1
00:00:12,960 --> 00:00:18,320
and now we welcome to the science data

2
00:00:15,679 --> 00:00:21,119
and analytics track of pycon australia

3
00:00:18,320 --> 00:00:24,000
2021 marlene mangami

4
00:00:21,119 --> 00:00:24,960
malin is a zimbabwean software engineer

5
00:00:24,000 --> 00:00:26,960
and

6
00:00:24,960 --> 00:00:29,359
engineer exploring speaker based on the

7
00:00:26,960 --> 00:00:31,279
city of harare an advocate for using

8
00:00:29,359 --> 00:00:32,960
science and technology for social good

9
00:00:31,279 --> 00:00:34,880
she's a director and the vice chair for

10
00:00:32,960 --> 00:00:36,800
the python software foundation and is

11
00:00:34,880 --> 00:00:39,440
currently working as a software engineer

12
00:00:36,800 --> 00:00:42,480
with the rapids team at nvidia

13
00:00:39,440 --> 00:00:44,160
in 2017 chico founded zimbo pi a

14
00:00:42,480 --> 00:00:45,920
non-profit organization that gives

15
00:00:44,160 --> 00:00:48,000
zimbabwean young women access to

16
00:00:45,920 --> 00:00:49,840
resources in the field of technology she

17
00:00:48,000 --> 00:00:51,520
is currently pursuing a computer science

18
00:00:49,840 --> 00:00:53,600
degree with a concentration in machine

19
00:00:51,520 --> 00:00:54,800
learning and ai at the university of

20
00:00:53,600 --> 00:00:56,960
london

21
00:00:54,800 --> 00:00:59,520
her her talk is recorded so there won't

22
00:00:56,960 --> 00:01:01,199
be an introduction or q a but you can

23
00:00:59,520 --> 00:01:04,080
find her at home on the web

24
00:01:01,199 --> 00:01:06,880
marlenemangami.com

25
00:01:04,080 --> 00:01:09,119
the talk is about python and gpus maleen

26
00:01:06,880 --> 00:01:10,560
will give us a brief history of python

27
00:01:09,119 --> 00:01:12,400
and parallel computing including

28
00:01:10,560 --> 00:01:14,240
python's infamous gill

29
00:01:12,400 --> 00:01:16,320
and she'll also share how and why

30
00:01:14,240 --> 00:01:19,520
pythonistas can get started with using

31
00:01:16,320 --> 00:01:22,159
gpus to do more work and faster welcome

32
00:01:19,520 --> 00:01:24,240
to python australia marlene mangami and

33
00:01:22,159 --> 00:01:28,640
with all of you a beginner's guide to

34
00:01:24,240 --> 00:01:28,640
gpu for by gpus for pythonistas

35
00:01:31,840 --> 00:01:39,280
i'm so excited to be here with you today

36
00:01:34,240 --> 00:01:42,399
for the virtual version of pycon line

37
00:01:39,280 --> 00:01:45,520
for pycon australia um the title of my

38
00:01:42,399 --> 00:01:48,240
talk today is a beginner's guide to gpus

39
00:01:45,520 --> 00:01:49,680
for python easters i'm very excited

40
00:01:48,240 --> 00:01:52,159
about this talk if you have any

41
00:01:49,680 --> 00:01:53,360
questions after i give my talk feel free

42
00:01:52,159 --> 00:01:56,320
to

43
00:01:53,360 --> 00:01:57,920
find me on twitter and just send me a dm

44
00:01:56,320 --> 00:02:00,240
i think i might be

45
00:01:57,920 --> 00:02:03,200
around after this talk as well just to

46
00:02:00,240 --> 00:02:06,079
answer any questions too

47
00:02:03,200 --> 00:02:06,960
first thing something about me is that i

48
00:02:06,079 --> 00:02:09,599
am

49
00:02:06,960 --> 00:02:11,840
the current vice chair and a director at

50
00:02:09,599 --> 00:02:13,920
the python software foundation with the

51
00:02:11,840 --> 00:02:15,040
which is the non-profit organization

52
00:02:13,920 --> 00:02:16,800
behind

53
00:02:15,040 --> 00:02:19,200
python programming language it's a very

54
00:02:16,800 --> 00:02:20,800
cool organization that i encourage you

55
00:02:19,200 --> 00:02:23,120
to look into

56
00:02:20,800 --> 00:02:24,959
something else about myself is that i am

57
00:02:23,120 --> 00:02:27,599
a co-founder of a non-profit

58
00:02:24,959 --> 00:02:29,760
organization here in zimbabwe which is

59
00:02:27,599 --> 00:02:32,239
actually where i'm calling you from

60
00:02:29,760 --> 00:02:34,720
i'm called zimbo-pai and this

61
00:02:32,239 --> 00:02:38,239
organization just tries to get girls

62
00:02:34,720 --> 00:02:39,519
excited about programming in python and

63
00:02:38,239 --> 00:02:41,519
then uh

64
00:02:39,519 --> 00:02:43,599
something else about myself is that

65
00:02:41,519 --> 00:02:46,160
a final thing about myself actually

66
00:02:43,599 --> 00:02:48,160
so that this isn't too long is that i'm

67
00:02:46,160 --> 00:02:51,440
currently interning as a software

68
00:02:48,160 --> 00:02:55,920
engineer with the rapids team at nvidia

69
00:02:51,440 --> 00:02:58,000
and i am working on a python um gpu

70
00:02:55,920 --> 00:03:00,400
based data frame library that i will

71
00:02:58,000 --> 00:03:02,239
talk a little bit about um

72
00:03:00,400 --> 00:03:05,280
some of the work that i'm doing in this

73
00:03:02,239 --> 00:03:06,959
talk as well and the library i work on

74
00:03:05,280 --> 00:03:08,800
is open source so you can also check out

75
00:03:06,959 --> 00:03:11,760
the code after if you'd like

76
00:03:08,800 --> 00:03:13,519
so let's go ahead and dive right in i

77
00:03:11,760 --> 00:03:15,840
think a good place to start is

78
00:03:13,519 --> 00:03:20,400
understanding why python easter should

79
00:03:15,840 --> 00:03:23,440
be interested in gpus in the first place

80
00:03:20,400 --> 00:03:26,799
typically when people think about gpus

81
00:03:23,440 --> 00:03:30,400
they immediately their mind goes to

82
00:03:26,799 --> 00:03:33,040
graphics cards or gaming and i know for

83
00:03:30,400 --> 00:03:35,519
myself i was really surprised when i

84
00:03:33,040 --> 00:03:37,920
started reading about this that

85
00:03:35,519 --> 00:03:40,959
gpus were being used for anything else

86
00:03:37,920 --> 00:03:43,519
outside of the gaming world and this

87
00:03:40,959 --> 00:03:44,400
assumption is actually not too far off

