1
00:00:00,000 --> 00:00:08,460
hi unfortunately we had to cancel the

2
00:00:03,600 --> 00:00:10,919
the talk that we had planned for GDC but

3
00:00:08,460 --> 00:00:15,120
we do have some super exciting news

4
00:00:10,919 --> 00:00:17,160
about ps5 and who better to bring that

5
00:00:15,119 --> 00:00:18,929
to you than the one and only mark Cerny

6
00:00:17,160 --> 00:00:24,390
without further ado over to you mark

7
00:00:18,929 --> 00:00:25,980
thank you Jim there will be lots of

8
00:00:24,390 --> 00:00:28,710
chances later on this year to look at

9
00:00:25,980 --> 00:00:30,329
the PlayStation 5 games today I want to

10
00:00:28,710 --> 00:00:32,279
talk a bit about our goals for the

11
00:00:30,329 --> 00:00:34,379
PlayStation 5 hardware and how they

12
00:00:32,279 --> 00:00:36,719
influenced the development of the

13
00:00:34,380 --> 00:00:38,730
console I think you all know I'm a big

14
00:00:36,719 --> 00:00:41,489
believer in console generations once

15
00:00:38,729 --> 00:00:43,019
every 5 or 6 or 7 years a console

16
00:00:41,488 --> 00:00:45,509
arrives with substantially new

17
00:00:43,020 --> 00:00:47,520
capabilities there's a lot of learning

18
00:00:45,509 --> 00:00:49,799
by the game developers hopefully not too

19
00:00:47,520 --> 00:00:52,559
overwhelming and soon there's games that

20
00:00:49,799 --> 00:00:54,659
could never have been created before now

21
00:00:52,558 --> 00:00:56,849
it used to be that as a console designer

22
00:00:54,659 --> 00:00:58,799
you'd somehow Intuit what would be the

23
00:00:56,850 --> 00:01:00,989
best set of capabilities for the new

24
00:00:58,799 --> 00:01:03,479
console and then build it in complete

25
00:01:00,988 --> 00:01:05,849
secrecy for the PlayStation consoles

26
00:01:03,479 --> 00:01:08,429
that period lasted through PlayStation 3

27
00:01:05,849 --> 00:01:10,769
a powerful and groundbreaking console

28
00:01:08,430 --> 00:01:12,540
but also one that caused quite a lot of

29
00:01:10,769 --> 00:01:15,929
heartache as it was initially difficult

30
00:01:12,540 --> 00:01:17,640
to develop games for so starting with

31
00:01:15,930 --> 00:01:19,650
PlayStation 4 we've taken a different

32
00:01:17,640 --> 00:01:22,259
approach roughly centered around three

33
00:01:19,650 --> 00:01:24,118
principles the first of these is

34
00:01:22,259 --> 00:01:26,519
listening to the developers which is to

35
00:01:24,118 --> 00:01:28,469
say that a lot of what we put into a

36
00:01:26,519 --> 00:01:31,078
console derives directly from the needs

37
00:01:28,469 --> 00:01:33,060
and aspirations of the game creators we

38
00:01:31,078 --> 00:01:35,250
definitely do have some ideas of our own

39
00:01:33,060 --> 00:01:37,320
but at the core of our philosophy for

40
00:01:35,250 --> 00:01:40,109
designing consoles is that game players

41
00:01:37,319 --> 00:01:42,029
are here for the fantastic games which

42
00:01:40,109 --> 00:01:45,060
is to say that game creators matter

43
00:01:42,030 --> 00:01:46,890
anything we can do to make life easier

44
00:01:45,060 --> 00:01:50,430
for the game creators or help them

45
00:01:46,890 --> 00:01:52,829
realize their dreams we will do so about

46
00:01:50,430 --> 00:01:54,479
once every two years I take a tour of

47
00:01:52,828 --> 00:01:56,849
the industry I go to the various

48
00:01:54,478 --> 00:01:58,349
developers and publishers sit down and

49
00:01:56,849 --> 00:01:59,849
discuss how they're doing with the

50
00:01:58,349 --> 00:02:02,578
current consoles and what they'd like to

51
00:01:59,849 --> 00:02:04,769
see in future consoles this requires

52
00:02:02,578 --> 00:02:06,658
weeks on the road as reaching the bulk

53
00:02:04,769 --> 00:02:08,008
of the game creators involves talking to

54
00:02:06,659 --> 00:02:10,110
well over a hundred people it's

55
00:02:08,008 --> 00:02:13,669
something like two dozen publishers and

56
00:02:10,110 --> 00:02:17,220
developers and it is incredibly valuable

57
00:02:13,669 --> 00:02:20,369
by the way the feature most requested by

58
00:02:17,219 --> 00:02:22,319
the developers that was an SST which we

59
00:02:20,370 --> 00:02:24,990
were very happy to put in hardware but a

60
00:02:22,319 --> 00:02:27,239
lot of problem solving was required I'll

61
00:02:24,990 --> 00:02:29,430
be doing a deep dive on the SSD and

62
00:02:27,240 --> 00:02:32,450
surrounding systems later on in this

63
00:02:29,430 --> 00:02:34,500
talk it's also key to make a

64
00:02:32,449 --> 00:02:36,479
generational leaf while keeping the

65
00:02:34,500 --> 00:02:38,400
console sufficiently familiar to game

66
00:02:36,479 --> 00:02:40,819
developers I think about this in terms

67
00:02:38,400 --> 00:02:43,860
of balancing evolution and revolution

68
00:02:40,819 --> 00:02:45,840
now with Playstation one two and three

69
00:02:43,860 --> 00:02:48,150
the target was a revolution each time

70
00:02:45,840 --> 00:02:50,099
with a brand new feature set that was

71
00:02:48,150 --> 00:02:51,599
great in many ways but time for the

72
00:02:50,099 --> 00:02:53,789
developers to get up and running got

73
00:02:51,599 --> 00:02:56,939
longer with each console in the past

74
00:02:53,789 --> 00:02:59,489
I've called this time to triangle here's

75
00:02:56,939 --> 00:03:01,349
what I had for those three consoles to

76
00:02:59,490 --> 00:03:03,780
be clear I'm not talking about time to

77
00:03:01,349 --> 00:03:06,000
make a game developers will be ambitious

78
00:03:03,780 --> 00:03:08,460
and it may take them six years or so to

79
00:03:06,000 --> 00:03:10,919
realize their vision what I'm talking

80
00:03:08,460 --> 00:03:12,689
about is that dead time before graphics

81
00:03:10,919 --> 00:03:14,789
and other aspects of game development

82
00:03:12,689 --> 00:03:17,699
are up and running and trying to

83
00:03:14,789 --> 00:03:19,530
minimize that on the other hand if we're

84
00:03:17,699 --> 00:03:21,269
trying to reduce that dead time to 0

85
00:03:19,530 --> 00:03:22,800
that means the hardware architecture

86
00:03:21,270 --> 00:03:25,469
can't change at all

87
00:03:22,800 --> 00:03:27,660
we're handcuffed we need to judge for

88
00:03:25,469 --> 00:03:29,310
each feature what value it adds and

89
00:03:27,659 --> 00:03:31,549
whether it's worth the increase in

90
00:03:29,310 --> 00:03:34,110
developer time needed to support it so

91
00:03:31,550 --> 00:03:35,580
with Playstation 4 we were able to

92
00:03:34,110 --> 00:03:37,680
strike a pretty good balance between

93
00:03:35,580 --> 00:03:39,900
performance and familiarity we got

94
00:03:37,680 --> 00:03:43,770
required learning back to PlayStation 1

95
00:03:39,900 --> 00:03:45,450
levels with PS 5 the GPU was definitely

96
00:03:43,770 --> 00:03:48,090
the area that we felt the most tension

97
00:03:45,449 --> 00:03:50,849
between adding new features and keeping

98
00:03:48,090 --> 00:03:52,379
a familiar programming model ultimately

99
00:03:50,849 --> 00:03:54,269
I think we've ended up with something

100
00:03:52,379 --> 00:03:56,459
under a month of getting up to speed

101
00:03:54,270 --> 00:03:58,290
that feels like we're striking about the

