and it.

149
00:19:49.078 --> 00:19:55.138
So, that's what I like about it and C plus plus is currently being updated every 3 years.

150
00:19:55.138 --> 00:19:58.739
So the latest standard is C plus plus 20.

151
00:19:58.739 --> 00:20:09.058
Um, the standard is out there, the most advanced compilers are starting to include some features for that. Previous 1 was C plus plus 17.

152
00:20:09.058 --> 00:20:20.364
Um, they were not as far as I remember, they were not major extensions there. There was not really plus 14, not really simple. Plus 11 brought in some major parallel computing ID.

153
00:20:20.364 --> 00:20:26.273
So the thing keeps getting updated and it gets instantiated and videos. Compiler. Um.

154
00:20:26.699 --> 00:20:30.929
Extension software we talked much then extending on top of that.

155
00:20:30.929 --> 00:20:36.989
For multi core over programming in general, there's something called open ACC.

156
00:20:36.989 --> 00:20:43.679
Look ahead of me for an open ACC dot Org and that's an extension to C. plus plus to help.

157
00:20:43.679 --> 00:20:51.269
If you have an algorithm, that's part of the hardest part being in this subject is to have algorithms. You told them that are parallel.

158
00:20:51.269 --> 00:20:57.388
And so if you have a loop, the different iterations of the loop are not interfering with each other. And.

159
00:20:57.388 --> 00:21:03.384
Um, 1, iteration is not creating not writing, just something in the next iteration will read, for example, then it gets hard to parallelize.

160
00:21:03.443 --> 00:21:04.314
So C,

161
00:21:04.523 --> 00:21:05.513
open ACC,

162
00:21:05.513 --> 00:21:06.834
our set of extensions,

163
00:21:06.834 --> 00:21:10.284
small extensions plus plus language,

164
00:21:10.614 --> 00:21:14.993
some practice into function calls to go in parallel with different back end,

165
00:21:15.624 --> 00:21:16.374
multi core,

166
00:21:16.374 --> 00:21:19.344
or the many multi core meeting to see on the mini core meeting.

167
00:21:19.949 --> 00:21:27.298
Typically video. Okay, I'll talk about the open ACC. Some odd.

168
00:21:27.624 --> 00:21:41.153
It's also a living thing, it's been extended there's another thing called open MTP I didn't put here. It's a slightly lower level thing that I used to teach in previous versions of the course. And that also adds primitives to your program.

169
00:21:41.574 --> 00:21:54.263
Seamless open MP MP for multi programming between open is to some extent replace in an open. Mp was very slow to work with Jeff a back end of the video.

170
00:21:54.598 --> 00:22:00.959
Another, um, parallel oh, another hardware thing I'll talk about.

171
00:22:00.959 --> 00:22:10.318
Um, I used to talk about, I don't know, um, Intel for some years, had a parallel thing called the, um.

172
00:22:10.318 --> 00:22:17.278
Color coat process of is he on 5 code processor? It was a board you'd plug into your, um.

173
00:22:17.278 --> 00:22:25.798
Into your until the machine that added, it would have a few 100 cores. The board had sort of Intel zones on it, but they were stripped it down.

174
00:22:25.798 --> 00:22:32.338
They did not have, um, pipelining and so on, but this was a way until had you have a few 100.

175
00:22:32.338 --> 00:22:37.288
Had a few 100 cores they brought the product line so that's why.

176
00:22:37.584 --> 00:22:51.144
I'm not teaching anymore. Okay, so we've got, um, the open ACC, another general parallel thing is called thrust, um, many of, you know, about the standard template library or all of you.

177
00:22:51.144 --> 00:22:54.473
Perhaps it has some paradigms and some standard.

178
00:22:54.689 --> 00:23:01.378
Ways to work with that. If the data looks like a vector, preferably vector some plain old data set.

179
00:23:01.378 --> 00:23:05.459
Plain old data would be in some floats and so on. So, um.

180
00:23:06.719 --> 00:23:10.888
Um, the standard template library gives you tools for.

181
00:23:10.888 --> 00:23:24.479
Doing a reduction over it and so on and working with that. So trust is a parallel extension of the. So we can work with vectors of data on a parallel machine. And, um.

182
00:23:25.679 --> 00:23:31.888
And we'll see some paradigms for parallel programming. So a paradigm is a.

183
00:23:49.588 --> 00:24:01.108
I think some of the back end stuff might be proprietary. I can't I don't jump past enough to know what's happening. They keep changing these things around, but the back end for thrust could be just a normal.

184
00:24:01.108 --> 00:24:06.898
Sequential machine, or it could be the many cores or the multi core.

185
00:24:06.898 --> 00:24:15.929
In video now, in theory, you change a compiler flag and re, compile your program. But, you know, the joke about.

186
00:24:15.929 --> 00:24:20.699
In theory, it's not difference in theory and practice on practice. There is a difference. Um.

187
00:24:20.699 --> 00:24:29.699
I find that the program sheet twiddling to work with different backend. But trust is cool. It's and to see their ideas and to see specific thing.

188
00:24:29.699 --> 00:24:36.118
Those are higher level things. I mentioned math lab and Mathematica, because they have.

189
00:24:36.118 --> 00:24:40.108
Some parallel extension so Matlab, for example.

190
00:24:40.108 --> 00:24:44.608
Their most common operations, like, for the matrix.

191
00:24:44.963 --> 00:24:59.933
You can do that in parallel. If Matlab lets be posted on a parallel machine. The Matlab, um, has licensing restrictions. Last I checked. No matter how many cores your machine had your host had.

192
00:25:00.028 --> 00:25:06.148
Bad, I would not use more than 8 of them at 1 point or something. Mathematica is, um.

193
00:25:06.148 --> 00:25:09.929
Most probably show mathematical later for people who aren't, um.

194
00:25:09.929 --> 00:25:18.088
Who are familiar with it? I don't know what's going on here. Um.

195
00:25:21.028 --> 00:25:31.019
Yeah, my screen, um, I got a newer message saying, attempted to start an invalid broadcast session.

196
00:25:31.019 --> 00:25:34.259
I don't know what's going on here. Um.

197
00:25:37.528 --> 00:25:47.249
Okay.

198
00:25:52.348 --> 00:25:59.969
We try and figure out what's going on here.

199
00:26:01.618 --> 00:26:13.199
Thanks very good. Um, the reason I'm broadcasting, uh, from the iPad is in case, I want to write stuff with good notes.

200
00:26:13.199 --> 00:26:21.239
And I haven't had an application on Linux, which is like, good notes. So, for 1 thing, my iPad is a touch screen.

201
00:26:21.239 --> 00:26:24.239
In this laptop does not, um.

202
00:26:24.239 --> 00:26:28.048
I'm going to try to broadcast from the other 1. um.

203
00:26:28.048 --> 00:26:37.078
I have a 2nd, I actually iPad the 2nd thing that here, which I could do just in case, but I don't want to start it. Oh, let me see what happens here.

204
00:26:44.999 --> 00:26:52.439
Okay, let's see.

205
00:27:01.618 --> 00:27:12.749
Okay, um, this is all I have 2 machines in front of me. Um, so what I just did is that I'm sharing from my Linux.

206
00:27:12.749 --> 00:27:18.659
Pad now oh, until that crash is then I'll go back to the iPad.

207
00:27:18.659 --> 00:27:24.298
I still have the chat window open.

208
00:27:25.439 --> 00:27:36.179
So, where it was, okay.

209
00:27:36.179 --> 00:27:39.628
Well, it was hardware, um.

210
00:27:42.509 --> 00:27:47.878
Yeah, very good. If you can tell me more later about how buggy iPads, and it says, workarounds.

211
00:27:47.878 --> 00:27:52.919
I don't know, I'm also not wedded to Webex if there's other tools out there, but it just.

212
00:27:52.919 --> 00:27:55.949
It works largely.

213
00:27:55.949 --> 00:28:05.159
Okay, so I told you about thrust, it's an extension to, um, standard template library. Um, 2nd, here.

214
00:28:05.159 --> 00:28:14.038
You're crashes no mind doesn't crash. Every time it gets. Let's make a deep into the lecture and then it stops working.

215
00:28:14.038 --> 00:28:18.838
Okay, I don't love computers. I'm, um.

216
00:28:18.838 --> 00:28:22.618
I like to joke that when I was an undergrad, um.