88
00:03:43,519 --> 00:03:47,440
base

89
00:03:44,400 --> 00:03:49,760
in fact if we did a quick search on

90
00:03:47,440 --> 00:03:52,480
wikipedia which as many of us know is

91
00:03:49,760 --> 00:03:55,840
the source of all truth

92
00:03:52,480 --> 00:03:58,720
we find that gpu is actually an acronym

93
00:03:55,840 --> 00:04:01,040
for graphics processing unit and so the

94
00:03:58,720 --> 00:04:04,239
original purpose of the gpu was to

95
00:04:01,040 --> 00:04:05,760
create images for computer graphics and

96
00:04:04,239 --> 00:04:08,000
gaming consoles

97
00:04:05,760 --> 00:04:10,000
so that assumption that i had and i

98
00:04:08,000 --> 00:04:12,319
think many people around the world have

99
00:04:10,000 --> 00:04:15,200
is is actually fairly accurate since the

100
00:04:12,319 --> 00:04:19,440
early 2010s we saw a sharp rise in the

101
00:04:15,200 --> 00:04:21,759
use of gp gbus or general purpose

102
00:04:19,440 --> 00:04:25,199
graphic processing units

103
00:04:21,759 --> 00:04:28,639
and as one can imagine general purpose

104
00:04:25,199 --> 00:04:30,960
gpus are gpus that can be used

105
00:04:28,639 --> 00:04:32,720
for more general purposes

106
00:04:30,960 --> 00:04:34,800
and typically

107
00:04:32,720 --> 00:04:37,440
these sorts of gpus are used instead of

108
00:04:34,800 --> 00:04:40,960
a cpu for executing tasks that can be

109
00:04:37,440 --> 00:04:43,120
run in parallel so i think i have a

110
00:04:40,960 --> 00:04:45,280
pretty good example uh to just

111
00:04:43,120 --> 00:04:47,759
demonstrate what this means so one of

112
00:04:45,280 --> 00:04:50,560
the things that i absolutely love about

113
00:04:47,759 --> 00:04:51,360
conferences being remote these days is

114
00:04:50,560 --> 00:04:54,479
that

115
00:04:51,360 --> 00:04:56,880
i don't have to wait in line for a cup

116
00:04:54,479 --> 00:04:59,120
of coffee like if i if i would like

117
00:04:56,880 --> 00:05:01,199
coffee right now i just can get up and

118
00:04:59,120 --> 00:05:02,880
go to my kitchen and have some coffee

119
00:05:01,199 --> 00:05:05,440
but sometimes

120
00:05:02,880 --> 00:05:08,000
you'll attend these conferences that are

121
00:05:05,440 --> 00:05:09,680
great and and not all the time as well

122
00:05:08,000 --> 00:05:11,440
but you can find yourself in these

123
00:05:09,680 --> 00:05:13,759
really long lines

124
00:05:11,440 --> 00:05:16,240
waiting to use a coffee machine and

125
00:05:13,759 --> 00:05:18,240
sometimes it's a single coffee machine

126
00:05:16,240 --> 00:05:20,400
and now you're standing in line and

127
00:05:18,240 --> 00:05:23,120
you're debating whether you should go

128
00:05:20,400 --> 00:05:26,560
and give your talk like a mature adult

129
00:05:23,120 --> 00:05:28,560
or just stay and wait for a precious cup

130
00:05:26,560 --> 00:05:30,960
of coffee goodness

131
00:05:28,560 --> 00:05:32,240
but there are a number of solutions to

132
00:05:30,960 --> 00:05:33,360
this problem

133
00:05:32,240 --> 00:05:35,120
and

134
00:05:33,360 --> 00:05:37,039
one solution that a conference

135
00:05:35,120 --> 00:05:39,600
organizers if they can afford it can

136
00:05:37,039 --> 00:05:42,240
consider is to add a

137
00:05:39,600 --> 00:05:44,320
an extra coffee machine into the mix

138
00:05:42,240 --> 00:05:46,720
and this is great because it allows

139
00:05:44,320 --> 00:05:48,080
things to move much faster because the

140
00:05:46,720 --> 00:05:50,560
cues

141
00:05:48,080 --> 00:05:53,360
like generally become shorter and and

142
00:05:50,560 --> 00:05:56,319
maybe are even cut in half or even into

143
00:05:53,360 --> 00:05:59,199
a third and people are able to access

144
00:05:56,319 --> 00:06:01,680
their coffee grab it and go and get on

145
00:05:59,199 --> 00:06:03,759
with their day and be more productive so

146
00:06:01,680 --> 00:06:05,759
we can think of the cores in our

147
00:06:03,759 --> 00:06:07,759
computers in the same way we would think

148
00:06:05,759 --> 00:06:11,039
about these coffee machines in many

149
00:06:07,759 --> 00:06:13,680
cases utilizing more cores can make us

150
00:06:11,039 --> 00:06:16,720
more productive in the work that we do

151
00:06:13,680 --> 00:06:19,680
so a typical cpu contains anywhere from

152
00:06:16,720 --> 00:06:23,759
between four to eight cores whereas a

153
00:06:19,680 --> 00:06:28,000
gpu can contain thousands of them

154
00:06:23,759 --> 00:06:30,319
so having a higher number of cores

155
00:06:28,000 --> 00:06:33,600
increases the computational power that

156
00:06:30,319 --> 00:06:35,120
we have and allows us to do more work

157
00:06:33,600 --> 00:06:36,240
faster

158
00:06:35,120 --> 00:06:39,759
so

159
00:06:36,240 --> 00:06:40,960
just before i get a slew

160
00:06:39,759 --> 00:06:43,919
of

161
00:06:40,960 --> 00:06:46,840
emails from people who love cpus this is

162
00:06:43,919 --> 00:06:49,919
not me throwing cpus under the bus i use

163
00:06:46,840 --> 00:06:52,960
cpus actually there are some tasks that

164
00:06:49,919 --> 00:06:56,000
are much better suited to using a cpu as

165
00:06:52,960 --> 00:06:58,080
compared to a gpu so for example if you

166
00:06:56,000 --> 00:07:00,720
have a task that requires a lot of

167
00:06:58,080 --> 00:07:03,599
serial processing where you need to

168
00:07:00,720 --> 00:07:07,120
purchase things one by one in a sequence

169
00:07:03,599 --> 00:07:11,039
it would definitely be best to use a cpu

170
00:07:07,120 --> 00:07:13,440
in that case um gpus are great but they

171
00:07:11,039 --> 00:07:16,160
expect to run things simultaneously and

172
00:07:13,440 --> 00:07:18,080
so if your computations cannot be done

173
00:07:16,160 --> 00:07:21,919
simultaneously then it's abyss to stick

174
00:07:18,080 --> 00:07:23,840
to a cpu and by the way you actually can