102
00:03:56,459 --> 00:04:00,090
right balance I'll go into a bit more

103
00:03:58,289 --> 00:04:02,789
detail later today about our philosophy

104
00:04:00,090 --> 00:04:06,689
with the GPU and the specific feature

105
00:04:02,789 --> 00:04:09,060
set that resulted from it it's also very

106
00:04:06,689 --> 00:04:10,919
important for us as the hardware team to

107
00:04:09,060 --> 00:04:13,890
find new dreams by which I mean

108
00:04:10,919 --> 00:04:16,469
something other than CPU performance GPU

109
00:04:13,889 --> 00:04:18,298
performance and the amount of RAM the

110
00:04:16,470 --> 00:04:19,440
increase in graphics performance over

111
00:04:18,298 --> 00:04:21,719
the past two decades has been

112
00:04:19,439 --> 00:04:23,219
astonishing but there are other areas in

113
00:04:21,720 --> 00:04:25,169
which we can innovate and provide a

114
00:04:23,220 --> 00:04:26,330
significant value to the game creators

115
00:04:25,168 --> 00:04:29,089
and through them

116
00:04:26,329 --> 00:04:30,589
the players that's why the SSD was very

117
00:04:29,089 --> 00:04:32,628
much on our list of directions to

118
00:04:30,589 --> 00:04:34,638
explore regardless of what came out of

119
00:04:32,629 --> 00:04:37,610
the conversations with game developers

120
00:04:34,639 --> 00:04:39,829
and publishers the biggest feature in

121
00:04:37,610 --> 00:04:42,288
this category is the custom engine for

122
00:04:39,829 --> 00:04:45,050
audio that's today's final topic the

123
00:04:42,288 --> 00:04:47,658
push for vastly improved audio and in

124
00:04:45,050 --> 00:04:49,129
particular 3d audio isn't something that

125
00:04:47,658 --> 00:04:51,168
came out of the developer meetings it's

126
00:04:49,129 --> 00:04:53,810
much more the case that we had a dream

127
00:04:51,168 --> 00:04:55,818
of what might be possible five years

128
00:04:53,810 --> 00:04:57,769
from now and then worked out a number of

129
00:04:55,819 --> 00:05:00,468
steps we could take to set us on that

130
00:04:57,769 --> 00:05:02,870
path so here again are the three

131
00:05:00,468 --> 00:05:05,120
principles the first being enabling the

132
00:05:02,870 --> 00:05:08,689
desires of developers to drive the

133
00:05:05,120 --> 00:05:10,728
hardware design to me the SSD really is

134
00:05:08,689 --> 00:05:13,430
the key to the next generation it's a

135
00:05:10,728 --> 00:05:15,709
game changer and it was the number one

136
00:05:13,430 --> 00:05:18,590
ask from developers for PlayStation 5 as

137
00:05:15,709 --> 00:05:22,008
in we know it's probably impossible but

138
00:05:18,589 --> 00:05:23,299
can you put an SSD in it that was a

139
00:05:22,009 --> 00:05:25,098
discussion we were also having

140
00:05:23,300 --> 00:05:26,569
internally it was clear that the

141
00:05:25,098 --> 00:05:28,550
presence of a hard drive in every

142
00:05:26,569 --> 00:05:30,949
PlayStation 4 was having a positive

143
00:05:28,550 --> 00:05:33,199
impact a lot of things that would simply

144
00:05:30,949 --> 00:05:35,449
have been impossible at blu-ray disc

145
00:05:33,199 --> 00:05:38,389
speeds were now possible at the same

146
00:05:35,449 --> 00:05:39,430
time though in 2015 and 2016 when we

147
00:05:38,389 --> 00:05:41,810
were having these conversations

148
00:05:39,430 --> 00:05:43,370
developers were already banging up

149
00:05:41,810 --> 00:05:45,649
against the limits of the hard drive and

150
00:05:43,370 --> 00:05:50,209
a lot of developer time was being spent

151
00:05:45,649 --> 00:05:52,609
designing around slow load speeds I want

152
00:05:50,209 --> 00:05:54,948
to focus in on just one number here

153
00:05:52,610 --> 00:05:57,740
which is how long it takes to load a

154
00:05:54,949 --> 00:05:59,810
gigabyte of data from a hard drive the

155
00:05:57,740 --> 00:06:02,150
difficulty being that hard drives are

156
00:05:59,810 --> 00:06:04,848
neither particularly fast nor flexible

157
00:06:02,149 --> 00:06:07,370
if all your data is in one block which

158
00:06:04,848 --> 00:06:09,860
is frankly not very likely you can load

159
00:06:07,370 --> 00:06:11,899
50 to 100 megabytes a second depending

160
00:06:09,860 --> 00:06:13,310
on where the data is located on the hard

161
00:06:11,899 --> 00:06:15,049
drive let's assume it's on the outer

162
00:06:13,310 --> 00:06:17,389
edge which means loading a gigabyte

163
00:06:15,050 --> 00:06:19,728
takes 10 seconds if you compress your

164
00:06:17,389 --> 00:06:21,829
game packages you can fit more data on

165
00:06:19,728 --> 00:06:24,228
the blu-ray disc and also effectively

166
00:06:21,829 --> 00:06:26,990
boost your hard drive read speed by the

167
00:06:24,228 --> 00:06:28,939
compression ratio we support Z Lib

168
00:06:26,990 --> 00:06:31,129
decompression on PlayStation 4 that gets

169
00:06:28,939 --> 00:06:34,838
you something like 50% more data on the

170
00:06:31,129 --> 00:06:37,099
disk and 50% higher effective read speed

171
00:06:34,838 --> 00:06:39,408
unfortunately though it's highly likely

172
00:06:37,098 --> 00:06:41,569
that your data is scattered around in

173
00:06:39,408 --> 00:06:43,550
files on the hard drive as well as

174
00:06:41,569 --> 00:06:46,309
sourced from multiple locations within

175
00:06:43,550 --> 00:06:49,699
those files so lots of Sikhs are needed

176
00:06:46,309 --> 00:06:51,588
at two to fifty ish milliseconds each my

177
00:06:49,699 --> 00:06:53,478
rule of thumb is that the hard drive is

178
00:06:51,588 --> 00:06:56,478
spending two-thirds of its time seeking

179
00:06:53,478 --> 00:06:58,639
and only a third of its time actually

180
00:06:56,478 --> 00:07:01,428
loading data putting all of that

181
00:06:58,639 --> 00:07:05,059
together a gigabyte is very roughly 20

182
00:07:01,428 --> 00:07:07,638
seconds to load from a hard drive now a

183
00:07:05,059 --> 00:07:09,800
gigabyte is not much data games are

184
00:07:07,639 --> 00:07:12,019
using five or six gigabytes of RAM on

185
00:07:09,800 --> 00:07:14,449
PlayStation 4 so boot times and load

186
00:07:12,019 --> 00:07:16,819
tons can get pretty grim or to put that

187
00:07:14,449 --> 00:07:19,389
differently as a player you wait for the

188
00:07:16,819 --> 00:07:22,399
game to boot wait for the game to load

189
00:07:19,389 --> 00:07:24,110
wait for the level to reload every time

190
00:07:22,399 --> 00:07:25,399
you die and you wait for what is

191
00:07:24,110 --> 00:07:28,819
euphemistically called

192
00:07:25,399 --> 00:07:31,218
fast travel and all of that leads to the

193
00:07:28,819 --> 00:07:34,309
dream what if we could have not just an

194
00:07:31,218 --> 00:07:37,129
SSD but a blindingly fast Assisting if

195
00:07:34,309 --> 00:07:40,219
we could load five gigabytes a second

196
00:07:37,129 --> 00:07:41,539
from it what would change now SSDs are

197
00:07:40,218 --> 00:07:42,769
completely different from hard drives

198
00:07:41,538 --> 00:07:45,259
they don't have sikhs

199
00:07:42,769 --> 00:07:47,449
as such if you have a five gigabyte a

200
00:07:45,259 --> 00:07:49,218
second SSD you can read data from a

201
00:07:47,449 --> 00:07:52,699