217
00:28:22.618 --> 00:28:26.068
I took more physics and anything else actually, but.

218
00:28:26.068 --> 00:28:40.499
I wasn't smart enough to do the theoretical physics and I was too klutzy to do the experimental physics. So I went into computer science because at that time, if you were in the same computer program twice, you got the same answer.

219
00:28:40.499 --> 00:28:43.919
Well, that was before networking time sharing.

220
00:28:43.919 --> 00:28:48.959
Now, we run the same thing twice. We don't get the same answer. Any case.

221
00:28:48.959 --> 00:28:53.189
So, the next tool I'll talk about somebody's CUDA.

222
00:28:53.189 --> 00:28:59.669
Could, uh, you might say is invidious assembly language programming.

223
00:28:59.669 --> 00:29:05.969
For their GP use and you're down at the low level you can and.

224
00:29:05.969 --> 00:29:16.078
You've got much more detailed control over the parallelism and so now, down at the, it's an interesting architecture.

225
00:29:16.078 --> 00:29:20.878
And we'll talk about it more later, but to give you a teaser, um.

226
00:29:20.878 --> 00:29:27.419
The way an Intel works is they throw billions of Gates some.

227
00:29:27.419 --> 00:29:34.528
Transistors at solving the problem of doing things in parallel and Z on on.

228
00:29:34.528 --> 00:29:41.638
If you, it's, it's got in 1 cycle can do about 5 different things. I think it can do in the future.

229
00:29:41.638 --> 00:29:51.989
An manager, operation, logic, Boolean test, it could also do a floating addition and also a floating modification.

230
00:29:51.989 --> 00:29:59.608
And do all those things in parallel with 1 cycle that's the Super scaler idea. And they got a lot of complicated logic architecture.

231
00:29:59.608 --> 00:30:02.969
Um, in the, um.

232
00:30:02.969 --> 00:30:15.898
On the, the NVIDIA GPU is a totally different philosophy. The CUDA course they're much simpler. In fact, for comp, they have separate cores. Um.

233
00:30:15.898 --> 00:30:24.148
The holding point would be a separate core, and it might take more than 1 cycle to do, say, a floating operation.

234
00:30:24.148 --> 00:30:31.588
so in one cycle they can do much less work but you might have several thousand 

235
00:30:31.588 --> 00:30:44.003
Several 1000, um, could, of course that can run in parallel and they've got 5 different levels of memory, which are down at the CUDA level are visible to you, um, each thread.

236
00:30:44.034 --> 00:30:51.384
Oh, another thing they do is, they bundled, uh, thread is like 1 execution sequence. Okay, thread has a, um.

237
00:30:51.689 --> 00:31:04.409
Instruction pointer that points to the program counter routing points to the next is the address in memory where you're going to get the next instruction you got to execute and.

238
00:31:04.409 --> 00:31:14.189
And the thread has, and it's been executed on a local code core, and it's got, it's got a view into the memory.

239
00:31:14.189 --> 00:31:27.568
And, but in video bundles, the threads up and 32 threads, formal war analogy from where you can clock, and all 32 threads have to execute the same instruction.

240
00:31:27.568 --> 00:31:31.259
Or they might be flagged to be idle for that instruction.

241
00:31:31.259 --> 00:31:34.858
So, what this means is that.

242
00:31:34.858 --> 00:31:41.999
You look at the architecture, the real estate on the invidia GPU they have, um.

243
00:31:41.999 --> 00:31:45.898
They just space on the.

244
00:31:45.898 --> 00:31:56.759
You know, on the Silicon that encodes the instruction. Well, that space is amortized over 32 cooler. Course. So it's a much smaller fraction.

245
00:31:56.759 --> 00:32:05.094
Of space is used for decoding instructions, and also the video doesn't have the level of pipeline and so on.

246
00:32:06.023 --> 00:32:15.084
So the video spends relatively more space on looks fast registers and stuff like that. And special purposes that are not on. Is he on, like, the, um.

247
00:32:15.388 --> 00:32:23.669
Text her processors and so on. So, right at the top level, the have it.

248
00:32:23.669 --> 00:32:33.328
Quite a different policy for how you spend your resources and the resources are square millimeters on the Silicon.

249
00:32:33.328 --> 00:32:38.249
So, you're down with a low, low level and what they do, they got like, 5 levels the memory.

250
00:32:38.249 --> 00:32:44.699
So, there, each thread has some small cash register space. It's 256 words.

251
00:32:44.699 --> 00:32:50.759
And then it's got some larger local space. It's very slow and some shared space. It's fast and.

252
00:32:50.759 --> 00:33:02.128
It goes up at 5 levels, and these different levels of memory have a 100 to 1 ratio and their speed. So, a good NVIDIA. Um.

253
00:33:02.128 --> 00:33:09.209
Card might have 24 gigabytes now of memory on the GPU card.

254
00:33:10.163 --> 00:33:20.364
But it's 100 times slower to get to, you can stream it very fast, but it's got a very large agency to start moving it. So, Jonathan was it CUDA you can work with stuff like that.

255
00:33:20.364 --> 00:33:32.963
Also, in video has, like, 3 different levels of hierarchy for the threads, local threads that grouped into orgs and then you could have a block of threads called thread. Block might have 1000 threads.

256
00:33:34.199 --> 00:33:38.308
And the different warps are scheduled asynchronously they could.

257
00:33:38.308 --> 00:33:46.078
We're in separately of each other, but they couldn't pick up their local memory and then the whole block has.

258
00:33:46.078 --> 00:33:51.239
A, a set of shared memory that's visible to everything in the block.

259
00:33:51.239 --> 00:33:56.368
See, that the local work a threat, but thread you got to work with here. I took off the thread block.

260
00:33:56.368 --> 00:34:02.189
And your GPU now can have a number of thread blocked, might've dozens of thread blocks and they run asynchronously.

261
00:34:02.189 --> 00:34:07.169
And they have nothing in common with each other, except they can both access the global memory.

262
00:34:07.169 --> 00:34:19.018
And that's and these together make a parallel program, which Nvidia their terminology changes could be called a kernel sometimes. And then the GPU can have several colonels running.

263
00:34:19.018 --> 00:34:25.349
And, yeah, the past is supercomputer costs.

264
00:34:25.349 --> 00:34:30.568
Okay, yeah, in a few minutes I'll talk about that. Yeah, so, um.

265
00:34:31.043 --> 00:34:45.563
You've got this hierarchy of threats and on the GPU is is an operating system. They don't actually call it that that coordinates all this. So these things there's queues of stuff waiting to run. So, in the block, There'll be a queue of, um.

266
00:34:45.869 --> 00:34:51.268
Threads that are waiting to run because the threat needs some resources before. Good run.

267
00:34:51.268 --> 00:34:57.239
There may be some data from the global memory, has to be brought into a cash or something and.

268
00:34:57.239 --> 00:35:01.259
And then the whole colonel was at the top level, parallel programming some.

269
00:35:01.259 --> 00:35:04.978
It's resources of everything, so I could, of course so there's a couple of stuff.

270
00:35:04.978 --> 00:35:16.108
Who is a stock waiting to run and with the CUDA, you can get access to some of that. So kudo there's some C plus plus functions and there's also some minor extensions to the, um.

271
00:35:17.159 --> 00:35:20.489
Minor extensions to the language.

272
00:35:21.599 --> 00:35:28.259
Okay, so that gets into relating to that question.

273
00:35:28.259 --> 00:35:32.489
From Alan, these techniques are used and things like.

274
00:35:32.489 --> 00:35:36.059
Look at the fastest machines here.

275
00:35:38.219 --> 00:35:47.068
Okay, so there's a list here top 500 dot Org of the of the past is known machines.

276
00:35:47.068 --> 00:35:52.559
I say known machines because everyone figures that, um.

277
00:35:52.559 --> 00:35:58.768
There are very fast machines out there that people there almost don't talk about the Chinese just.

278
00:35:58.768 --> 00:36:03.509
Recently announced too fast machines that no 1 had known about and.

279
00:36:03.509 --> 00:36:10.619
Also here a 3 letter agency, you might perhaps in a very fast machine that no 1 talks about.

280
00:36:10.619 --> 00:36:20.278
And if you're an oil company in Houston, you may have passed Carl machines, you don't talk about. So it's not just classified purposes. So, in any case.