175
00:07:21,919 --> 00:07:26,960
take advantage of

176
00:07:23,840 --> 00:07:29,199
more cores in your cpu

177
00:07:26,960 --> 00:07:31,840
by using things like multi-threading or

178
00:07:29,199 --> 00:07:33,759
async io among other things to boost

179
00:07:31,840 --> 00:07:35,840
your performance

180
00:07:33,759 --> 00:07:38,080
locally with the cpu so from what we've

181
00:07:35,840 --> 00:07:40,080
talked about so far we've established

182
00:07:38,080 --> 00:07:40,880
that if you're a pythony style who needs

183
00:07:40,080 --> 00:07:43,039
to

184
00:07:40,880 --> 00:07:44,400
do some computations that can be run

185
00:07:43,039 --> 00:07:46,639
simultaneously

186
00:07:44,400 --> 00:07:50,240
gpus are a really great option for you

187
00:07:46,639 --> 00:07:51,919
to consider because uh ultimately your

188
00:07:50,240 --> 00:07:54,000
work will be done much faster because

189
00:07:51,919 --> 00:07:56,560
your code will be running much faster

190
00:07:54,000 --> 00:07:58,400
and since this is a beginner's guide

191
00:07:56,560 --> 00:08:00,720
some people watching someone watching

192
00:07:58,400 --> 00:08:03,120
might be uh thinking to themselves

193
00:08:00,720 --> 00:08:04,319
marlene i have been using python for a

194
00:08:03,120 --> 00:08:07,599
long time

195
00:08:04,319 --> 00:08:10,240
and i don't understand why i would need

196
00:08:07,599 --> 00:08:11,120
to speak python up it seems pretty fast

197
00:08:10,240 --> 00:08:13,599
to me

198
00:08:11,120 --> 00:08:14,800
and that's you know i think a good

199
00:08:13,599 --> 00:08:18,560
question

200
00:08:14,800 --> 00:08:19,919
uh to to to pose to to that question as

201
00:08:18,560 --> 00:08:23,120
well is

202
00:08:19,919 --> 00:08:25,199
um is python slow

203
00:08:23,120 --> 00:08:27,840
and uh believe it or not this is

204
00:08:25,199 --> 00:08:30,240
actually quite a contentious question

205
00:08:27,840 --> 00:08:32,800
and when i looked online for answers to

206
00:08:30,240 --> 00:08:34,000
this question i found some opinions that

207
00:08:32,800 --> 00:08:36,800
were helpful

208
00:08:34,000 --> 00:08:39,200
but others that were to be honest

209
00:08:36,800 --> 00:08:41,440
probably borderline slander

210
00:08:39,200 --> 00:08:43,919
and i personally came to the conclusion

211
00:08:41,440 --> 00:08:46,240
that the slowness of python is actually

212
00:08:43,919 --> 00:08:49,120
fairly relative and many developers will

213
00:08:46,240 --> 00:08:52,959
likely never encounter this slowness

214
00:08:49,120 --> 00:08:54,399
that people talk about um but to be fair

215
00:08:52,959 --> 00:08:57,600
there is

216
00:08:54,399 --> 00:08:59,440
some level of truth to that statement

217
00:08:57,600 --> 00:09:02,320
and i do think that it's important that

218
00:08:59,440 --> 00:09:04,560
when we are talking about python um

219
00:09:02,320 --> 00:09:06,720
potentially being so that we give some

220
00:09:04,560 --> 00:09:10,800
context and explain what we're comparing

221
00:09:06,720 --> 00:09:12,959
python to so last year in 2020 anthony

222
00:09:10,800 --> 00:09:16,720
shaw gave a really great talk at python

223
00:09:12,959 --> 00:09:19,279
us and the talk was titled why is python

224
00:09:16,720 --> 00:09:20,560
slow and in that talk

225
00:09:19,279 --> 00:09:22,480
he

226
00:09:20,560 --> 00:09:25,040
put this this really great graph that's

227
00:09:22,480 --> 00:09:27,360
on the screen right now um and this is a

228
00:09:25,040 --> 00:09:29,360
graph that shows the in body algorithm

229
00:09:27,360 --> 00:09:32,800
and it shows how long different

230
00:09:29,360 --> 00:09:35,680
languages take to run this algorithm and

231
00:09:32,800 --> 00:09:38,160
uh the program models the orbits of

232
00:09:35,680 --> 00:09:41,519
jupiter saturn uranus and neptune and is

233
00:09:38,160 --> 00:09:43,200
pretty uh sort of intense python for

234
00:09:41,519 --> 00:09:46,000
some of you some of you may know that

235
00:09:43,200 --> 00:09:48,720
python is a an interpreted language that

236
00:09:46,000 --> 00:09:51,360
is written and compiled in c

237
00:09:48,720 --> 00:09:52,480
but if you look at the chart and you

238
00:09:51,360 --> 00:09:56,160
compare

239
00:09:52,480 --> 00:09:59,920
the time it takes for c to

240
00:09:56,160 --> 00:10:03,200
run this algorithm c takes seven seconds

241
00:09:59,920 --> 00:10:05,519
while python takes 14 minutes

242
00:10:03,200 --> 00:10:08,000
i don't even know how many times

243
00:10:05,519 --> 00:10:10,880
slower that is and you know someone

244
00:10:08,000 --> 00:10:13,440
might be thinking okay but c is actually

245
00:10:10,880 --> 00:10:15,279
a strongly typed and compiled language

246
00:10:13,440 --> 00:10:18,560
while we know that python is dynamically

247
00:10:15,279 --> 00:10:20,800
typed and interpreted so you know that's

248
00:10:18,560 --> 00:10:23,839
not really a fair comparison but if we

249
00:10:20,800 --> 00:10:25,360
look at another language like node.js

250
00:10:23,839 --> 00:10:27,200
which is also

251
00:10:25,360 --> 00:10:30,800
dynamically typed

252
00:10:27,200 --> 00:10:33,760
uh similarly to python we still see that

253
00:10:30,800 --> 00:10:34,800
python is much slower than node.js

254
00:10:33,760 --> 00:10:36,959
so

255
00:10:34,800 --> 00:10:38,240
if you'd if you're really interested in

256
00:10:36,959 --> 00:10:41,120
this specific

257
00:10:38,240 --> 00:10:43,120
example and would want to find out some

258
00:10:41,120 --> 00:10:45,760
specific reasons about why this is the

259
00:10:43,120 --> 00:10:48,240
case i definitely recommend you watching

260
00:10:45,760 --> 00:10:49,040
anthony's talk it's a fantastic talk

261
00:10:48,240 --> 00:10:51,440
but

262
00:10:49,040 --> 00:10:53,680
something to note is that even though