thousand different locations in that

202
00:07:49,218 --> 00:07:54,889
second pretty much at speed as for time

203
00:07:52,699 --> 00:07:56,929
to load a gigabyte this is next-gen

204
00:07:54,889 --> 00:07:59,149
we're talking about so memory is bigger

205
00:07:56,928 --> 00:08:01,788
instead we should be asking how long to

206
00:07:59,149 --> 00:08:04,519
load two gigabytes and the answer is

207
00:08:01,788 --> 00:08:06,528
about a quarter of a second that's

208
00:08:04,519 --> 00:08:10,218
amazing we're talking two orders of

209
00:08:06,528 --> 00:08:13,158
magnitude meaning very roughly 100 times

210
00:08:10,218 --> 00:08:15,079
faster which means at five gigabytes a

211
00:08:13,158 --> 00:08:18,019
second for the SSD the potential is that

212
00:08:15,079 --> 00:08:20,149
the game boots in a second there are no

213
00:08:18,019 --> 00:08:22,278
load screens the game just fades down

214
00:08:20,149 --> 00:08:24,559
while loading a half dozen gigabytes and

215
00:08:22,278 --> 00:08:26,449
fades back up again same for a reload

216
00:08:24,559 --> 00:08:29,449
you're immediately back in the action

217
00:08:26,449 --> 00:08:31,759
after you die and fast travel becomes so

218
00:08:29,449 --> 00:08:33,740
fast it's blinking you miss it as game

219
00:08:31,759 --> 00:08:35,629
creators we go from trying to distract

220
00:08:33,740 --> 00:08:37,818
the player from how long fast travel is

221
00:08:35,629 --> 00:08:40,519
taking like those spider-man subway

222
00:08:37,818 --> 00:08:42,078
rides to being so blindingly fast that

223
00:08:40,519 --> 00:08:45,289
we might even have to slow that

224
00:08:42,078 --> 00:08:47,808
transition down pretty cool right but

225
00:08:45,289 --> 00:08:50,269
for me this is not the primary reason to

226
00:08:47,808 --> 00:08:52,339
change from a hard drive to an SSD the

227
00:08:50,269 --> 00:08:54,889
primary reason for an ultra

228
00:08:52,340 --> 00:08:57,769
SST is that it gives the game designer

229
00:08:54,889 --> 00:09:00,049
freedom or to put that differently with

230
00:08:57,769 --> 00:09:02,299
a hard drive the twenty seconds that it

231
00:09:00,049 --> 00:09:04,879
takes to load a gigabyte can sabotage

232
00:09:02,299 --> 00:09:07,759
the game that the developer is trying to

233
00:09:04,879 --> 00:09:09,500
create I think almost all of us in the

234
00:09:07,759 --> 00:09:11,569
room have experienced this maybe in

235
00:09:09,500 --> 00:09:13,220
different ways say we're making an

236
00:09:11,570 --> 00:09:15,230
adventure game and we have two rich

237
00:09:13,220 --> 00:09:17,840
environments where we each want enough

238
00:09:15,230 --> 00:09:19,370
textures and models to fill memory what

239
00:09:17,840 --> 00:09:22,220
you can do as long as you have a long

240
00:09:19,370 --> 00:09:23,870
staircase or elevator ride or a windy

241
00:09:22,220 --> 00:09:26,509
corridor where you can ditch the old

242
00:09:23,870 --> 00:09:29,659
assets and then take 30 seconds or so to

243
00:09:26,509 --> 00:09:31,819
load the new assets having a 30 second

244
00:09:29,659 --> 00:09:33,829
elevator ride is a little extreme

245
00:09:31,820 --> 00:09:35,900
more realistically we'd probably chop

246
00:09:33,830 --> 00:09:37,820
the world into a number of smaller

247
00:09:35,899 --> 00:09:40,699
pieces and then do some calculations

248
00:09:37,820 --> 00:09:42,920
with sightlines and run speeds like we

249
00:09:40,700 --> 00:09:45,800
did for Haven city when we were making

250
00:09:42,919 --> 00:09:48,379
jak 2 the game is 20 years old but not

251
00:09:45,799 --> 00:09:50,629
much has changed since then all those

252
00:09:48,379 --> 00:09:52,399
twisty passages are there for a reason

253
00:09:50,629 --> 00:09:54,919
there's a whole subset of level design

254
00:09:52,399 --> 00:09:57,470
dedicated to this sort of work but still

255
00:09:54,919 --> 00:10:01,069
it's a giant distraction for a team that

256
00:09:57,470 --> 00:10:03,710
just wants to make their game so when I

257
00:10:01,070 --> 00:10:05,780
talked about the dream of an SSD part of

258
00:10:03,710 --> 00:10:08,480
the reason for that 5 gigabyte a second

259
00:10:05,779 --> 00:10:10,339
target was to eliminate loads but also

260
00:10:08,480 --> 00:10:14,149
part of the reason for that target was

261
00:10:10,340 --> 00:10:16,580
streaming as in what if the SSD is so

262
00:10:14,149 --> 00:10:18,409
fast that as the player is turning

263
00:10:16,580 --> 00:10:20,300
around it's possible to load textures

264
00:10:18,409 --> 00:10:22,850
for everything behind the player in that

265
00:10:20,299 --> 00:10:24,620
split second if you figure that it takes

266
00:10:22,850 --> 00:10:26,389
half a second to turn that's four

267
00:10:24,620 --> 00:10:28,820
gigabytes of compressed data you can

268
00:10:26,389 --> 00:10:31,100
load that sounds about right for next

269
00:10:28,820 --> 00:10:32,570
gen anyway back to the hard drive

270
00:10:31,100 --> 00:10:34,610
another strategy for increasing

271
00:10:32,570 --> 00:10:37,310
effective read speed is to make big

272
00:10:34,610 --> 00:10:39,379
sequential chunks of data for example we

273
00:10:37,309 --> 00:10:41,629
might group all the data together for

274
00:10:39,379 --> 00:10:44,090
each city block that removes most of the

275
00:10:41,629 --> 00:10:45,980
seeks and the streaming gets faster but

276
00:10:44,090 --> 00:10:48,230
there's a downside too which is that

277
00:10:45,980 --> 00:10:49,879
frequently used data is included in many

278
00:10:48,230 --> 00:10:51,759
chunks and therefore is on the hard

279
00:10:49,879 --> 00:10:54,409
drive many many times

280
00:10:51,759 --> 00:10:56,120
Marvel spider-men uses this strategy and

281
00:10:54,409 --> 00:10:58,699
though it works very well for increasing

282
00:10:56,120 --> 00:11:00,860
the streaming speed there's a massive

283
00:10:58,700 --> 00:11:03,259
duplication as a result some of the

284
00:11:00,860 --> 00:11:04,550
objects like mailboxes or newsracks are

285
00:11:03,259 --> 00:11:07,669
on the hard drive for

286
00:11:04,549 --> 00:11:09,049
hundred times what I'm describing here

287
00:11:07,669 --> 00:11:11,509
are things that cramp

288
00:11:09,049 --> 00:11:13,039
a creative director style either level

289
00:11:11,509 --> 00:11:15,588
design gets a little bit boring in

290
00:11:13,039 --> 00:11:17,240
places or the data is duplicated so many

291
00:11:15,589 --> 00:11:19,550
times that it no longer fits on the

292
00:11:17,240 --> 00:11:21,350
blu-ray disk and you end up with hard

293
00:11:19,549 --> 00:11:23,689
limits on the players run speed or

294
00:11:21,350 --> 00:11:25,670
driving speed the player can't go faster

295
00:11:23,690 --> 00:11:28,370
than the load speed from the hard drive

296
00:11:25,669 --> 00:11:30,500
and finally I'm sure many of you have

297
00:11:28,370 --> 00:11:32,389
noticed that after a patch download the

298
00:11:30,500 --> 00:11:35,600
PlayStation 4 will sometimes take a long

299
00:11:32,389 --> 00:11:37,338
time to install the patch that's because

300
00:11:35,600 --> 00:11:39,019
when just part of a file has been