281
00:36:20.278 --> 00:36:23.818
So, if we look at some of them, um.

282
00:36:23.818 --> 00:36:30.389
Say, number 2, American machines, um.

283
00:36:32.429 --> 00:36:39.208
Okay, let's look at Summit. Let's say it's a faster. Okay, so the question was about the economy.

284
00:36:39.208 --> 00:36:50.248
Of these things add, economics are how that is funded and used. Okay. So the national labs in the United States, um.

285
00:36:50.248 --> 00:36:54.869
There's several.

286
00:36:54.869 --> 00:36:59.969
Um, um, look 1 more.

287
00:36:59.969 --> 00:37:04.648
Um, they started out in World War 2.

288
00:37:04.648 --> 00:37:08.998
To, uh, basically to test the atomic bombs.

289
00:37:08.998 --> 00:37:14.398
And so Oakridge was a.

290
00:37:16.079 --> 00:37:20.369
Was you're at 1 of the uranium verification?

291
00:37:20.369 --> 00:37:23.938
Techniques, um, so what the U. S get.

292
00:37:23.938 --> 00:37:27.028
And this is an interesting engineering thing.

293
00:37:27.028 --> 00:37:36.659
Is the U. S through money at the project in order to save time? So it wasn't known which.

294
00:37:36.659 --> 00:37:39.929
What was the good way to, um.

295
00:37:41.458 --> 00:37:44.878
To purify, separate.

296
00:37:44.878 --> 00:37:50.518
Which is like, point 7% of the year anymore. So they tried a couple of methods in parallel.

297
00:37:50.518 --> 00:37:54.659
It wasn't known whether uranium or plutonium bomb would work.

298
00:37:54.659 --> 00:37:59.548
So both were devised in parallel and these were the different Labs.

299
00:37:59.548 --> 00:38:05.668
And it turned out that almost every almost everything that was tried and ended up working.

300
00:38:05.668 --> 00:38:10.619
Well, I'm testing using the same ideas, for example, with space X.

301
00:38:10.619 --> 00:38:15.480
Throw money at the problem and try several things in parallel. Okay.

302
00:38:15.480 --> 00:38:23.550
So so, after the war, that was the cold war, they're working on, if any H bomb and so on and then they move to civilian uses.

303
00:38:23.550 --> 00:38:28.139
And so the new general, Super computing now, and.

304
00:38:28.139 --> 00:38:34.949
For solving other problems, we'll take, we'll talk about applications, so.

305
00:38:34.949 --> 00:38:42.449
They're funded by the Department of energy and the energy Labs Department of energy. So that's where the money comes.

306
00:38:42.449 --> 00:38:51.989
No, basically for these IBM, they had the blue jeans, for example, I knew were machines here and.

307
00:38:51.989 --> 00:39:04.650
So, IBM prompts the money in order to sell it to customers and customers that want very big things would be for example, I mentioned the oil companies that are modeling. We.

308
00:39:04.650 --> 00:39:08.639
Technology of the rocks to try to predict oil.

309
00:39:08.639 --> 00:39:15.659
Flight Simulator is mobile, so that's that's who pays for it.

310
00:39:15.659 --> 00:39:20.610
And I are the machine from the IBM, um.

311
00:39:20.610 --> 00:39:29.579
I think our gets some, you'd have to talk to Professor kruthers but, um, those are companies that buy time on the, um.

312
00:39:29.579 --> 00:39:32.940
On our computer for all. I know that might be, you know.

313
00:39:32.940 --> 00:39:40.320
A large amount of the income, so that's the how it's funded.

314
00:39:40.320 --> 00:39:45.599
For used now the energy Labs, you can apply to use them for free.

315
00:39:45.599 --> 00:39:49.559
So these are there, there's various.

316
00:39:49.559 --> 00:39:54.750
Government funded natural science foundation also enough to start of energy.

317
00:39:54.750 --> 00:39:58.019
Parallel machine Super machines out there in the country.

318
00:39:58.019 --> 00:40:01.349
That you can apply and.

319
00:40:01.349 --> 00:40:11.070
If you're in an application to say what you want to use it for, and if they like your ideas, they will give you an allocation of time for free.

320
00:40:11.070 --> 00:40:17.940
And it will say Mark guys thing. No, you can just you go out some time on the.

321
00:40:17.940 --> 00:40:23.610
Computer yeah, you so there's an application form on their website. I need to.

322
00:40:23.610 --> 00:40:35.849
Say, here's the time. I know I know a large amount of time. There was 1 site. I looked at, um, their initial they would give any 150,000 CPU hours.

323
00:40:35.849 --> 00:40:38.909
Just for asking, so this is, you know.

324
00:40:38.909 --> 00:40:50.190
250,000 hours was so small. It wasn't. You sort of like, hey, you've heard of parallel computing. You think you'd like to learn it is 50,000 hours.

325
00:40:50.190 --> 00:40:54.210
And if you want more than apply for it okay, so.

326
00:40:54.210 --> 00:41:02.730
So we look at the architecture for the summit here so it's from IBM. So you have a smaller version. This is they did a big version here.

327
00:41:02.730 --> 00:41:09.900
Um, look at the speed, 148 flops. It's a nice speed here.

328
00:41:09.900 --> 00:41:16.619
The benchmarks benchmarks for a lot of these things are very specific. They are, um.

329
00:41:16.619 --> 00:41:20.099
Typically matrix multiplication.

330
00:41:20.099 --> 00:41:33.719
Or it breaks down to doing a thought products of of vectors cause that's what makes qualification. There's lots of dark products of rows and columns. So it's a very specific benchmark. And in fact, hardware people can game it.

331
00:41:33.719 --> 00:41:36.719
Laptop is still considered to be useful.

332
00:41:36.719 --> 00:41:41.219
So, if you look at this, this structure here, so 4,000, no, it's it.

333
00:41:41.219 --> 00:41:45.239
The power 9 is an IBM is it's a very nice architecture.

334
00:41:45.239 --> 00:41:50.610
And it's IBM competitor to the Intel you might say.

335
00:41:50.610 --> 00:41:54.329
Well, as soon aside the until Zee on, um.

336
00:41:54.329 --> 00:42:02.070
That's not your standard Intel. It's on my laptop here is here, but it's a very good laptop. Um, your typical.

337
00:42:02.070 --> 00:42:08.130
Consumer machines don't have to the step below it. Well, to see on adds to the.

338
00:42:08.130 --> 00:42:14.429
Yeah, the core whatever machines it adds some hardware support for doing.

339
00:42:14.429 --> 00:42:25.380
Multi programming and virtual program, and more than the lower level of machines do it that support for more memory and stuff like that. Okay so we look at how the summit is organized here.

340
00:42:25.380 --> 00:42:30.329
So, 4,000 nodes with each node is a board with 2 power. 9.

341
00:42:30.329 --> 00:42:34.860
Is it power? I'm very nice architecture and they're multi core. Okay.

342
00:42:34.860 --> 00:42:40.469
And the power line has, um, 22 cores.

343
00:42:40.469 --> 00:42:44.820
And then 6 video, um.

344
00:42:44.820 --> 00:42:49.320
Um, now.

345
00:42:49.320 --> 00:42:55.679
Some more details here and video goes through generations. Um.

346
00:42:56.880 --> 00:43:00.570
The Tesla is like an architecture.

347
00:43:00.570 --> 00:43:03.659
Level, so, um.

348
00:43:03.659 --> 00:43:18.655
The lowest level for a video are the cheapest ones that are used for gaming and above the slide, there's the Quadro line, which has support for scientific computing. It will do.

349
00:43:19.914 --> 00:43:34.614
Say standard floating point um, I triple standard floating points so routing to be done correctly and so on. And that takes hardware to do that right? Floating point. So, that's what you get from the video going from the G force.

350
00:43:34.829 --> 00:43:42.989
To the Quadro, although the exact capital simple overlap and you go from the quarter up to the Tesla.

351
00:43:42.989 --> 00:43:48.539
The test the line that adds more support for the high end computing.

352
00:43:48.744 --> 00:44:03.655
And a Tesla JP might not have any video at all. I mean, these things here on the subject is gonna be no video I'm guessing at all. So, test is super peutic thing, but orthogonal to the G, fours quadrant test all the results on mobile live.