263
00:10:51,440 --> 00:10:56,320
this might be the case according to

264
00:10:53,680 --> 00:10:58,079
stackoverflow's developer survey python

265
00:10:56,320 --> 00:11:00,399
beats out almost every single other

266
00:10:58,079 --> 00:11:04,000
language in this chart in terms of

267
00:11:00,399 --> 00:11:05,680
popularity and the truth is python to me

268
00:11:04,000 --> 00:11:06,560
at least i think to a lot of other

269
00:11:05,680 --> 00:11:09,360
people

270
00:11:06,560 --> 00:11:12,800
is a really good and fun language to

271
00:11:09,360 --> 00:11:16,480
program in python is such a joy to use

272
00:11:12,800 --> 00:11:19,200
and so if you really need that um if you

273
00:11:16,480 --> 00:11:21,839
would want that speed up and making sure

274
00:11:19,200 --> 00:11:24,640
that you have uh your code running as

275
00:11:21,839 --> 00:11:28,000
fast as possible but also want to

276
00:11:24,640 --> 00:11:30,800
continue enjoying coding in python there

277
00:11:28,000 --> 00:11:33,120
are lots of ways to be able to do that

278
00:11:30,800 --> 00:11:34,959
earlier today i talked about how

279
00:11:33,120 --> 00:11:37,279
utilizing more cores through things like

280
00:11:34,959 --> 00:11:39,120
multi-threading and concurrency can help

281
00:11:37,279 --> 00:11:41,839
us get better performance from our

282
00:11:39,120 --> 00:11:43,760
python code and in the past python

283
00:11:41,839 --> 00:11:45,760
hasn't actually had a very friendly

284
00:11:43,760 --> 00:11:49,440
reputation with things like

285
00:11:45,760 --> 00:11:52,399
multi-threading and parallelism because

286
00:11:49,440 --> 00:11:54,079
of uh something called

287
00:11:52,399 --> 00:11:54,959
the guild

288
00:11:54,079 --> 00:11:56,399
so

289
00:11:54,959 --> 00:11:59,920
you know i wouldn't

290
00:11:56,399 --> 00:12:02,639
really have talked about python and

291
00:11:59,920 --> 00:12:04,639
concurrency python and speed even

292
00:12:02,639 --> 00:12:07,440
if i don't at least

293
00:12:04,639 --> 00:12:10,800
mention the gill exactly is going on

294
00:12:07,440 --> 00:12:14,079
with python skill and if you search

295
00:12:10,800 --> 00:12:17,440
online for python talks on the gill you

296
00:12:14,079 --> 00:12:18,480
get this very strange mix of suggestions

297
00:12:17,440 --> 00:12:20,320
um

298
00:12:18,480 --> 00:12:23,200
as you can see you know some people are

299
00:12:20,320 --> 00:12:25,680
talking about removing it completely

300
00:12:23,200 --> 00:12:27,360
um others are talking about

301
00:12:25,680 --> 00:12:29,360
understanding it more trying to help

302
00:12:27,360 --> 00:12:32,240
others understand it i think

303
00:12:29,360 --> 00:12:35,040
and even even more you know recently

304
00:12:32,240 --> 00:12:36,959
eric snow gave a talk asking whether to

305
00:12:35,040 --> 00:12:39,680
kill or not to kill

306
00:12:36,959 --> 00:12:41,040
so generally it seems like it has this

307
00:12:39,680 --> 00:12:44,240
really big

308
00:12:41,040 --> 00:12:47,120
reputation so so what is the gill

309
00:12:44,240 --> 00:12:51,120
um well the girl is python's global

310
00:12:47,120 --> 00:12:53,360
interpreter lock and this is a lock in c

311
00:12:51,120 --> 00:12:55,680
python that preserves its internal

312
00:12:53,360 --> 00:12:57,200
shared global state and what

313
00:12:55,680 --> 00:12:59,760
occasionally makes the gill

314
00:12:57,200 --> 00:13:03,680
controversial is that because of it no

315
00:12:59,760 --> 00:13:06,000
more than one thread can run at a time

316
00:13:03,680 --> 00:13:08,959
and for people that are wanting to

317
00:13:06,000 --> 00:13:10,480
utilize things like multi-threading um

318
00:13:08,959 --> 00:13:12,320
you know which is running many different

319
00:13:10,480 --> 00:13:14,880
threads at the same time to speed up

320
00:13:12,320 --> 00:13:17,519
their work the girl can stand in their

321
00:13:14,880 --> 00:13:19,519
way potentially however

322
00:13:17,519 --> 00:13:22,399
in recent years the girl has become less

323
00:13:19,519 --> 00:13:24,560
and less of a barrier for people who

324
00:13:22,399 --> 00:13:25,839
want to speed up their python code in

325
00:13:24,560 --> 00:13:29,200
this respect

326
00:13:25,839 --> 00:13:31,680
and gpus are only one of a number of

327
00:13:29,200 --> 00:13:32,880
different ways to get around it and i

328
00:13:31,680 --> 00:13:35,519
thought

329
00:13:32,880 --> 00:13:37,760
this uh this next clip that i have on

330
00:13:35,519 --> 00:13:41,760
the screen is a clip from raymond

331
00:13:37,760 --> 00:13:44,079
hittinger's 2017 keynote at pi bay and i

332
00:13:41,760 --> 00:13:46,399
thought it was really helpful in terms

333
00:13:44,079 --> 00:13:48,320
of helping understand why the gill is

334
00:13:46,399 --> 00:13:49,760
important and some ways that you can

335
00:13:48,320 --> 00:13:51,519
actually get around it

336
00:13:49,760 --> 00:13:53,360
our friend larry hastings is working on

337
00:13:51,519 --> 00:13:55,040
a project called the gillectomy to

338
00:13:53,360 --> 00:13:56,880
remove the global interpreter lock from

339
00:13:55,040 --> 00:13:58,320
python do you think that's a difficult

340
00:13:56,880 --> 00:14:01,120
project

341
00:13:58,320 --> 00:14:03,680
that is an incorrect hypothesis

342
00:14:01,120 --> 00:14:05,120
uh takes about a day of work

343
00:14:03,680 --> 00:14:07,600
to remove the gill

344
00:14:05,120 --> 00:14:09,920
he removed the gill on the first day it

345
00:14:07,600 --> 00:14:12,000
was no problem taking it out

346
00:14:09,920 --> 00:14:13,760
the problem then becomes all of the

347
00:14:12,000 --> 00:14:16,320
locks that you have to put in everywhere

348
00:14:13,760 --> 00:14:18,320
else in order to get python uh to

349
00:14:16,320 --> 00:14:20,560
function it turns out that's not