301
00:11:37,339 --> 00:11:41,779
changed the new data can be downloaded

302
00:11:39,019 --> 00:11:43,700
pretty quickly but before the game boots

303
00:11:41,778 --> 00:11:45,409
up a brand new file has to be

304
00:11:43,700 --> 00:11:47,959
constructed that includes the changed

305
00:11:45,409 --> 00:11:50,899
portion otherwise every change would add

306
00:11:47,958 --> 00:11:52,909
a seek or two even so you can

307
00:11:50,899 --> 00:11:55,100
occasionally see this happening on game

308
00:11:52,909 --> 00:11:58,219
titles they start to hitch once they get

309
00:11:55,100 --> 00:12:00,889
patched enough with an SSD though no

310
00:11:58,220 --> 00:12:02,600
seeks so no need to make brand new files

311
00:12:00,889 --> 00:12:05,659
with the changes incorporated into them

312
00:12:02,600 --> 00:12:09,950
which means no installs as you know them

313
00:12:05,659 --> 00:12:12,049
today there's yet one more benefit which

314
00:12:09,950 --> 00:12:15,350
is that system memory can be used much

315
00:12:12,049 --> 00:12:17,000
more efficiently on PlayStation 4 game

316
00:12:15,350 --> 00:12:19,579
data on the hard drive feels very

317
00:12:17,000 --> 00:12:21,169
distant and difficult to use by the time

318
00:12:19,578 --> 00:12:23,239
you realize you need a piece of data

319
00:12:21,169 --> 00:12:25,250
it's much too late to go out and load it

320
00:12:23,240 --> 00:12:27,139
so system memory has to contain all of

321
00:12:25,250 --> 00:12:30,019
the data that could be used in the next

322
00:12:27,139 --> 00:12:32,028
30 seconds or so of gameplay that means

323
00:12:30,019 --> 00:12:34,399
a lot of the eight gigabytes of system

324
00:12:32,028 --> 00:12:38,899
memory is idle it's just waiting there

325
00:12:34,399 --> 00:12:40,970
to be potentially used on PlayStation 5

326
00:12:38,899 --> 00:12:43,250
though the SSD is very close to being

327
00:12:40,970 --> 00:12:45,019
like more RAM typically it's fast enough

328
00:12:43,250 --> 00:12:47,179
that when you realize you need a piece

329
00:12:45,019 --> 00:12:49,698
of data you can just load it from the

330
00:12:47,179 --> 00:12:51,588
SSD and use it there's no need to have

331
00:12:49,698 --> 00:12:53,769
lots of data parked in system memory

332
00:12:51,589 --> 00:12:56,029
waiting to potentially be used a

333
00:12:53,769 --> 00:12:58,698
different way of saying that is that

334
00:12:56,028 --> 00:13:01,338
most of Ram is working on the game's

335
00:12:58,698 --> 00:13:05,299
behalf this is one of the reasons that

336
00:13:01,339 --> 00:13:08,180
16 gigabytes of gddr5 for playstation 5

337
00:13:05,299 --> 00:13:10,759
the presence of the SSD reduces the need

338
00:13:08,179 --> 00:13:14,539
for a massive intergenerational increase

339
00:13:10,759 --> 00:13:16,439
in size so back to the dream of the SSD

340
00:13:14,539 --> 00:13:19,139
here's the set of targets

341
00:13:16,440 --> 00:13:22,529
buuut the game in a second no load

342
00:13:19,139 --> 00:13:25,649
screens design freedom meaning no twisty

343
00:13:22,528 --> 00:13:27,720
passages or long corridors more game on

344
00:13:25,649 --> 00:13:32,100
the disk and more game on the SSD and

345
00:13:27,720 --> 00:13:34,649
finally those patch installs go away the

346
00:13:32,100 --> 00:13:36,689
reality though is that the SSD is just

347
00:13:34,649 --> 00:13:38,519
one piece of the puzzle there's a lot of

348
00:13:36,688 --> 00:13:40,649
places where bottlenecks can occur in

349
00:13:38,519 --> 00:13:42,959
between the SSD and the game code that

350
00:13:40,649 --> 00:13:45,568
uses the data you can see this on

351
00:13:42,958 --> 00:13:47,369
PlayStation 4 if I use an SSD with 10

352
00:13:45,568 --> 00:13:50,458
times the speed of a standard hard drive

353
00:13:47,370 --> 00:13:54,060
I probably see only double the loading

354
00:13:50,458 --> 00:13:56,549
speed if that for PlayStation 5 our goal

355
00:13:54,059 --> 00:13:58,799
was not just that the SSD itself be a

356
00:13:56,549 --> 00:14:01,049
hundred times faster it was that game

357
00:13:58,799 --> 00:14:03,899
loads and streaming would be a hundred

358
00:14:01,049 --> 00:14:05,758
times faster so every single potential

359
00:14:03,899 --> 00:14:08,610
bottleneck needed to be addressed and

360
00:14:05,759 --> 00:14:10,050
there are a lot of them let's look at

361
00:14:08,610 --> 00:14:11,810
check-in and what happens when its

362
00:14:10,049 --> 00:14:14,490
overhead gets a hundred times larger

363
00:14:11,809 --> 00:14:16,708
conceptually check-in is a pretty simple

364
00:14:14,490 --> 00:14:18,720
process data is loaded into system

365
00:14:16,708 --> 00:14:21,208
memory from the hard drive or SSD it's

366
00:14:18,720 --> 00:14:23,278
examined a few values are tweaked to

367
00:14:21,208 --> 00:14:26,219
check it in and then it's moved to its

368
00:14:23,278 --> 00:14:28,948
final location at the SSD speeds we're

369
00:14:26,220 --> 00:14:30,689
talking about that last part moving the

370
00:14:28,948 --> 00:14:32,938
data meaning copying it from one

371
00:14:30,688 --> 00:14:36,088
location to another takes roughly an

372
00:14:32,938 --> 00:14:37,649
entire next-gen CPU core and that's just

373
00:14:36,089 --> 00:14:39,029
the tip of the iceberg if all the

374
00:14:37,649 --> 00:14:41,159
overheads get a hundred times larger

375
00:14:39,028 --> 00:14:42,990
that will cripple the framerate as soon

376
00:14:41,159 --> 00:14:45,870
as the player moves and that massive

377
00:14:42,990 --> 00:14:48,480
stream of data starts coming off the SSD

378
00:14:45,870 --> 00:14:51,269
so to solve all of that we built a lot

379
00:14:48,480 --> 00:14:53,399
of custom hardware namely a custom flash

380
00:14:51,269 --> 00:14:56,220
controller and a number of custom units

381
00:14:53,399 --> 00:14:58,559
in our main chip the flash controller in

382
00:14:56,220 --> 00:15:01,079
the SSD was designed for smooth and

383
00:14:58,559 --> 00:15:03,000
bottleneck free operation but also with

384
00:15:01,078 --> 00:15:04,919
games in mind for example there are six

385
00:15:03,000 --> 00:15:07,470
levels of priority when reading from the

386
00:15:04,919 --> 00:15:09,899
SSD priority is very important you can

387
00:15:07,470 --> 00:15:11,370
imagine the player heading into some new

388
00:15:09,899 --> 00:15:13,828
location in the world and the game

389
00:15:11,370 --> 00:15:15,899
requesting a few gigabytes of textures

390
00:15:13,828 --> 00:15:18,659
and while those textures are being

391
00:15:15,899 --> 00:15:21,389
loaded an enemy is shot and has to speak

392
00:15:18,659 --> 00:15:23,100
a few dying words having multiple

393
00:15:21,389 --> 00:15:25,730
priority levels let's the audio for

394
00:15:23,100 --> 00:15:28,949
those dying words get loaded immediately

395
00:15:25,730 --> 00:15:30,300
on one side that flash controller

396
00:15:28,948 --> 00:15:32,639
connects to the actual

397
00:15:30,299 --> 00:15:34,889
- dyes that supply of the storage to

398
00:15:32,639 --> 00:15:37,289
reach art bandwidth target of 5

399
00:15:34,889 --> 00:15:39,840