353
00:44:03.775 --> 00:44:04.824
But you're not talking about.

354
00:44:05.070 --> 00:44:18.150
My Avenue with a couple of times cars by Audi cars. Um, 2 in the in it, I think it's got 2 displays. Okay so orthogonal to that. Then there's things like the, um.

355
00:44:18.150 --> 00:44:24.300
Various slide, um, V means full time. So, as the years go on, they've got.

356
00:44:24.864 --> 00:44:39.775
For me, caterpillar, Maxwell, Pascal Volt and err and every, like, 2 year or 2 years that come up with another generation. And that's what the V stands for. And the 100 is just within the wholesale line.

357
00:44:41.880 --> 00:44:46.050
Talks about the size of the GPU, the number of credit cards and so on.

358
00:44:46.050 --> 00:44:51.809
Okay, and the hierarchy inside the course something called a streaming multi processor.

359
00:44:51.809 --> 00:44:58.050
Inside it and an extremely multi processor. Well, have hundreds of.

360
00:44:58.050 --> 00:45:03.690
For 50 or whatever could, of course, there's just several level hierarchy that she was talking about that.

361
00:45:03.690 --> 00:45:07.920
And.

362
00:45:07.920 --> 00:45:13.320
So, I can't remember the V100 it's got a I have a few 1000 who the cars.

363
00:45:13.320 --> 00:45:16.500
So you're gonna have a few 1000 times 6.

364
00:45:16.500 --> 00:45:19.559
It was 2 times 4,000.

365
00:45:19.559 --> 00:45:25.170
That is going to be up at a nice number of code, of course, and a very fast bandwidth.

366
00:45:25.170 --> 00:45:29.969
That's not the organization works and the thing is going to be happening. Um.

367
00:45:29.969 --> 00:45:34.320
Power consumption is, um, and the megawatt range, so.

368
00:45:35.670 --> 00:45:39.329
Professor Carlos about his electricity bill sometimes.

369
00:45:39.329 --> 00:45:45.449
And then in the summer, of course, he has to pay twice because he has to pay for the electricity.

370
00:45:45.449 --> 00:45:48.900
That's generating megawatts of heat then he has to pay for the.

371
00:45:48.900 --> 00:45:53.849
Electricity to run the air conditioner to take all that out. So he has to pay twice for the electricity.

372
00:45:53.849 --> 00:45:57.750
He's got a nice I'm guessing it's power bills millions of dollars here.

373
00:45:57.750 --> 00:46:01.170
That's a little bit how the real world things work.

374
00:46:04.230 --> 00:46:12.449
Okay, and so the top 500 list, 205 and 6 on the top few are.

375
00:46:14.039 --> 00:46:22.320
And is so, um, oh, 1, more thing a top 100 list if we look at the software, the operating system, they're running.

376
00:46:22.320 --> 00:46:27.389
They're all running some version of Linux or whatever.

377
00:46:27.389 --> 00:46:31.829
I used to be able to say that there was 1 machine running. Um.

378
00:46:31.829 --> 00:46:36.750
Wanting Microsoft Windows, but then it stopped.

379
00:46:36.750 --> 00:46:41.460
There was cut off at the top 500 so there are 0 machines running. Um.

380
00:46:41.664 --> 00:46:53.275
Windows on the list, including the machine at Microsoft owns. So, you know, Microsoft is starting to adopt Linux a little, you know, they've got Windows top system for Linux, um, under Windows.

381
00:46:53.275 --> 00:46:58.885
Now it's a virtual machine and that runs Linux on um.

382
00:47:00.869 --> 00:47:06.000
It runs Linux under Windows and I played with it a little.

383
00:47:06.000 --> 00:47:09.239
It's, it's a good Linux. It's virtualized at a low level.

384
00:47:09.239 --> 00:47:16.320
It's not just a PowerShell layer on top. Like they used to have this is running Linux at a low level. So I played with it so much.

385
00:47:16.320 --> 00:47:20.250
Oh, you like a Tom? Okay. Great.

386
00:47:20.250 --> 00:47:26.639
I thought about putting it on my latest laptop and then I got a little frightened that it was too fragile. I could not go on it.

387
00:47:26.639 --> 00:47:29.820
Type a wrong command and delete Michael off.

388
00:47:29.820 --> 00:47:35.880
System, so I went back to dual booting, which means I have to do over the top.

389
00:47:35.880 --> 00:47:40.199
1 of my machine, I've also got a licensed version of VMware.

390
00:47:40.199 --> 00:47:43.260
On my big machine here and that.

391
00:47:45.750 --> 00:47:56.579
And that, um, you know, right don't use a gooey. Oh, right. I have also another complaint I had was.

392
00:47:58.019 --> 00:48:07.469
Well, I, I couldn't mount an external partition inside it. It was really weird like my desk at different partitions and, um.

393
00:48:08.820 --> 00:48:14.519
For some reason, and of course, I'm using CFS, it's a file system. I like the best that I could not.

394
00:48:14.519 --> 00:48:19.980
So, it looked like a, I don't I wouldn't do it for any more. Okay.

395
00:48:19.980 --> 00:48:25.530
Well, let me just show you my, what my current system is actually, um.

396
00:48:27.090 --> 00:48:39.719
Let's see here. So this is a laptop now it's dual 6 core so it's running. Um, it's 12 cores here.

397
00:48:39.719 --> 00:48:42.900
Look at the memory.

398
00:48:42.900 --> 00:48:49.170
This machine has has 128 gigabytes of.

399
00:48:49.170 --> 00:49:02.369
I like memory, I don't swap he'll swap space because swap swapping is obsolete. It takes too much time to swap. It would take seconds to swap that amount of physical memory Internet once. So.

400
00:49:02.369 --> 00:49:11.489
Okay, you can see what we're looking at. I'm sorry I'm.

401
00:49:14.969 --> 00:49:26.849
Just a 2nd, um, it works better if I share applications, not the entire screen, but that requires.

402
00:49:29.550 --> 00:49:33.059
On okay.

403
00:49:33.059 --> 00:49:36.989
Um, that's running H, talk.

404
00:49:36.989 --> 00:49:50.429
Yeah, I, I can see it now. Okay, so this is a laptop. Now I've got 6. it's a Z on 6 core. Is he type for theaters? That's 12 threads.

405
00:49:50.429 --> 00:49:59.340
And look over here, 12,125 using binary 100 by the kernel used by the.

406
00:49:59.340 --> 00:50:03.000
I also, so it's 128 gigabytes of.

407
00:50:03.000 --> 00:50:06.000
Acc ACC.

408
00:50:06.000 --> 00:50:10.829
You know, so this is a very nice a laptop, I think.

409
00:50:10.829 --> 00:50:16.139
Okay, all the Chromium tabs.

410
00:50:16.139 --> 00:50:22.380
Well, that's the security thing in browsers. Now they run on each, they have many separate processes.

411
00:50:22.380 --> 00:50:30.480
And the hope is, in a sense, have walled gardens around each tab on your browser. And I think it works to some extent.

412
00:50:30.480 --> 00:50:36.809
Okay, just a 2nd, here back to the browser I I know it stopped.

413
00:50:54.655 --> 00:51:00.985
Okay, this appears to work for live, but if I share separate applications and a picture of the whole screen.

414
00:51:01.230 --> 00:51:12.239
Okay reference it can a little about me. I got my bachelors at Toronto, master's and PhD at Harvard.

415
00:51:12.239 --> 00:51:15.750
Technically, my Harvard PhD says applied bath.

416
00:51:15.750 --> 00:51:18.900
I diplomacy as a podcast, but it's really, um.

417
00:51:18.900 --> 00:51:25.710
In the computer science, so, and you don't want to know how long I've been programming.

418
00:51:25.710 --> 00:51:32.820
Okay, emails the best office out of our virtual to online for office hours.

419
00:51:32.820 --> 00:51:38.369
Basically, it holding me an email. I have many different email addresses.

420
00:51:38.369 --> 00:51:52.980
Address there and I have I've had my own domain now for 20 years or something. No girlfriend and her daughter. I also W, Randall Franklin go to work. I also, if you would like to send an encrypted email.

421
00:51:52.980 --> 00:52:01.739
Um, I have also a proton mail account over effective P. M. E. and we could also so proton mail set payments with.