350
00:14:18,320 --> 00:14:22,639
particularly hard either and a few days

351
00:14:20,560 --> 00:14:26,399
later he had that done

352
00:14:22,639 --> 00:14:28,320
so the gill is gone and replaced by uh

353
00:14:26,399 --> 00:14:30,800
lots of other little locks on smaller

354
00:14:28,320 --> 00:14:33,120
data structures problem solved any

355
00:14:30,800 --> 00:14:33,120
questions

356
00:14:34,240 --> 00:14:36,720
oh

357
00:14:35,120 --> 00:14:38,000
are locks expensive to acquire in

358
00:14:36,720 --> 00:14:40,480
release

359
00:14:38,000 --> 00:14:43,279
in fact they are and so the good news is

360
00:14:40,480 --> 00:14:45,360
the gill is gone it's free threaded and

361
00:14:43,279 --> 00:14:47,279
it's uh dozens of times slower than

362
00:14:45,360 --> 00:14:49,199
regular uh python

363
00:14:47,279 --> 00:14:52,000
so you actually get a payoff for the

364
00:14:49,199 --> 00:14:53,760
guild and the payoff is you don't pay

365
00:14:52,000 --> 00:14:56,480
all of the performance cost of all of

366
00:14:53,760 --> 00:14:58,959
these individual lock acquires and our

367
00:14:56,480 --> 00:15:02,000
releases is actually a really nice thing

368
00:14:58,959 --> 00:15:03,839
uh to have it gets in the way of us free

369
00:15:02,000 --> 00:15:06,000
threading but we have ways of solving

370
00:15:03,839 --> 00:15:08,639
that problem if you can't fully free

371
00:15:06,000 --> 00:15:10,880
thread one python why don't i run eight

372
00:15:08,639 --> 00:15:12,720
pythons in parallel each with their own

373
00:15:10,880 --> 00:15:14,160
threads

374
00:15:12,720 --> 00:15:16,880
and then it's no problem i'm taking

375
00:15:14,160 --> 00:15:18,000
advantage of all of other cores or you

376
00:15:16,880 --> 00:15:20,240
can combine threading and

377
00:15:18,000 --> 00:15:24,160
multi-processing uh there's a number of

378
00:15:20,240 --> 00:15:26,160
ways to go in fact at some point most

379
00:15:24,160 --> 00:15:28,480
folks just get over the python has a

380
00:15:26,160 --> 00:15:30,720
global interpreter lock go ahead and

381
00:15:28,480 --> 00:15:32,320
saturate all eight cores to a hundred

382
00:15:30,720 --> 00:15:34,000
percent and get full advantage of the

383
00:15:32,320 --> 00:15:35,759
other machine and just simply ignore the

384
00:15:34,000 --> 00:15:37,839
problem there's lots of ways to ignore

385
00:15:35,759 --> 00:15:40,240
the global interpreter lock it is not

386
00:15:37,839 --> 00:15:44,000
that big of deep okay so like raymond

387
00:15:40,240 --> 00:15:45,920
said when we utilize all of our cores we

388
00:15:44,000 --> 00:15:49,199
don't really need to worry about the

389
00:15:45,920 --> 00:15:51,600
guilt and with gpus we have thousands of

390
00:15:49,199 --> 00:15:54,959
cores so if we decide we wanted to even

391
00:15:51,600 --> 00:15:57,759
just run one thread on each core we

392
00:15:54,959 --> 00:16:00,399
would still have a really significant

393
00:15:57,759 --> 00:16:03,040
speed up okay so

394
00:16:00,399 --> 00:16:05,839
great now that we've looked at some of

395
00:16:03,040 --> 00:16:08,959
the big performance issues that would

396
00:16:05,839 --> 00:16:11,360
lead us to possibly using a gpu when

397
00:16:08,959 --> 00:16:13,759
we're running python let's go ahead and

398
00:16:11,360 --> 00:16:15,680
look at some practical examples of how

399
00:16:13,759 --> 00:16:18,800
we can do this

400
00:16:15,680 --> 00:16:21,279
so modern datasets can have as much as

401
00:16:18,800 --> 00:16:24,880
millions even billions of data points

402
00:16:21,279 --> 00:16:26,800
that need processing and for the average

403
00:16:24,880 --> 00:16:29,199
data scientist that is working with

404
00:16:26,800 --> 00:16:31,279
these large data sets just using

405
00:16:29,199 --> 00:16:34,560
ordinary python for processing this

406
00:16:31,279 --> 00:16:36,079
information is not very efficient and

407
00:16:34,560 --> 00:16:38,720
most of the time you know they would

408
00:16:36,079 --> 00:16:41,519
spend if they chose to to run python

409
00:16:38,720 --> 00:16:43,920
only they could spend a lot of time just

410
00:16:41,519 --> 00:16:46,160
waiting around and not being productive

411
00:16:43,920 --> 00:16:48,880
but these days data scientists wanting

412
00:16:46,160 --> 00:16:51,120
to perform larger computations on

413
00:16:48,880 --> 00:16:54,160
datasets or wanting to filter through

414
00:16:51,120 --> 00:16:55,839
information in larger quantities

415
00:16:54,160 --> 00:16:58,079
there are a number of different

416
00:16:55,839 --> 00:17:01,519
libraries available so pandas is

417
00:16:58,079 --> 00:17:02,639
probably the most popular solution for

418
00:17:01,519 --> 00:17:04,480
this

419
00:17:02,639 --> 00:17:06,880
it's a data science library written in

420
00:17:04,480 --> 00:17:09,520
python c and cython

421
00:17:06,880 --> 00:17:12,559
that uses vectorized operations in order

422
00:17:09,520 --> 00:17:15,039
to run functions at very high speeds and

423
00:17:12,559 --> 00:17:16,880
although it doesn't use gpus

424
00:17:15,039 --> 00:17:19,439
let's quickly take a look at how much of

425
00:17:16,880 --> 00:17:22,160
a speed up we get when we try to create

426
00:17:19,439 --> 00:17:23,120
a data table with pandas versus with

427
00:17:22,160 --> 00:17:25,439
python

428
00:17:23,120 --> 00:17:28,480
at the moment we are starting off on

429
00:17:25,439 --> 00:17:30,080
google colab and this specifically for

430
00:17:28,480 --> 00:17:32,480
context is

431
00:17:30,080 --> 00:17:33,760
the rapids notebook rapids has a really

432
00:17:32,480 --> 00:17:37,039
great um

433
00:17:33,760 --> 00:17:39,200
free notebook on um if you just go to

434
00:17:37,039 --> 00:17:41,039
the get started page

435
00:17:39,200 --> 00:17:43,840