gigabytes a second we ended up with a 12

400
00:15:37,289 --> 00:15:42,209
channel interface 8 channels wouldn't be

401
00:15:39,840 --> 00:15:43,980
enough the resulting bandwidth we have

402
00:15:42,210 --> 00:15:46,620
achieved is actually five and a half

403
00:15:43,980 --> 00:15:48,810
gigabytes a second with a 12 channel

404
00:15:46,620 --> 00:15:53,159
interface the most natural size that

405
00:15:48,809 --> 00:15:56,039
emerges for an SSD is 825 gigabytes the

406
00:15:53,159 --> 00:15:58,169
key question for us was is that enough

407
00:15:56,039 --> 00:16:00,179
it's tempting to add more but flash

408
00:15:58,169 --> 00:16:02,370
certainly doesn't come cheap and we have

409
00:16:00,179 --> 00:16:04,349
a responsibility to our gaming audience

410
00:16:02,370 --> 00:16:07,529
to be cost effective with regards to

411
00:16:04,350 --> 00:16:09,240
what we put in the console ultimately we

412
00:16:07,529 --> 00:16:11,850
resolve this question by looking at the

413
00:16:09,240 --> 00:16:14,549
play patterns of a broad range of gamers

414
00:16:11,850 --> 00:16:15,840
we examined the specific games that they

415
00:16:14,549 --> 00:16:18,269
were playing over the course of a

416
00:16:15,840 --> 00:16:20,700
weekend or week or a month and whether

417
00:16:18,269 --> 00:16:23,490
that set of games would fit properly on

418
00:16:20,700 --> 00:16:25,800
the SSD we were able to establish that

419
00:16:23,490 --> 00:16:28,350
the friction caused by reinstalled or

420
00:16:25,799 --> 00:16:29,669
read downloads would be quite low and so

421
00:16:28,350 --> 00:16:32,879
we locked in on that

422
00:16:29,669 --> 00:16:34,949
825 gigabyte size while also preparing

423
00:16:32,879 --> 00:16:37,080
multiple strategies so that those who

424
00:16:34,950 --> 00:16:38,930
want more storage can add it I'll go

425
00:16:37,080 --> 00:16:41,400
through the details in a moment

426
00:16:38,929 --> 00:16:43,049
back to the flash controller on the

427
00:16:41,399 --> 00:16:45,870
other side it connects to our main

428
00:16:43,049 --> 00:16:48,750
custom chip via four lanes of Gen 4 PCIe

429
00:16:45,870 --> 00:16:51,560
and inside the main custom chip is a

430
00:16:48,750 --> 00:16:53,700
pretty hefty unit dedicated to i/o

431
00:16:51,559 --> 00:16:56,179
before we talk about what that does

432
00:16:53,700 --> 00:16:58,650
let's talk compression for a moment

433
00:16:56,179 --> 00:17:00,989
PlayStation 4 used Z live as its

434
00:16:58,649 --> 00:17:04,200
compression format we decided to use it

435
00:17:00,990 --> 00:17:05,970
again on PlayStation 5 but on my 2017

436
00:17:04,200 --> 00:17:09,360
tour of developers I learned about a new

437
00:17:05,970 --> 00:17:12,589
format called Kraken from rad game tools

438
00:17:09,359 --> 00:17:15,449
it's like Z libs smarter cousin simple

439
00:17:12,589 --> 00:17:18,209
similar types of algorithms but about

440
00:17:15,450 --> 00:17:20,910
10% better compression which is pretty

441
00:17:18,210 --> 00:17:24,090
big that means 10% more game on the UHD

442
00:17:20,910 --> 00:17:25,529
blu-ray disc or on the SSD kraken had

443
00:17:24,089 --> 00:17:27,929
only been out for a year that it was

444
00:17:25,529 --> 00:17:29,879
already becoming a de-facto industry

445
00:17:27,930 --> 00:17:31,799
standard half of the teams I talked to

446
00:17:29,880 --> 00:17:34,650
or either using it or getting ready to

447
00:17:31,799 --> 00:17:37,019
evaluate it so we hustled and built a

448
00:17:34,650 --> 00:17:40,019
custom decompressor into the i/o unit

449
00:17:37,019 --> 00:17:40,929
one capable of handling over 5 gigabytes

450
00:17:40,019 --> 00:17:43,778
of crack and

451
00:17:40,929 --> 00:17:45,519
format input data a second after

452
00:17:43,778 --> 00:17:47,558
decompression that typically becomes

453
00:17:45,519 --> 00:17:49,749
eight or nine gigabytes but the unit

454
00:17:47,558 --> 00:17:52,749
itself is capable of outputting as much

455
00:17:49,749 --> 00:17:54,778
as 22 gigabytes a second if the data

456
00:17:52,749 --> 00:17:57,639
happened to compress particularly well

457
00:17:54,778 --> 00:18:00,278
by the way in terms of performance that

458
00:17:57,638 --> 00:18:01,418
custom decompressor equates to nine of

459
00:18:00,278 --> 00:18:02,980
our Xen two cores

460
00:18:01,419 --> 00:18:04,749
that's what it would take to decompress

461
00:18:02,980 --> 00:18:08,288
the kraken stream with a conventional

462
00:18:04,749 --> 00:18:10,868
CPU there's a lot more in the custom IO

463
00:18:08,288 --> 00:18:12,819
unit including a dedicated DMA

464
00:18:10,868 --> 00:18:14,648
controller the game can direct exactly

465
00:18:12,819 --> 00:18:17,230
where it wants to send the data coming

466
00:18:14,648 --> 00:18:19,569
off of the SSD this equates to another

467
00:18:17,230 --> 00:18:21,759
Xen two core or two in terms of its copy

468
00:18:19,569 --> 00:18:24,908
performance its primary purpose is to

469
00:18:21,759 --> 00:18:27,249
remove check in as a bottleneck there's

470
00:18:24,909 --> 00:18:28,360
two dedicated i/o coprocessors in a

471
00:18:27,249 --> 00:18:29,919
large ramp pool

472
00:18:28,359 --> 00:18:31,569
these are tins in 2 cores there are

473
00:18:29,919 --> 00:18:34,629
there principally to direct the variety

474
00:18:31,569 --> 00:18:36,788
of custom hardware around them one of

475
00:18:34,628 --> 00:18:39,759
the co processors is dedicated to SSD

476
00:18:36,788 --> 00:18:41,558
i/o this lets us bypass traditional file

477
00:18:39,759 --> 00:18:44,259
i/o and it's bottlenecks when reading

478
00:18:41,558 --> 00:18:46,388
from the SSD the other is responsible

479
00:18:44,259 --> 00:18:48,128
for memory mapping which I know doesn't

480
00:18:46,388 --> 00:18:50,740
sound like anything related to the SSD

481
00:18:48,128 --> 00:18:53,618
but a lot of developers map and remap

482
00:18:50,740 --> 00:18:56,378
memory as part of file i/o and this too

483
00:18:53,618 --> 00:18:58,480
can become a bottleneck there are

484
00:18:56,378 --> 00:19:00,908
coherency engines to assist the

485
00:18:58,480 --> 00:19:03,190
coprocessors coherency comes up a lot in

486
00:19:00,909 --> 00:19:05,909
places probably the biggest coherency

487
00:19:03,190 --> 00:19:08,558
issue is stale data in the GPU caches

488
00:19:05,909 --> 00:19:11,289
flushing all of the GPU caches whenever

489
00:19:08,558 --> 00:19:13,599
the SSD is read is an unattractive

490
00:19:11,288 --> 00:19:15,700
option it could really hurt the GPU

491
00:19:13,599 --> 00:19:18,548
performance so we've implemented a

492
00:19:15,700 --> 00:19:21,519
gentler way of doing things where the

493
00:19:18,548 --> 00:19:23,950
coherency engines inform the GPU of the

494
00:19:21,519 --> 00:19:26,740
overwritten address ranges and custom

495
00:19:23,950 --> 00:19:29,319
scrubbers in several dozen GPU caches do

496
00:19:26,740 --> 00:19:31,749
pinpoint evictions of just those address

497
00:19:29,319 --> 00:19:33,668
ranges the best thing is as a game

498
00:19:31,749 --> 00:19:35,919
developer when you read from the SSD you

499
00:19:33,669 --> 00:19:37,570