422
00:52:01.739 --> 00:52:08.550
Encrypted service, and also we can exchange public.

423
00:52:08.550 --> 00:52:15.090
Keys if you'd like, and you can send me if you want, we could set up a new and send encrypted email. So.

424
00:52:15.090 --> 00:52:19.530
Um, I generally send from non RPI accounts.

425
00:52:19.530 --> 00:52:32.610
Cause I just gets so tedious to set up our all the time and you're welcome to send. And in fact, if you give me an RPA account, I'll add it to my list. So you'll get emails. I sent.

426
00:52:32.610 --> 00:52:46.469
Mostly, I, if I'm asking the whole class, I'll send it from my account probably, because otherwise will tag it as spam for the recipients on our API. The spammed it um.

427
00:52:46.469 --> 00:52:53.579
Spam assassin has its problems I mean, at 1 point it crap to an individual message from the provost to me.

428
00:52:53.579 --> 00:52:57.960
Not a multiple into it or message to her. It's it's critical what it does so.

429
00:52:57.960 --> 00:53:09.269
I mean, I check it spam filter every few days and that's all I know. But so you're welcome if you send me an email from some other account, it's really nice if your header makes it obvious who you are.

430
00:53:09.269 --> 00:53:17.789
And don't say Buffy the Vampire Slayer or something and if you tag it with the hashtag.

431
00:53:17.789 --> 00:53:26.130
And we can use Webex face time or whatever talking if you'd like, um.

432
00:53:26.130 --> 00:53:32.280
I deprecate FaceTime with required who use? Webex? Um, I have, um.

433
00:53:32.280 --> 00:53:42.150
My own domain here you see the blog, um, the nice quick way to, um, get to it is just Google.

434
00:53:44.039 --> 00:53:47.550
There it is. Okay. Um.

435
00:53:49.920 --> 00:53:57.840
No, no.

436
00:53:57.840 --> 00:54:06.960
There we go. Okay, office hours. I'm teaching probabilities starting at 2, so I don't have a lot of time after my parallel glass.

437
00:54:06.960 --> 00:54:16.469
But after probability we can talk as long as anyone wants to talk informal meetings, I'm willing to talk to people. If it's any talk. If it's legal and ethical, I'll just have to talk about it.

438
00:54:16.469 --> 00:54:21.510
Professor like talking, so, um, even after you leave.

439
00:54:21.510 --> 00:54:29.940
I'm teaching the course I created this course a few some years ago, so I'd like the topics. That's why I'm teaching it yet.

440
00:54:29.940 --> 00:54:34.889
I used to teach computer graphics. There's a slight overlap because I got confused and both of them.

441
00:54:34.889 --> 00:54:42.239
I got the webpage here, reading material. There's nothing that you have to buy.

442
00:54:42.239 --> 00:54:47.429
There's a few things that are on, you you welcome to buy.

443
00:54:47.429 --> 00:54:52.409
If you wish you don't have to, and I'll use a little from some of that.

444
00:54:52.409 --> 00:54:57.719
Now, there's a problem here when the field is changing. 1st really good books.

445
00:54:57.719 --> 00:55:02.699
Standards book, it's they, they have an updated it lately so there's lots of examples.

446
00:55:02.699 --> 00:55:06.449
But they haven't updated it. Another thing is that.

447
00:55:06.449 --> 00:55:17.250
Could have changing and even in hideous example programs don't catch up. Don't keep up with it. Now. Why it is relevant.

448
00:55:17.250 --> 00:55:20.940
Can I explain BCC? Okay. Yeah, I just saw that. Um.

449
00:55:20.940 --> 00:55:24.719
In a minute, um, okay.

450
00:55:24.719 --> 00:55:29.610
They're adding again is the video folks.

451
00:55:29.610 --> 00:55:35.489
There's a reason the company is such a high market cap stock is so high, is that they're doing good things.

452
00:55:35.489 --> 00:55:43.800
They're adding value, so they look at how their stuff is used and they look at the problems and they try to help.

453
00:55:43.800 --> 00:55:46.800
So you have a parallel.

454
00:55:46.800 --> 00:55:50.039
Program, so you have your, um.

455
00:55:51.059 --> 00:55:58.889
Okay um, okay, let me talk about this. So I'll come back to.

456
00:55:58.889 --> 00:56:05.610
Bcc, is there a correcting memory? Um, it's not just parity checking. It can correct and error.

457
00:56:05.610 --> 00:56:10.110
And it's just in case, you know, 128 gigabytes of memory. That's a lot of Gates.

458
00:56:10.110 --> 00:56:14.550
I can adapt with cause mccray's and so on then.

459
00:56:14.550 --> 00:56:18.449
And the flip a bed, it's nice that the fact that it doesn't.

460
00:56:18.449 --> 00:56:21.840
My file system or something, so.

461
00:56:21.840 --> 00:56:27.210
Why high amount or 2 reasons it's more necessary.

462
00:56:27.210 --> 00:56:41.454
When you have a lot of D round, cause the chance of an error is large, I think most most errors are not going to cause you're not even going to know it. It tends to be real, you know, computers are not that you flip a few random bits. Probably nothing will happen. Really? Okay.

463
00:56:41.545 --> 00:56:44.605
And, um, if something happens, it'll probably be minor.

464
00:56:44.880 --> 00:56:51.239
But still, you know, I like the liability on my file system. I mostly use DFS.

465
00:56:51.239 --> 00:56:56.369
Error correcting and checks off and everything and it also.

466
00:56:56.369 --> 00:57:03.090
In fact, you CFS Warren's that the greatest vulnerability of CFS now is, um.

467
00:57:04.230 --> 00:57:10.590
Took bets on the seat. Okay. The 2nd thing is if people large amounts of direct their bigger projects.

468
00:57:10.590 --> 00:57:18.750
So, they're not hobbyist trying to get a machine for 500 dollars. They've got more money so they can afford to pay for it.

469
00:57:19.800 --> 00:57:25.050
And cause it does cost money and the other thing is, I mentioned, I got the Z on on this laptop.

470
00:57:25.050 --> 00:57:28.980
And the lower end, like the core of view.

471
00:57:28.980 --> 00:57:32.880
I went up to 911, I guess I lose track of it.

472
00:57:32.880 --> 00:57:36.539
They cannot use memory, so.

473
00:57:36.539 --> 00:57:42.210
So, getting into ACC, it's part of the next level of spending money.

474
00:57:42.210 --> 00:57:50.280
So actually, my laptop here, when I configured it with an old, it's custom, it's not start.

475
00:57:50.280 --> 00:57:56.099
I got a call from their fraud person, you know, such a fancy machine. So.

476
00:57:56.099 --> 00:58:06.869
Okay, okay. I'm getting back now. It's an important point of problem here with so, in video they're adding stuff.

477
00:58:06.869 --> 00:58:21.420
To, um, they're adding stuff to their kudo. Okay. A big thing is, I'm talking about various well, I've heard of so you got your CPO and you got your GPU.

478
00:58:21.420 --> 00:58:25.800
And they've got separate memories. Okay. So.

479
00:58:25.800 --> 00:58:34.829
Another thing I'll show you here is I also have an on this laptop. It's not none of this integrated.

480
00:58:34.829 --> 00:58:37.920
Graphics thing, or it's, you know, part of Intel.

481
00:58:37.920 --> 00:58:41.159
It it has a separate.

482
00:58:41.159 --> 00:58:44.400
A separate NVIDIA.

483
00:58:44.400 --> 00:58:47.670
Let me show you let me.

484
00:58:47.670 --> 00:58:51.719
Up again, here just a 2nd, um.

485
00:59:00.690 --> 00:59:11.489
Okay, should thank you. So, is it in video.

486
00:59:11.489 --> 00:59:17.219
To the, um, here to Quadro 5,000.

487
00:59:17.219 --> 00:59:22.860
And there are 16, um, gigabytes.

488
00:59:22.860 --> 00:59:25.889
On the GPU.

489
00:59:25.889 --> 00:59:32.130
I like the laptop, so the Intel ziam is 12,825.

490
00:59:32.130 --> 00:59:38.849
And then the GPU with 16 gigabytes and so these are 2 separate memory spaces.

491
00:59:38.849 --> 00:59:47.010
And what you would have to do in the past is explicitly coffee data back and forth. So you're going to create a program.