you can have access to this notebook and

436
00:17:41,039 --> 00:17:47,280
basically it allows you to test out a

437
00:17:43,840 --> 00:17:49,679
gpu and the different rapids

438
00:17:47,280 --> 00:17:51,200
libraries very quickly and

439
00:17:49,679 --> 00:17:53,280
we'll just walk you through some of the

440
00:17:51,200 --> 00:17:55,840
basics of that

441
00:17:53,280 --> 00:17:58,720
we are not yet starting off with our gpu

442
00:17:55,840 --> 00:18:01,039
but just to give you some extra context

443
00:17:58,720 --> 00:18:02,720
at the moment we are going to we are

444
00:18:01,039 --> 00:18:06,320
importing our libraries and then we're

445
00:18:02,720 --> 00:18:08,720
creating um some data using uh numpy

446
00:18:06,320 --> 00:18:10,080
some random data and then

447
00:18:08,720 --> 00:18:13,360
we're going to start off by creating a

448
00:18:10,080 --> 00:18:16,160
pandas data frame and this is you know a

449
00:18:13,360 --> 00:18:18,240
very simple way to do it and then after

450
00:18:16,160 --> 00:18:21,120
we create this data frame we're going to

451
00:18:18,240 --> 00:18:22,480
see how long it takes using

452
00:18:21,120 --> 00:18:24,320
the time module we're going to see how

453
00:18:22,480 --> 00:18:27,120
long it takes for

454
00:18:24,320 --> 00:18:29,760
pandas to calculate the mean

455
00:18:27,120 --> 00:18:31,200
of the data in this column

456
00:18:29,760 --> 00:18:34,160
okay so

457
00:18:31,200 --> 00:18:36,480
i'm just going to go ahead and do that

458
00:18:34,160 --> 00:18:40,160
i'm going to create a

459
00:18:36,480 --> 00:18:43,679
separate column in the pandas data frame

460
00:18:40,160 --> 00:18:47,440
and i'm going to go ahead and

461
00:18:43,679 --> 00:18:49,760
use the pandas rolling method um to

462
00:18:47,440 --> 00:18:52,559
calculate the mean and this is a great

463
00:18:49,760 --> 00:18:53,520
method that really speeds things up

464
00:18:52,559 --> 00:18:56,880
um

465
00:18:53,520 --> 00:18:58,240
yeah so let's go ahead and see how long

466
00:18:56,880 --> 00:19:00,960
it takes for

467
00:18:58,240 --> 00:19:02,160
um pandas to calculate the mean oh that

468
00:19:00,960 --> 00:19:04,720
was really quick

469
00:19:02,160 --> 00:19:07,520
and it takes about 0.03

470
00:19:04,720 --> 00:19:09,600
seconds as you can see there which is

471
00:19:07,520 --> 00:19:11,679
really fast um

472
00:19:09,600 --> 00:19:13,760
now the next thing we're going to do is

473
00:19:11,679 --> 00:19:15,520
we're going to figure out how long it's

474
00:19:13,760 --> 00:19:18,799
going to take for

475
00:19:15,520 --> 00:19:21,520
python to do the same thing and instead

476
00:19:18,799 --> 00:19:24,480
of using the python rolling method we're

477
00:19:21,520 --> 00:19:28,240
going to create as i'm doing now

478
00:19:24,480 --> 00:19:30,559
we are going to create a for loop and

479
00:19:28,240 --> 00:19:33,120
basically what this for loop is going to

480
00:19:30,559 --> 00:19:36,880
do is it's uh we're going to use the

481
00:19:33,120 --> 00:19:39,039
mean loop column in the data frame and

482
00:19:36,880 --> 00:19:40,400
we are going to just loop through all of

483
00:19:39,039 --> 00:19:43,120
the values

484
00:19:40,400 --> 00:19:44,960
and calculate their means

485
00:19:43,120 --> 00:19:45,760
and add them to

486
00:19:44,960 --> 00:19:48,960
the

487
00:19:45,760 --> 00:19:51,440
uh to the data frame as well um but this

488
00:19:48,960 --> 00:19:52,559
time of course using python and not

489
00:19:51,440 --> 00:19:54,799
pandas

490
00:19:52,559 --> 00:19:58,480
so as you can guess this is probably

491
00:19:54,799 --> 00:20:00,000
going to be longer we're going to um

492
00:19:58,480 --> 00:20:01,520
just going to fill this up but we're

493
00:20:00,000 --> 00:20:02,799
going to

494
00:20:01,520 --> 00:20:04,400
um

495
00:20:02,799 --> 00:20:06,799
yeah we're just going to we're going to

496
00:20:04,400 --> 00:20:09,360
calculate this and also time it and see

497
00:20:06,799 --> 00:20:11,919
how long it's going to take

498
00:20:09,360 --> 00:20:14,640
for all of these loops to run let's just

499
00:20:11,919 --> 00:20:17,039
comment out that code

500
00:20:14,640 --> 00:20:20,280
and let's run it to see how long that's

501
00:20:17,039 --> 00:20:20,280
going to take

502
00:20:21,840 --> 00:20:25,440
so

503
00:20:23,440 --> 00:20:26,880
the loop here

504
00:20:25,440 --> 00:20:30,640
using

505
00:20:26,880 --> 00:20:32,720
python looping over took a full one

506
00:20:30,640 --> 00:20:35,039
second

507
00:20:32,720 --> 00:20:37,679
and 16 seconds it's just

508
00:20:35,039 --> 00:20:39,039
do with both that's a lot that's a big

509
00:20:37,679 --> 00:20:41,039
difference

510
00:20:39,039 --> 00:20:42,000
um that's a really significant

511
00:20:41,039 --> 00:20:43,840
difference

512
00:20:42,000 --> 00:20:47,679
so as you can see

513
00:20:43,840 --> 00:20:51,440
it's possible to get a speed up of about

514
00:20:47,679 --> 00:20:54,000
500 times just by switching to this

515
00:20:51,440 --> 00:20:55,520
library this probably explains why most

516
00:20:54,000 --> 00:20:57,520
data scientists choose to work with

517
00:20:55,520 --> 00:20:59,520
pandas as compared to writing their code

518
00:20:57,520 --> 00:21:01,520
in pure python

519
00:20:59,520 --> 00:21:04,640
now even though you'll get a significant

520
00:21:01,520 --> 00:21:06,080
speed up just by switching to this sort

521
00:21:04,640 --> 00:21:09,200
of workflow

522
00:21:06,080 --> 00:21:12,960
from python to pandas making use of a

523
00:21:09,200 --> 00:21:15,679
gpu especially for larger datasets can

524
00:21:12,960 --> 00:21:18,240
take things even further than earlier i

525
00:21:15,679 --> 00:21:20,880
have been interning as a software

526
00:21:18,240 --> 00:21:23,600
engineer with the rapids team at nvidia