don't need to know any of this you don't

500
00:19:35,919 --> 00:19:40,389
even need to know that your data is

501
00:19:37,569 --> 00:19:42,128
compressed you just indicate what data

502
00:19:40,388 --> 00:19:44,079
you'd like to read from your original

503
00:19:42,128 --> 00:19:45,878
uncompressed file and where you'd like

504
00:19:44,079 --> 00:19:48,849
to put it and the whole process of

505
00:19:45,878 --> 00:19:51,759
loading it happens invisibly to you and

506
00:19:48,849 --> 00:19:54,139
at very high speed back to the dream

507
00:19:51,759 --> 00:19:56,660
thanks to all of that surrounding hard

508
00:19:54,140 --> 00:19:58,520
our 5.5 gigabytes a second really should

509
00:19:56,660 --> 00:20:01,460
translate into something like a hundred

510
00:19:58,519 --> 00:20:04,009
times faster IO than ps4 and allow the

511
00:20:01,460 --> 00:20:06,259
dream of no load screens and superfast

512
00:20:04,009 --> 00:20:09,259
streaming to become a reality

513
00:20:06,259 --> 00:20:11,359
having said that expandability of our

514
00:20:09,259 --> 00:20:13,250
SSD is going to be quite important

515
00:20:11,359 --> 00:20:15,439
flash is costly and you may very well

516
00:20:13,250 --> 00:20:19,009
want to add storage to whatever we put

517
00:20:15,440 --> 00:20:20,960
in the console now the kind of storage

518
00:20:19,009 --> 00:20:22,940
you need depends on how you're going to

519
00:20:20,960 --> 00:20:25,009
use it if you have an extensive

520
00:20:22,940 --> 00:20:26,360
PlayStation 4 library and you'd like to

521
00:20:25,009 --> 00:20:28,039
take advantage of backwards

522
00:20:26,359 --> 00:20:30,918
compatibility to play those games on

523
00:20:28,039 --> 00:20:33,798
PlayStation 5 then a large external hard

524
00:20:30,919 --> 00:20:35,480
drive is ideal you can leave your games

525
00:20:33,798 --> 00:20:38,119
on the hard drive and play them directly

526
00:20:35,480 --> 00:20:40,190
from there thus saving the pricier SSD

527
00:20:38,119 --> 00:20:42,500
storage for your PlayStation 5 titles or

528
00:20:40,190 --> 00:20:44,990
you can copy your active PlayStation 4

529
00:20:42,500 --> 00:20:46,279
titles to the SSD if your purpose in

530
00:20:44,990 --> 00:20:49,160
adding more storage is to play

531
00:20:46,279 --> 00:20:52,039
playstation 5 titles though ideally you

532
00:20:49,160 --> 00:20:55,130
would add to your SSD storage we will be

533
00:20:52,039 --> 00:20:56,899
supporting certain m2 SSDs these are

534
00:20:55,130 --> 00:20:59,630
internal drives that you can get on the

535
00:20:56,900 --> 00:21:02,059
open market and install in a bay in the

536
00:20:59,630 --> 00:21:04,160
playstation 5 as for which ones we

537
00:21:02,058 --> 00:21:06,589
support and when I'll get to that in a

538
00:21:04,160 --> 00:21:09,769
moment they connect through the custom

539
00:21:06,589 --> 00:21:11,269
io unit just like our SSD does so they

540
00:21:09,769 --> 00:21:13,789
can take full advantage of the

541
00:21:11,269 --> 00:21:15,789
decompression io coprocessors and all

542
00:21:13,789 --> 00:21:18,678
the other features I was talking about

543
00:21:15,789 --> 00:21:21,289
here's the catch though that commercial

544
00:21:18,679 --> 00:21:23,929
drive has to be at least as fast as ours

545
00:21:21,289 --> 00:21:26,058
games that rely on the speed of our SSD

546
00:21:23,929 --> 00:21:28,970
need to work flawlessly with nem to

547
00:21:26,058 --> 00:21:31,160
drive when I gave the Wired interview

548
00:21:28,970 --> 00:21:33,769
last year I said that the PlayStation 5

549
00:21:31,160 --> 00:21:36,890
SSD was faster than anything available

550
00:21:33,769 --> 00:21:40,308
on PC at the time commercial M 2 drives

551
00:21:36,890 --> 00:21:43,130
used pcie 3.0 and four lanes of that cap

552
00:21:40,308 --> 00:21:46,819
out at 3.5 gigabytes a second in other

553
00:21:43,130 --> 00:21:51,020
words no pcie 3.0 Drive can hit the

554
00:21:46,819 --> 00:21:52,970
required spec M 2 drives with PCI u 4.0

555
00:21:51,019 --> 00:21:56,480
or now out in the market we're getting

556
00:21:52,970 --> 00:21:58,759
our in samples and seeing four or five

557
00:21:56,480 --> 00:22:00,558
gigabytes a second from them by year's

558
00:21:58,759 --> 00:22:03,379
end I expect there will be drives that

559
00:22:00,558 --> 00:22:05,808
saturate 4.0 and support seven gigabytes

560
00:22:03,380 --> 00:22:06,800
a second having said that we are

561
00:22:05,808 --> 00:22:08,930
comparing

562
00:22:06,799 --> 00:22:10,879
apples and oranges though because that

563
00:22:08,930 --> 00:22:12,650
commercial m2 Drive will have its own

564
00:22:10,880 --> 00:22:15,590
architecture its own flash controller

565
00:22:12,650 --> 00:22:17,720
and so on for example the nvme

566
00:22:15,589 --> 00:22:19,849
specification lays out a priority scheme

567
00:22:17,720 --> 00:22:23,089
for requests that the m2 drives can use

568
00:22:19,849 --> 00:22:26,359
and that scheme is pretty nice but it

569
00:22:23,089 --> 00:22:29,480
only has two true priority levels our

570
00:22:26,359 --> 00:22:31,039
drive supports six we can hook up a

571
00:22:29,480 --> 00:22:33,559
drive with only two priority levels

572
00:22:31,039 --> 00:22:35,629
definitely but our custom IO unit has to

573
00:22:33,559 --> 00:22:37,700
arbitrate the extra priorities rather

574
00:22:35,630 --> 00:22:39,950
than the m2 drives flash controller and

575
00:22:37,700 --> 00:22:41,809
so the m2 drive needs a little extra

576
00:22:39,950 --> 00:22:44,269
speed to take care of issues arising

577
00:22:41,809 --> 00:22:45,710
from the different approach that

578
00:22:44,269 --> 00:22:47,960
Commercial Drive also needs to

579
00:22:45,710 --> 00:22:50,380
physically fit inside of the bay we

580
00:22:47,960 --> 00:22:53,029
created in PlayStation 5 for m2 drives

581
00:22:50,380 --> 00:22:55,580
unlike internal hard drives there's

582
00:22:53,029 --> 00:22:58,190
unfortunately no standard for the height

583
00:22:55,579 --> 00:23:00,409
of an m2 Drive and some m2 drives have

584
00:22:58,190 --> 00:23:03,650
giant heat sinks in fact some of them

585
00:23:00,410 --> 00:23:05,360
even have their own fans right now we're

586
00:23:03,650 --> 00:23:07,850
getting em to drive samples and

587
00:23:05,359 --> 00:23:09,949
benchmarking them in various ways when

588
00:23:07,849 --> 00:23:11,629
games hit in beta as they get ready for

589
00:23:09,950 --> 00:23:13,730
the PlayStation 5 launch at year-end

590
00:23:11,630 --> 00:23:15,020
we'll also be doing some compatibility

591
00:23:13,730 --> 00:23:17,329
testing to make sure that the

592
00:23:15,019 --> 00:23:18,980
architecture of particular M 2 drives

593
00:23:17,329 --> 00:23:20,809
isn't too foreign for the games to

594
00:23:18,980 --> 00:23:22,940
handle once we've done that

595
00:23:20,809 --> 00:23:24,740
compatibility testing we should be able

596
00:23:22,940 --> 00:23:26,630
to start letting you know which drives

597
00:23:24,740 --> 00:23:28,400
will physically fit and which drive

598
00:23:26,630 --> 00:23:31,280
samples have benchmark appropriately

599
00:23:28,400 --> 00:23:32,930