492
00:59:47.010 --> 00:59:51.599
You want to do something on GPU 1st you have to copy the data to the G.

493
00:59:51.599 --> 00:59:56.190
And then you run your job, and then you copy of the result back.

494
00:59:56.190 --> 00:59:59.190
It's very tedious all this copy back and forth.

495
01:00:00.420 --> 01:00:04.739
Higher level crosstalk make it easy. So so, if the copy so now.

496
01:00:04.739 --> 01:00:12.869
What video now goes they've got virtual memory that is a unified address space with paging.

497
01:00:12.869 --> 01:00:15.989
That combines the.

498
01:00:15.989 --> 01:00:20.369
So, if you have an array on the heat on.

499
01:00:20.369 --> 01:00:26.309
Then you can refer to it. 1st, it's a unified address phase.

500
01:00:26.309 --> 01:00:30.780
It's a 4,949 bid virtual address.

501
01:00:30.780 --> 01:00:39.389
And that was the 1st step the 2nd step is a demand paging. So, if your GPU program.

502
01:00:39.389 --> 01:00:43.800
Refers to an address that is physically on the.

503
01:00:43.800 --> 01:00:46.800
The GPU pages it over to the GPU.

504
01:00:46.800 --> 01:00:55.199
And this is just new, it's a couple of years old. Oh, by the way, the hardest thing to get this implemented, I believe was Linux.

505
01:00:55.199 --> 01:01:00.090
Because this requires support in the Linux colonel.

506
01:01:00.090 --> 01:01:05.639
And it took years before the Linux colonel added the support and released it. I think.

507
01:01:05.639 --> 01:01:09.119
They were possibly worried that if it was done.

508
01:01:09.119 --> 01:01:12.570
If it had a bug in, it would really screw over the users.

509
01:01:12.570 --> 01:01:18.659
But so there is this, um, support analytics colonel for the on demand paging.

510
01:01:18.659 --> 01:01:22.320
Okay, so this makes writing.

511
01:01:22.320 --> 01:01:25.860
A graph a GPU program a lot easier.

512
01:01:25.860 --> 01:01:35.250
However, since this is new, a lot of your examples don't use it so you look at your examples could've by example.

513
01:01:35.250 --> 01:01:44.639
They do not the examples. Don't use this tool. So their examples I teach you all sleep technique. You don't have to worry who's any more probably.

514
01:01:44.639 --> 01:01:48.119
So, that's I actually just incentive to use the.

515
01:01:48.119 --> 01:01:52.230
Some of those books.

516
01:01:52.230 --> 01:01:58.170
Show you 1 other thing before I go back, um, let's see.

517
01:02:00.599 --> 01:02:04.769
So this is a lower so this just shows you, uh.

518
01:02:04.769 --> 01:02:12.239
What is this again? My laptop. Okay. Cuda has different versions of 11 point something.

519
01:02:12.239 --> 01:02:18.269
And it's got another type of every time they add a new function or something above the version.

520
01:02:18.269 --> 01:02:22.050
And the capability tells in another sense.

521
01:02:22.050 --> 01:02:25.320
What is available in the coda?

522
01:02:25.320 --> 01:02:29.039
Um, 16 gigabytes of memory.

523
01:02:29.039 --> 01:02:32.940
The laptop is 3000 could, of course.

524
01:02:32.940 --> 01:02:37.679
And so the multi processes, those are the streaming multi processors.

525
01:02:37.679 --> 01:02:41.369
At 48 of them each 1 and 60 could of course.

526
01:02:41.369 --> 01:02:47.369
Um, it has a cash it's 4 megabytes. Um.

527
01:02:47.369 --> 01:02:54.659
Ignore the texture um, so these are architecture decisions here. So if you're at NVIDIA.

528
01:02:54.659 --> 01:03:00.960
And you're inventing the next level, you've got to decide how much memory global remember you, you're going to have.

529
01:03:00.960 --> 01:03:05.190
Um, it's available to all of the processors.

530
01:03:05.190 --> 01:03:15.989
And how many so, the multi processors, how many each multi process? It's true. It also called the streaming multi processes. Um, is a big area on the chip.

531
01:03:15.989 --> 01:03:19.860
And how many of these obviously decide the clock rate.

532
01:03:19.860 --> 01:03:23.010
The cache is passed, but expensive.

533
01:03:23.010 --> 01:03:27.840
They take a piece of the cache and they have.

534
01:03:27.840 --> 01:03:33.480
A cons a read, only constant memory it's implemented as a cash view into the global memory.

535
01:03:33.480 --> 01:03:36.809
So, it's small, but it's the thing is, it's.

536
01:03:36.809 --> 01:03:40.349
It's it's shared by all of your threads.

537
01:03:41.579 --> 01:03:50.789
And then I mentioned the, your current on your parallel program is divided into independent blocks inside each block. There's some past shared memory.

538
01:04:07.170 --> 01:04:13.679
There are 65,000 registers in the block, the block and I've talked to a 1000 threads, 70, 60.

539
01:04:13.679 --> 01:04:17.909
564 registers per thread, but if you've got 2 or 3 ID script.

540
01:04:17.909 --> 01:04:21.119
There I kind of more registers, um.

541
01:04:21.119 --> 01:04:30.269
So you may your program may run past if you use your threads, but they need to spread it's more fast registered space because it doesn't have to spill over to the flow.

542
01:04:30.269 --> 01:04:39.059
Slower global memory and maximum. This is 32 threads all run the same instructions that.

543
01:04:39.059 --> 01:04:42.329
Um, so they have different program calendars.

544
01:04:42.329 --> 01:04:46.409
But, no, they have the same program counter.

545
01:04:46.409 --> 01:04:50.250
Okay, but different local memory. So they're running the same.

546
01:04:50.250 --> 01:04:53.400
The program memory is the same for all the threat to the war.

547
01:04:53.400 --> 01:04:56.969
They're running the same instruction, but their access in different memory.

548
01:04:56.969 --> 01:05:00.000
Remove on.

549
01:05:00.000 --> 01:05:03.300
And that's basically that.

550
01:05:03.300 --> 01:05:08.400
Unified address say is this carbon address base.

551
01:05:08.400 --> 01:05:12.059
Oh, the GPU has no, um.

552
01:05:12.059 --> 01:05:19.800
Unified memory is the common 49 bit address space for the host. That's the Intel Z on the device.

553
01:05:19.800 --> 01:05:23.159
That's the video. Okay. Um.

554
01:05:23.159 --> 01:05:27.929
And you can look at the high order bits and know which it is.

555
01:05:27.929 --> 01:05:32.039
The managed memory is the, um.

556
01:05:32.039 --> 01:05:37.050
Is the on demand paging between the host and device? This is the thing that's fairly new.

557
01:05:37.050 --> 01:05:42.210
A couple of years old, 2, years old or something, and other stuff. Um.

558
01:05:43.619 --> 01:05:47.909
I'm not completely certain what this is here. Okay.

559
01:05:47.909 --> 01:05:52.349
Colonels can start on the kernels and so on.

560
01:05:52.349 --> 01:05:55.500
Okay, back to the website.

561
01:05:57.480 --> 01:06:02.219
Here and again I'm sharing applications.

562
01:06:02.219 --> 01:06:05.369
Not the whole screen because it's more reliable.

563
01:06:07.469 --> 01:06:11.519
Give it a 2nd, come on.

564
01:06:11.519 --> 01:06:21.150
Gotcha. Okay so I'll go point to free material on the web here.

565
01:06:21.150 --> 01:06:24.599
Oh, the computer systems I'm using.

566
01:06:24.599 --> 01:06:36.300
What I'll talk on next is a good introduction to parallel computing. Is this free online tutorial and it's some Lawrence Livermore National Labs that.

567
01:06:36.300 --> 01:06:39.840
Again, Lawrence, Livermore, they started out making.

568
01:06:39.840 --> 01:06:46.230
Very energetic devices. Very, very, very energetic devices. Um.

569
01:06:46.230 --> 01:06:52.139
Let me tell you a sign I tell you stories, but real world engineering issues.

570
01:06:53.340 --> 01:06:59.190
You have programming mistakes. Okay. You write a computer program.

571
01:06:59.190 --> 01:07:03.059
That screws up, um, does something you did not expect um.