527
00:21:20,880 --> 00:21:25,360
and i've really enjoyed working uh with

528
00:21:23,600 --> 00:21:28,240
the team there i've enjoyed working on

529
00:21:25,360 --> 00:21:33,440
this project um qdf is an open source

530
00:21:28,240 --> 00:21:34,400
python based gpu data frame library and

531
00:21:33,440 --> 00:21:36,400
for

532
00:21:34,400 --> 00:21:38,080
data scientists pythonistas that are

533
00:21:36,400 --> 00:21:40,159
wanting to work with data

534
00:21:38,080 --> 00:21:42,480
it allows you to do a lot of those

535
00:21:40,159 --> 00:21:44,000
operations like loading or joining

536
00:21:42,480 --> 00:21:45,760
aggregating and filtering and different

537
00:21:44,000 --> 00:21:48,000
things like that

538
00:21:45,760 --> 00:21:50,640
but it makes it possible for you to do

539
00:21:48,000 --> 00:21:51,600
those computations on the gpu

540
00:21:50,640 --> 00:21:52,559
and

541
00:21:51,600 --> 00:21:54,720
how

542
00:21:52,559 --> 00:21:58,640
the team has designed this

543
00:21:54,720 --> 00:22:00,960
qdf is to really mirror the panda's api

544
00:21:58,640 --> 00:22:02,400
so a lot of the things that if you can

545
00:22:00,960 --> 00:22:05,039
do something

546
00:22:02,400 --> 00:22:08,159
in pandas you should be able to do that

547
00:22:05,039 --> 00:22:11,039
same thing in korea so

548
00:22:08,159 --> 00:22:14,159
i think as well because the apis are so

549
00:22:11,039 --> 00:22:16,400
similar in how they've been written um

550
00:22:14,159 --> 00:22:18,640
it's really intuitive to make a switch

551
00:22:16,400 --> 00:22:21,840
so i'm going to go ahead and show you an

552
00:22:18,640 --> 00:22:22,640
example of what it looks like

553
00:22:21,840 --> 00:22:23,520
right

554
00:22:22,640 --> 00:22:26,720
so

555
00:22:23,520 --> 00:22:29,840
here we would like to compare pandas and

556
00:22:26,720 --> 00:22:31,760
qdf and see the difference in

557
00:22:29,840 --> 00:22:33,520
speed for this we're going to be using

558
00:22:31,760 --> 00:22:35,679
time it

559
00:22:33,520 --> 00:22:37,840
to just see how

560
00:22:35,679 --> 00:22:40,000
how long it takes for pandas to do some

561
00:22:37,840 --> 00:22:43,440
calculations and how long it takes for

562
00:22:40,000 --> 00:22:44,799
qdf to do calculations so as you can see

563
00:22:43,440 --> 00:22:46,880
um

564
00:22:44,799 --> 00:22:49,039
right now at the top we are creating a

565
00:22:46,880 --> 00:22:51,760
pandas data frame and this data frame

566
00:22:49,039 --> 00:22:54,159
has significantly more data than last

567
00:22:51,760 --> 00:22:56,960
time and then i want to show just how

568
00:22:54,159 --> 00:22:59,120
easy it is to convert a pandas data

569
00:22:56,960 --> 00:23:01,840
frame to qdf

570
00:22:59,120 --> 00:23:04,080
and uh it's it's a very similar

571
00:23:01,840 --> 00:23:06,240
instruction and then

572
00:23:04,080 --> 00:23:09,440
you're using the method from pandas and

573
00:23:06,240 --> 00:23:12,080
just putting in um your original pandas

574
00:23:09,440 --> 00:23:13,120
data frame you can also create your own

575
00:23:12,080 --> 00:23:15,520
um

576
00:23:13,120 --> 00:23:18,320
you can also create your own

577
00:23:15,520 --> 00:23:20,080
qdf data frames from scratch if you uh

578
00:23:18,320 --> 00:23:24,720
if you'd like to as well

579
00:23:20,080 --> 00:23:27,200
and so we are going to start off by

580
00:23:24,720 --> 00:23:28,080
you know using pandas to calculate the

581
00:23:27,200 --> 00:23:31,440
mean

582
00:23:28,080 --> 00:23:33,280
of the data that is in our pandas data

583
00:23:31,440 --> 00:23:35,600
frame and we're going to do the same

584
00:23:33,280 --> 00:23:38,080
thing as well with qdf and we're going

585
00:23:35,600 --> 00:23:40,960
to time both of these

586
00:23:38,080 --> 00:23:44,000
operations and see how long it takes um

587
00:23:40,960 --> 00:23:47,840
to do this with pandas and then to see

588
00:23:44,000 --> 00:23:48,799
how long it takes to do this with kudith

589
00:23:47,840 --> 00:23:50,400
so

590
00:23:48,799 --> 00:23:52,159
go ahead and run this

591
00:23:50,400 --> 00:23:55,279
and see

592
00:23:52,159 --> 00:23:55,279
what the time takes

593
00:23:56,720 --> 00:24:01,400
so it takes um

594
00:23:59,039 --> 00:24:05,880
that's great it takes about

595
00:24:01,400 --> 00:24:09,840
156 milliseconds per loop for pandas and

596
00:24:05,880 --> 00:24:13,840
3.56 milliseconds per loop for

597
00:24:09,840 --> 00:24:16,840
qdf which is really significant

598
00:24:13,840 --> 00:24:21,039
so i did want to also show you a second

599
00:24:16,840 --> 00:24:22,480
calculation um which is just merging um

600
00:24:21,039 --> 00:24:23,760
data together

601
00:24:22,480 --> 00:24:25,840
and

602
00:24:23,760 --> 00:24:28,240
i ran out of ram

603
00:24:25,840 --> 00:24:32,720
to do this on google collab because it

604
00:24:28,240 --> 00:24:33,440
is a much larger computation to do and

605
00:24:32,720 --> 00:24:35,120
so

606
00:24:33,440 --> 00:24:37,039
instead of trying to do it here we're

607
00:24:35,120 --> 00:24:38,240
just going to move over to

608
00:24:37,039 --> 00:24:39,120
um

609
00:24:38,240 --> 00:24:41,360
to

610
00:24:39,120 --> 00:24:44,640
a jupiter notebook that

611
00:24:41,360 --> 00:24:47,679
is using my own local gpu

612
00:24:44,640 --> 00:24:50,400
so here we are in our jeep to notebook

613
00:24:47,679 --> 00:24:51,279
and uh behind the scenes we are running

614
00:24:50,400 --> 00:24:52,559
a

615
00:24:51,279 --> 00:24:55,840
gpu

616
00:24:52,559 --> 00:24:58,320
and so we are still able to access our

617
00:24:55,840 --> 00:25:00,320
cdf gpu data frame

618
00:24:58,320 --> 00:25:02,080
we have a similar setup from what we had

619
00:25:00,320 --> 00:25:04,159
last time we're importing our libraries

620
00:25:02,080 --> 00:25:05,200