high in our testing it would be great if

600
00:23:31,279 --> 00:23:35,660
that happened by launch but it's likely

601
00:23:32,930 --> 00:23:37,940
to be a bit past it so please hold off

602
00:23:35,660 --> 00:23:40,960
on getting that M to drive until you

603
00:23:37,940 --> 00:23:43,880
hear from us ok back to our principles

604
00:23:40,960 --> 00:23:45,799
balancing evolution and revolution is

605
00:23:43,880 --> 00:23:49,970
the second of them this was definitely a

606
00:23:45,799 --> 00:23:52,490
recurring theme with the GPU we need new

607
00:23:49,970 --> 00:23:54,049
GPU features and capabilities if if we

608
00:23:52,490 --> 00:23:56,329
only have more performance it's not

609
00:23:54,049 --> 00:23:58,609
really a new generation of console of

610
00:23:56,329 --> 00:24:00,859
course many of these capabilities result

611
00:23:58,609 --> 00:24:03,049
in more performance that's part of why a

612
00:24:00,859 --> 00:24:05,209
Playstation 5 teraflop is more powerful

613
00:24:03,049 --> 00:24:06,799
than a Playstation 4 teraflop but we

614
00:24:05,210 --> 00:24:09,140
aren't just looking for the performance

615
00:24:06,799 --> 00:24:11,419
we also need the ability to do something

616
00:24:09,140 --> 00:24:14,360
with the GPU that could not have been

617
00:24:11,420 --> 00:24:15,560
done before and we need higher

618
00:24:14,359 --> 00:24:17,569
performance per watt

619
00:24:15,559 --> 00:24:20,609
every time we double the performance of

620
00:24:17,569 --> 00:24:22,500
some GPU component we don't want to find

621
00:24:20,609 --> 00:24:25,288
we've doubled the power consumed and the

622
00:24:22,500 --> 00:24:27,869
heat produced but at the same time we

623
00:24:25,288 --> 00:24:30,450
have to make sure the GPU can run ps4

624
00:24:27,869 --> 00:24:32,308
games and we have to ensure that the

625
00:24:30,450 --> 00:24:35,399
architecture is easy for the developers

626
00:24:32,308 --> 00:24:37,980
to adopt now backwards compatibility was

627
00:24:35,398 --> 00:24:40,109
handled masterfully by AMD they treated

628
00:24:37,980 --> 00:24:42,298
it as a key need throughout the design

629
00:24:40,109 --> 00:24:44,158
process as our solution to adding new

630
00:24:42,298 --> 00:24:46,408
features without blindsiding developers

631
00:24:44,159 --> 00:24:48,269
we made sure that if there were new

632
00:24:46,409 --> 00:24:51,330
significant features it would be

633
00:24:48,269 --> 00:24:52,859
optional to use them the GPU supports

634
00:24:51,329 --> 00:24:55,319
ray tracing but you don't have to use

635
00:24:52,859 --> 00:24:57,028
ray tracing to make your game the GPU

636
00:24:55,319 --> 00:24:58,918
supports primitive shaders but you can

637
00:24:57,028 --> 00:25:02,038
release your first game on PlayStation 5

638
00:24:58,919 --> 00:25:04,350
without making any use of them before I

639
00:25:02,038 --> 00:25:06,028
get into the capabilities of the GPU I'd

640
00:25:04,349 --> 00:25:09,538
like to make clear two points that can

641
00:25:06,028 --> 00:25:12,769
be quite confusing first we have a

642
00:25:09,538 --> 00:25:16,019
custom AMD GPU based on there are DNA 2

643
00:25:12,769 --> 00:25:18,089
technology what does that mean AMD is

644
00:25:16,019 --> 00:25:20,370
continuously improving and revising

645
00:25:18,089 --> 00:25:23,369
their tech for our DNA to their goals

646
00:25:20,369 --> 00:25:25,619
were roughly speaking to reduce power of

647
00:25:23,369 --> 00:25:27,209
consumption by rhe architecting the GPU

648
00:25:25,619 --> 00:25:30,178
to put data close to where it's needed

649
00:25:27,210 --> 00:25:32,548
to optimize the GPU for performance and

650
00:25:30,179 --> 00:25:36,028
to adding new more advanced feature set

651
00:25:32,548 --> 00:25:38,369
but that feature set is malleable which

652
00:25:36,028 --> 00:25:40,470
is to say that we have our own needs for

653
00:25:38,369 --> 00:25:43,109
PlayStation and that can factor into

654
00:25:40,470 --> 00:25:45,808
what the AMD roadmap becomes so

655
00:25:43,109 --> 00:25:47,759
collaboration is born if we bring

656
00:25:45,808 --> 00:25:50,038
concepts to AMD that are felt to be

657
00:25:47,759 --> 00:25:52,169
widely useful then they can be adopted

658
00:25:50,038 --> 00:25:55,888
into our DNA - and used broadly

659
00:25:52,169 --> 00:25:57,659
including in PC GPUs if the ideas are

660
00:25:55,888 --> 00:25:58,709
sufficiently specific to what we're

661
00:25:57,659 --> 00:26:00,929
trying to accomplish

662
00:25:58,710 --> 00:26:02,788
like the GPU cache scrubbers I was

663
00:26:00,929 --> 00:26:06,240
talking about then they end up being

664
00:26:02,788 --> 00:26:09,869
just for us if you see a similar

665
00:26:06,240 --> 00:26:11,700
discrete GPU available as a PC card at

666
00:26:09,869 --> 00:26:13,888
roughly the same time as we release our

667
00:26:11,700 --> 00:26:16,860
console that means our collaboration

668
00:26:13,888 --> 00:26:19,319
with AMD succeeded in producing

669
00:26:16,859 --> 00:26:21,509
technology useful in both worlds it

670
00:26:19,319 --> 00:26:23,788
doesn't mean that we as sony simply

671
00:26:21,509 --> 00:26:26,669
incorporated the pc part into our

672
00:26:23,788 --> 00:26:29,038
console this continuous improvement in

673
00:26:26,669 --> 00:26:30,870
AMD technology means it's dangerous to

674
00:26:29,038 --> 00:26:32,669
rely on teraflops as an absolute

675
00:26:30,869 --> 00:26:34,678
indicator of performance and

676
00:26:32,669 --> 00:26:37,710
see you count should be avoided as well

677
00:26:34,679 --> 00:26:39,989
in the case of CPUs we all understand

678
00:26:37,710 --> 00:26:42,690
this the PlayStation 4 and PlayStation 5

679
00:26:39,989 --> 00:26:43,919
each have eight CPUs but we never think

680
00:26:42,690 --> 00:26:46,558
that meant the capabilities and

681
00:26:43,919 --> 00:26:48,299
performance are equal it's the same for

682
00:26:46,558 --> 00:26:50,190
see use for one thing they've been

683
00:26:48,298 --> 00:26:52,739
getting much larger over time

684
00:26:50,190 --> 00:26:55,769
adding new features means adding lots of

685
00:26:52,739 --> 00:26:59,730
transistors in fact the transistor count

686
00:26:55,769 --> 00:27:01,259
for a Playstation 5 CU is 62% larger

687
00:26:59,730 --> 00:27:05,009
than the transistor count for a

688
00:27:01,259 --> 00:27:07,710
playstation 4 CU second the playstation

689
00:27:05,009 --> 00:27:10,618
5 and GPU is backwards compatible with

690
00:27:07,710 --> 00:27:11,999
playstation 4 what does that mean one

691
00:27:10,618 --> 00:27:14,038
way you can achieve backwards

692
00:27:11,999 --> 00:27:16,319
compatibility is to put the previous

693
00:27:14,038 --> 00:27:18,298
consoles chipset in the new console like

694
00:27:16,319 --> 00:27:20,548
we did with some PlayStation 3s but

695
00:27:18,298 --> 00:27:23,099
that's of course extremely expensive a

696
00:27:20,548 --> 00:27:24,720
better way is to incorporate any

697
00:27:23,099 --> 00:27:27,509
differences in the previous consoles

698
00:27:24,720 --> 00:27:30,149
logic into the new consoles custom chips

699
00:27:27,509 --> 00:27:32,249