572
01:07:03.059 --> 01:07:09.300
Tell you a story about how 11 day, and I think a 154 when the physicists.

573
01:07:09.300 --> 01:07:13.469
Screwed up. They were designing an H bomb.

574
01:07:13.469 --> 01:07:21.150
And the name was castle Bravo, and they designed the H bomb to be about 5, Mega tons.

575
01:07:22.559 --> 01:07:25.920
They made it easy change. See little.

576
01:07:25.920 --> 01:07:34.800
Physics error and their computation. Oh, who hasn't made in little physics errors and they.

577
01:07:34.800 --> 01:07:38.699
Ignored a particular reaction pathway.

578
01:07:38.699 --> 01:07:42.210
So, there fine make a ton bomb.

579
01:07:42.210 --> 01:07:48.570
Turned out to be 15, Mega times. It ended up to be 2 and a half times larger.

580
01:07:48.570 --> 01:07:53.429
Then they predicted, so, this was the largest H ball in the U. S. that were exploded.

581
01:07:53.429 --> 01:07:57.000
And, um, can show you in fact, um.

582
01:08:06.780 --> 01:08:15.900
Okay. Um, okay, so 15 Mega intended to be like, 5.

583
01:08:15.900 --> 01:08:23.609
5 or 6 turned out to be 15 because of a minor little physics error. So this is just make errors too.

584
01:08:23.609 --> 01:08:37.050
As a result of this error, um, the experimenter's on the island survive, because they were in an over design bunker, and they were trapped for a while, but it did irradiates the crewman on a Japanese fishing boat downwind.

585
01:08:37.050 --> 01:08:43.739
Um, and I think killed some of them integrated some islanders.

586
01:08:43.739 --> 01:08:48.689
And actually led to.

587
01:08:48.689 --> 01:08:56.579
Godzilla, so this idea of a, you know, big bomb cause, you know, waking up creatures and so on.

588
01:08:56.579 --> 01:09:00.539
So, Godzilla was inspired by this physics error.

589
01:09:00.539 --> 01:09:07.859
And, in fact, I think a cloud, some of the radiation would actually fall in and Troy, because it went up in the.

590
01:09:07.859 --> 01:09:16.229
stratosphere um and was precipitated out by the weather and it's something that might be in landed here any case so 

591
01:09:16.229 --> 01:09:21.210
You made computing error, your computing errors don't do this. So, you know.

592
01:09:21.210 --> 01:09:28.409
Could be life could be worse. Okay. And this is why the way it was the largest explosion that you guys did not do anything else.

593
01:09:28.409 --> 01:09:38.729
Okay, so, computer says so I'm just mentioning Lawrence Livermore, so okay, so computer systems I'm going to use, um.

594
01:09:42.989 --> 01:09:49.409
Okay, okay. So, um, I have a machine in my lab and cannot be a.

595
01:09:49.409 --> 01:09:53.369
So, to do a 14 CTC on.

596
01:09:53.369 --> 01:09:58.949
Now, each core is fairly, so you see 2 gigahertz. That's quite slow. Well, it.

597
01:09:58.949 --> 01:10:03.689
Over clocks that so, um, I hear interesting design things.

598
01:10:03.689 --> 01:10:08.279
The way these high end work is.

599
01:10:08.279 --> 01:10:12.239
The more parallel course they have the slower.

600
01:10:12.239 --> 01:10:19.859
They run because it's just hard getting the hardware working cross bar switches and all that stuff. So, the more parallel.

601
01:10:19.859 --> 01:10:29.760
The Intel processors that have more parallelism, have a have a slower clock speed. So when I was, I was checking out this machine to buy.

602
01:10:29.760 --> 01:10:33.600
Um, I went for more cores in the slower clock.

603
01:10:33.600 --> 01:10:37.319
And also keeping within my budget, of course, and now.

604
01:10:37.319 --> 01:10:42.270
The way it works is if you're not using the parallelism, the course, will actually over clock.

605
01:10:42.270 --> 01:10:48.720
So, if you're running a sequential program on this machine, then it goes much faster than that.

606
01:10:48.720 --> 01:10:56.250
Machine is several years old now. Okay. So this has a quarter terabyte of memory. So it's faster than.

607
01:10:56.250 --> 01:11:02.430
A lot better, be powered sort of thing and it has a better, um.

608
01:11:02.430 --> 01:11:06.989
It has the best Intel, the best and video.

609
01:11:06.989 --> 01:11:11.939
As of when I bought a couple of years ago.

610
01:11:11.939 --> 01:11:18.119
So, if so, this GPU has 48 gigabytes on the GPU card.

611
01:11:18.119 --> 01:11:30.600
Nice machine and a few more cores, but it has great tracing units on what the rate tracing unit does. It will do a quick test of array against the sphere or a box.

612
01:11:30.600 --> 01:11:34.649
And I'm going to stop, they could not going to update.

613
01:11:34.649 --> 01:11:45.210
Oh, okay. I have it and I have some machine machine, which we are from the cloud. So, what I'm going to do is give you all accounts on that to access from home.

614
01:11:45.210 --> 01:11:54.000
And again, we can, um, get data on and so, on here, um.

615
01:11:55.260 --> 01:12:00.539
If you wish, so, yeah, I'm not using the graphics on it. Um.

616
01:12:00.539 --> 01:12:06.479
Turing was a previous architecture before the vault because this is a couple years old.

617
01:12:06.479 --> 01:12:12.239
It's called a Quadro again, these different levels. Quadro Tesla.

618
01:12:12.239 --> 01:12:17.399
They overlap somewhat and so, and and video themselves. So.

619
01:12:18.479 --> 01:12:22.380
Actually, um.

620
01:12:22.380 --> 01:12:34.409
Nvidia themselves actually, they don't fall was wrong terminal because they keep changing things and some of the, um, the real time, very tracing cars. Um.

621
01:12:34.409 --> 01:12:47.640
Nice amount floating point performance here. 16 terra flops and a half performance. This is because again machine learning stuff doesn't have a lot of significant fits. So.

622
01:12:47.640 --> 01:13:01.380
Double precision to lot slower because they do that takes special hardware, tell you what an story about this. So we don't have to get 3 years ago. So I'm speaking out the machine to purchase. We have so purchasing thing.

623
01:13:01.380 --> 01:13:06.149
And the GPU here, and the lawyers.

624
01:13:06.149 --> 01:13:09.359
I saw this.

625
01:13:09.359 --> 01:13:20.430
And thought it was controlled technology. No, use the very detail export technology, export rules and they're surprisingly detail.

626
01:13:20.430 --> 01:13:27.779
And they have what the rules have been updated like, did they said any machine above? I can't remember the number.

627
01:13:27.779 --> 01:13:40.319
810 flocks or something was considered to be a military machine that was export control. So I got a note from our lawyers saying, um, well.

628
01:13:40.319 --> 01:13:45.720
Who is going to be using this machine and what's their citizenship and so on.

629
01:13:45.720 --> 01:13:54.270
And now, the thing about this card here is, at the time I bought this card, this card was this price or something like 600 dollars.

630
01:13:54.270 --> 01:13:57.960
Give or take, and it was available on Amazon.

631
01:13:57.960 --> 01:14:05.340
It was also available on Amazon, UK for 600 pounds. You know, the conversion um.

632
01:14:05.340 --> 01:14:17.845
Between an American, you know, high tech, pocket, computer product for 7 us and for sale, you know, you convert the price and I'm not joking. It's a character replaced when you change the dollar sign to a pound sign and you keep the digits the same.

633
01:14:18.265 --> 01:14:29.664
And that's what they in practice do with same, centrally. So 600 dollars in the state 600 pounds. So warriors I was saying, this is an export controlled machine, but it's available throughout the world on Amazon. So.

634
01:14:31.020 --> 01:14:37.470
I politely said that and said it's going to be used in class so they stopped bugging me but, um.

635
01:14:38.454 --> 01:14:52.675
You know, your work legal things. And the funny thing is that I also bought an Intel processor, the processor, which is even faster than this and they didn't complain about that because they didn't recognize it. They saw the purchase order and they stopped.

636
01:14:52.770 --> 01:14:58.590
But they saw also saw the I, I didn't know what his name was. I was saying.

637
01:14:58.590 --> 01:15:02.189
Okay, this is a nice fast machine here.