and creating the same data frame with

621
00:25:04,159 --> 00:25:07,120
pandas

622
00:25:05,200 --> 00:25:10,159
this time we're trying to merge the data

623
00:25:07,120 --> 00:25:13,440
together at first with pandas and then

624
00:25:10,159 --> 00:25:15,360
um with qdf so i'm just running that

625
00:25:13,440 --> 00:25:17,840
and

626
00:25:15,360 --> 00:25:19,919
yeah so i had to this was this took

627
00:25:17,840 --> 00:25:23,440
quite a long time to run

628
00:25:19,919 --> 00:25:24,799
particularly with pandas so i had to

629
00:25:23,440 --> 00:25:26,480
speed things up

630
00:25:24,799 --> 00:25:28,559
so i'm just going to jump forward a

631
00:25:26,480 --> 00:25:31,120
little bit

632
00:25:28,559 --> 00:25:32,720
all right so it did take quite a long

633
00:25:31,120 --> 00:25:35,360
time to

634
00:25:32,720 --> 00:25:37,679
get through the calculation

635
00:25:35,360 --> 00:25:39,120
with pandas i actually cut out a whole

636
00:25:37,679 --> 00:25:40,960
section of time

637
00:25:39,120 --> 00:25:44,080
and and

638
00:25:40,960 --> 00:25:47,679
here pandas to go through all of it took

639
00:25:44,080 --> 00:25:49,559
about um one minute 10 seconds

640
00:25:47,679 --> 00:25:53,120
plus or minus

641
00:25:49,559 --> 00:25:55,440
486 milliseconds per loop

642
00:25:53,120 --> 00:25:58,000
which is quite a lot of time

643
00:25:55,440 --> 00:26:00,960
to go through all of those loops

644
00:25:58,000 --> 00:26:03,039
it was it was a lot of data so it you

645
00:26:00,960 --> 00:26:05,840
know pandas was quite slow

646
00:26:03,039 --> 00:26:09,120
but next we can just let it go through

647
00:26:05,840 --> 00:26:12,400
and calculate and as you can see

648
00:26:09,120 --> 00:26:14,320
for um kudif it was much faster it took

649
00:26:12,400 --> 00:26:17,440
about 367

650
00:26:14,320 --> 00:26:19,840
milliseconds um to calculate to

651
00:26:17,440 --> 00:26:23,200
calculate that for per loop

652
00:26:19,840 --> 00:26:25,919
um which is a lot faster so definitely

653
00:26:23,200 --> 00:26:30,480
that's a fantastic way to get speed ups

654
00:26:25,919 --> 00:26:32,480
and from your python code but qdf is not

655
00:26:30,480 --> 00:26:35,600
the only way that

656
00:26:32,480 --> 00:26:38,559
python easters are using or can use gpus

657
00:26:35,600 --> 00:26:41,360
there are lots of fields where

658
00:26:38,559 --> 00:26:43,600
people using python can get speed ups in

659
00:26:41,360 --> 00:26:45,960
their python code one of those fields is

660
00:26:43,600 --> 00:26:49,039
mathematics you know we have in data

661
00:26:45,960 --> 00:26:51,679
visualization machine learning and ai

662
00:26:49,039 --> 00:26:54,320
gaming and graphics just lots of ways

663
00:26:51,679 --> 00:26:55,679
that python is can utilize gpus to get

664
00:26:54,320 --> 00:26:57,120
speed ups

665
00:26:55,679 --> 00:26:58,960
you might be watching this and thinking

666
00:26:57,120 --> 00:27:01,440
to yourself you know marlene

667
00:26:58,960 --> 00:27:03,200
i can't afford a gpu or i don't know

668
00:27:01,440 --> 00:27:06,159
where to get a gpu

669
00:27:03,200 --> 00:27:10,000
um you can definitely google it but um

670
00:27:06,159 --> 00:27:10,960
there are some great um resources online

671
00:27:10,000 --> 00:27:13,440
and

672
00:27:10,960 --> 00:27:15,840
right now i would say gpus are quite

673
00:27:13,440 --> 00:27:17,919
accessible online

674
00:27:15,840 --> 00:27:21,279
and in some cases you can actually get

675
00:27:17,919 --> 00:27:23,760
them for free or use them at a really

676
00:27:21,279 --> 00:27:24,880
good sort of cost rate

677
00:27:23,760 --> 00:27:27,200
and

678
00:27:24,880 --> 00:27:29,760
here on the screen i have some examples

679
00:27:27,200 --> 00:27:32,480
of places that you could access a gpu

680
00:27:29,760 --> 00:27:34,960
even right away i know for the rapids

681
00:27:32,480 --> 00:27:37,760
team we actually have a

682
00:27:34,960 --> 00:27:40,399
a google collab notebook on the website

683
00:27:37,760 --> 00:27:42,880
i'll put a link to the website there

684
00:27:40,399 --> 00:27:43,919
that you can actually take a look at and

685
00:27:42,880 --> 00:27:46,159
you can

686
00:27:43,919 --> 00:27:47,600
go to the colab notebook and actually

687
00:27:46,159 --> 00:27:50,559
try out the

688
00:27:47,600 --> 00:27:52,080
um the different libraries for free

689
00:27:50,559 --> 00:27:53,760
so definitely would encourage you to do

690
00:27:52,080 --> 00:27:56,720
that if you would like to

691
00:27:53,760 --> 00:27:59,200
um that is actually the end of my talk i

692
00:27:56,720 --> 00:28:02,960
hope you enjoyed listening to it

693
00:27:59,200 --> 00:28:05,440
if you have any questions feel free to

694
00:28:02,960 --> 00:28:07,120
let me know after this talk or

695
00:28:05,440 --> 00:28:10,640
feel free to also reach out to me on

696
00:28:07,120 --> 00:28:14,559
twitter i'm marlene underscores dmw

697
00:28:10,640 --> 00:28:14,559
thank you so much for for listening

698
00:28:15,440 --> 00:28:19,360
hello again uh reminder that the

699
00:28:17,679 --> 00:28:21,039
previous talk was recorded so there

700
00:28:19,360 --> 00:28:23,279
would be no questions

701
00:28:21,039 --> 00:28:25,520
but thank you marlene mangami anyway i

702
00:28:23,279 --> 00:28:26,880
mean thank you for presenting at

703
00:28:25,520 --> 00:28:28,559
pyconeyou

704
00:28:26,880 --> 00:28:31,440
uh there's a rest now there's a bit of a

705
00:28:28,559 --> 00:28:34,880
break for 45 minutes

706
00:28:31,440 --> 00:28:38,320
last block of the day will begin at

707
00:28:34,880 --> 00:28:40,640
4 15 australian eastern time

708
00:28:38,320 --> 00:28:44,520
so take a break maybe hydrate maybe

709
00:28:40,640 --> 00:28:44,520
stretch your legs see you then

710
00:28:46,640 --> 00:28:48,720
you