638
01:15:02.189 --> 01:15:05.340
Um, which you'll have accounts on. Okay.

639
01:15:05.340 --> 01:15:10.829
And again, um, so other stuff, um.

640
01:15:10.829 --> 01:15:18.270
We may do something on Amazon. I haven't planned to have Amazon actually has good values for some things. If you want to use.

641
01:15:18.270 --> 01:15:29.159
A high end video for some work, not for graphics latency. It'd be too big and bandwidth, but just for Super computing you can sometimes.

642
01:15:29.159 --> 01:15:36.210
Rented by the hour on Amazon for less than it would cost you to buy it advertised.

643
01:15:36.210 --> 01:15:44.010
At 1 point Amazon, I think had a GPU that would cost 10,000 dollars to buy. They were ready for a dollar an hour.

644
01:15:44.010 --> 01:15:53.880
Because you'd have to also figure out what for the CPO themselves that was good. I'll use grade scope for you used to submit home credit. Scopes is very nice for logging stuff in and quick creating stuff.

645
01:15:53.880 --> 01:15:57.180
So, and.

646
01:15:57.180 --> 01:16:09.630
Webex I'm just going to use only for distributing computed rates. I think it was great scope is it has no way in it for me to create a new column so to speak with gray tie compute.

647
01:16:09.630 --> 01:16:13.050
Elements says that the elements are horrible for everything else.

648
01:16:13.050 --> 01:16:16.590
So, distributed computer query to the end of the semester.

649
01:16:16.590 --> 01:16:20.489
Other stuff, if I see a new tool.

650
01:16:20.489 --> 01:16:33.060
I may use it, so how do you get grades homework's in term project and class presentations so well, more than teams, if you like, submit do a great scope teams to make 1.

651
01:16:33.060 --> 01:16:40.470
You can change your teams off as much as you want. You will be a term project at the end of the course, um, teams of any size.

652
01:16:40.470 --> 01:16:47.369
combine it with your thesis work with your grad school you combine it with another course you such a tell everyone that you're doing 

653
01:16:47.369 --> 01:16:51.359
Yeah, okay. And you can build on existing work. You have to tell me.

654
01:16:51.359 --> 01:16:55.739
So, anything I've got on my website, it's an automatic yes.

655
01:16:55.739 --> 01:17:08.880
So she got to implement their demonstrate or document something related to cloud computing and implement something right up a paper and scientific paper style.

656
01:17:08.880 --> 01:17:14.100
and with some more detailed man and i want you to give a project proposal in class 

657
01:17:14.100 --> 01:17:20.640
Middle semester, and then a presentation at the end it's in progress reports right now. Okay.

658
01:17:20.640 --> 01:17:26.189
Size of code it's impossible to say.

659
01:17:26.189 --> 01:17:40.800
um i've got an eight line program i wrote the test at the points and a polygon and wrote it many decades ago and still quoted occasionally and it's much shorter than other people's lives and as a story about bill gates he once was collaborating with ibm 

660
01:17:40.800 --> 01:17:49.800
And, um, and Bill Gates is a very good program, a very good hacker. Okay. And, um.

661
01:17:49.800 --> 01:17:59.069
So, he took some IBM code, rewrote it to be shorter and it's probably faster simultaneously, but the unity was working with didn't like it because IBM.

662
01:17:59.069 --> 01:18:06.659
quantify work by lines of code produced and so bill gates had caused that unit to do negative work by their metrics so 

663
01:18:06.659 --> 01:18:18.630
for most procedures i'm starting a virtual now switch to red line with rpm tells me too if i get to know this i haven't been on triple vaccinated i haven't gotten 

664
01:18:18.630 --> 01:18:28.920
You know, I'll probably catch the wrong thing at some point in our will tell me to go virtual. So I'll try to keep you in the loop here. Um.

665
01:18:28.920 --> 01:18:35.760
I could set up. Yeah, you can I could I got so I could set up a Webex space if you wanted, but.

666
01:18:35.760 --> 01:18:39.210
Um, I might do some in class stuff. I don't know.

667
01:18:39.210 --> 01:18:47.159
30 warning system I'll use, so basically, if you tell me that you weren't able to do something.

668
01:18:47.159 --> 01:18:52.890
because of family problems i will probably tell the data students so that you're having problems 

669
01:18:52.890 --> 01:18:58.050
Um, academic integrity is nothing new here and feedback.

670
01:18:58.050 --> 01:19:01.050
I created this course, um.

671
01:19:01.050 --> 01:19:10.109
I like the course, so I want to give you a good experience. If you have suggestions.

672
01:19:10.109 --> 01:19:14.159
You know, tell me in person if I don't scare you too much.

673
01:19:14.875 --> 01:19:28.614
Or tell someone else looking forward to it honestly the way the organization or professor Lou is the undergrad head, he organizes the undergrad classes and does assignments and prompts and you can tell him or I can tell him.

674
01:19:31.619 --> 01:19:44.430
I haven't set up the great great scope link. Yeah, I'll do it now. Okay. Class 1 this will be class 2. um, I do have homework out. Um, well, let me show organization of the blog here.

675
01:19:44.430 --> 01:19:49.770
And the archive lists, all the posts that we just sold postings.

676
01:19:49.770 --> 01:19:56.850
They are tagged. Okay. We've got the syllabus up here in my home page.

677
01:19:56.850 --> 01:20:00.479
The videos I'll probably, um.

678
01:20:00.479 --> 01:20:06.659
Um, I'll lower the make media site or else to hear both of them.

679
01:20:08.579 --> 01:20:14.279
My experience is that my backups are more reliable than backups. Um.

680
01:20:14.279 --> 01:20:20.670
So that I keep a copy of everything that I do all, you know.

681
01:20:20.670 --> 01:20:25.829
And last spring, the I get hit by the hackers and final exam period.

682
01:20:25.829 --> 01:20:28.979
I would say we're doing now. Um.

683
01:20:30.060 --> 01:20:33.600
Of course, um, I could just, um.

684
01:20:35.279 --> 01:20:43.020
The probability I just moved the course over to another off site hosts because my authoritative copy of it.

685
01:20:43.020 --> 01:20:46.949
On machines files with the files, that's.

686
01:20:46.949 --> 01:20:52.949
Okay. Homework. Okay. I've done enough talking. I'm getting worse. I want you to do some talking and we.

687
01:20:52.949 --> 01:20:57.539
Going to have and so I want you to pick some topics.

688
01:20:57.539 --> 01:21:02.850
And email me with your choices and I'll give you 1 of them. I'd like you to talk about it.

689
01:21:02.850 --> 01:21:09.329
Next Thursday, so that's a week and a half from now because next Monday is a holiday so.

690
01:21:09.329 --> 01:21:21.270
And you can do it, you can do it in the team. If you want on 2. if you want 35 to 10. I'm not going to jerk you off stage at 10 minutes. But, you know.

691
01:21:22.409 --> 01:21:31.319
Again, it's 5 to 10 should be 3 to 5. I'm going to shorten that. Sorry you know, several minute talk. So, and again, I'm gonna short these numbers here.

692
01:21:31.319 --> 01:21:41.520
And it may spill over to the call. So you tell me and I'll tell you which 1 of the 2 I pick. And basically it's a 1st come 1st served. I'll give you the 1st choice.

693
01:21:41.520 --> 01:21:47.880
Um, okay, so I'll start going from the Lawrence Livermore.

694
01:21:47.880 --> 01:21:51.000
Tutorial next Thursday, other than that.

695
01:21:52.109 --> 01:21:55.229
Um, if there's any questions.

696
01:21:55.229 --> 01:22:00.479
Hang around, if you'd like to ask any questions other than that. Um.

697
01:22:01.680 --> 01:22:12.000
You know, we'll start off with more stuff, so I wanted to use the time today to actually, you know, do sideways things to do introductions into stocks. And so I took the whole class.

698
01:22:13.380 --> 01:22:19.170
Um, okay.

699
01:22:19.170 --> 01:22:24.630
If there is no, let's see questions.

700
01:22:28.800 --> 01:22:32.909
Not to.

701
01:22:34.859 --> 01:22:41.069
Virtually.

702
01:22:43.020 --> 01:22:51.840
You're welcome. Oh, okay. What else? Okay. Bye. And I'll try to say the.

703
01:22:52.979 --> 01:22:56.369
It's this works with a video.