meaning that even as the technology

700
00:27:30,148 --> 00:27:34,349
evolves the logic and feature set that

701
00:27:32,249 --> 00:27:36,808
PlayStation 4 and PlayStation 4 Pro

702
00:27:34,349 --> 00:27:39,388
titles rely on is still available in

703
00:27:36,808 --> 00:27:41,668
backwards compatibility modes one

704
00:27:39,388 --> 00:27:43,288
advantage of this strategy is that once

705
00:27:41,669 --> 00:27:46,528
backwards compatibility is in the

706
00:27:43,288 --> 00:27:48,960
console it's in if not as if a cost down

707
00:27:46,528 --> 00:27:51,839
will remove backwards compatibility like

708
00:27:48,960 --> 00:27:54,028
it did on PlayStation 3 achieving this

709
00:27:51,839 --> 00:27:56,668
unification of functionality took years

710
00:27:54,028 --> 00:27:58,409
of efforts by AMD as any roadmap

711
00:27:56,669 --> 00:28:02,309
advancement creates a potential

712
00:27:58,409 --> 00:28:04,169
divergence in logic running ps4 and PS 4

713
00:28:02,308 --> 00:28:06,628
titles at boosted frequencies has also

714
00:28:04,169 --> 00:28:08,340
added complexity the boost is truly

715
00:28:06,628 --> 00:28:10,709
massive this time around and some game

716
00:28:08,339 --> 00:28:13,249
code just can't handle it testing has to

717
00:28:10,710 --> 00:28:16,019
be done on the title by title basis

718
00:28:13,249 --> 00:28:17,399
results are excellent though we recently

719
00:28:16,019 --> 00:28:19,528
took a look at the top hundred

720
00:28:17,398 --> 00:28:22,349
PlayStation 4 titles as ranked by play

721
00:28:19,528 --> 00:28:24,179
time and we're expecting almost all of

722
00:28:22,349 --> 00:28:27,898
them to be playable at launch on

723
00:28:24,179 --> 00:28:30,119
playstation 5 with regards to new

724
00:28:27,898 --> 00:28:32,219
features as I said our strategy was to

725
00:28:30,118 --> 00:28:35,038
try to break new ground but at the same

726
00:28:32,220 --> 00:28:38,399
time not to require use of the new GPU

727
00:28:35,038 --> 00:28:40,259
capabilities for more than a decade GPUs

728
00:28:38,398 --> 00:28:42,538
have imposed a restriction on game

729
00:28:40,259 --> 00:28:43,980
engines software handles vertex

730
00:28:42,538 --> 00:28:45,400
processing but for the most part

731
00:28:43,980 --> 00:28:47,529
dedicated hardware

732
00:28:45,400 --> 00:28:49,960
responsible for the triangles and other

733
00:28:47,529 --> 00:28:52,660
geometry that the vertices form that

734
00:28:49,960 --> 00:28:54,789
means it's not possible to do even basic

735
00:28:52,660 --> 00:28:56,769
optimizations such as aborting

736
00:28:54,789 --> 00:29:00,879
processing of a vertex if all geometry

737
00:28:56,769 --> 00:29:03,369
that uses it is off screen PlayStation 5

738
00:29:00,880 --> 00:29:04,870
has a new unit called the geometry

739
00:29:03,369 --> 00:29:06,819
engine which brings handling of

740
00:29:04,869 --> 00:29:09,399
triangles and other primitives under

741
00:29:06,819 --> 00:29:11,079
full programmatic control as a game

742
00:29:09,400 --> 00:29:13,600
developer you're free to ignore its

743
00:29:11,079 --> 00:29:15,309
existence and use the PlayStation 5 GPU

744
00:29:13,599 --> 00:29:18,549
as if it were no more capable than the

745
00:29:15,309 --> 00:29:21,730
ps4 GPU or you can use this new

746
00:29:18,549 --> 00:29:23,769
intelligence in various ways simple

747
00:29:21,730 --> 00:29:25,599
usage could be performance optimizations

748
00:29:23,769 --> 00:29:28,119
such as removing back faced or

749
00:29:25,599 --> 00:29:30,099
off-screen vertices and triangles more

750
00:29:28,119 --> 00:29:31,869
complex usage involves something called

751
00:29:30,099 --> 00:29:34,480
primitive shaders which allow the game

752
00:29:31,869 --> 00:29:36,609
to synthesize geometry on-the-fly as

753
00:29:34,480 --> 00:29:39,370
it's being rendered it's a brand new

754
00:29:36,609 --> 00:29:41,409
capability using primitive shaders on

755
00:29:39,369 --> 00:29:43,989
PlayStation 5 will allow for a broad

756
00:29:41,410 --> 00:29:46,420
variety of techniques including smoothly

757
00:29:43,990 --> 00:29:48,400
varying level of detail addition of

758
00:29:46,420 --> 00:29:50,529
procedural detail to close up objects

759
00:29:48,400 --> 00:29:53,230
and improvements to particle effects and

760
00:29:50,529 --> 00:29:55,809
other visual special effects another

761
00:29:53,230 --> 00:29:58,539
major new feature of our custom our DNA

762
00:29:55,809 --> 00:30:01,649
2 based GPU is ray tracing using the

763
00:29:58,539 --> 00:30:05,109
same strategy as AMD's upcoming PC GPUs

764
00:30:01,650 --> 00:30:07,120
the ciues contain a new specialized unit

765
00:30:05,109 --> 00:30:09,099
called the intersection engine which can

766
00:30:07,119 --> 00:30:11,289
calculate the intersection of rays with

767
00:30:09,099 --> 00:30:13,389
boxes and triangles to use the

768
00:30:11,289 --> 00:30:15,670
intersection engine first you build what

769
00:30:13,390 --> 00:30:18,100
is called an acceleration structure its

770
00:30:15,670 --> 00:30:20,230
data in RAM that contains all of your

771
00:30:18,099 --> 00:30:22,689
geometry there's a specific set of

772
00:30:20,230 --> 00:30:25,750
formats you can use their variations on

773
00:30:22,690 --> 00:30:28,240
the same BVH concept then in your shader

774
00:30:25,750 --> 00:30:30,369
program you use a new instruction that

775
00:30:28,240 --> 00:30:33,730
asks the intersection engine to check

776
00:30:30,369 --> 00:30:35,589
array against the BVH while the

777
00:30:33,730 --> 00:30:37,539
intersection engine is processing the

778
00:30:35,589 --> 00:30:39,970
requested ray triangle or ray box

779
00:30:37,539 --> 00:30:42,250
intersections the shaders are free to do

780
00:30:39,970 --> 00:30:43,900
other work having said that the ray

781
00:30:42,250 --> 00:30:46,569
tracing instruction is pretty memory

782
00:30:43,900 --> 00:30:49,690
intensive so it's a good mix with logic

783
00:30:46,569 --> 00:30:51,939
heavy code there's of course no need to

784
00:30:49,690 --> 00:30:54,580
use ray tracing ps4 graphics engines

785
00:30:51,940 --> 00:30:56,860
will run just fine on PlayStation 5 but

786
00:30:54,579 --> 00:30:59,349
it presents an opportunity for those

787
00:30:56,859 --> 00:31:01,418
interested I'm thinking it'll

788
00:30:59,349 --> 00:31:03,308
take less than a million raise a second

789
00:31:01,419 --> 00:31:05,169
to have a big impact on audio that

790
00:31:03,308 --> 00:31:07,960
should be enough for audio occlusion and

791
00:31:05,169 --> 00:31:10,120
some reverb calculations with a bit more

792
00:31:07,960 --> 00:31:11,769
of the GPU invested in ray-tracing it

793
00:31:10,119 --> 00:31:13,769
should be possible to do some very nice

794
00:31:11,769 --> 00:31:16,660
global illumination

795
00:31:13,769 --> 00:31:18,819
having said that adding ray-traced

796
00:31:16,660 --> 00:31:20,590
shadows and reflections to a traditional

797
00:31:18,819 --> 00:31:22,720
graphics engine could easily take

798
00:31:20,589 --> 00:31:24,730
hundreds of millions of rays a second

799
00:31:22,720 --> 00:31:27,579
and full ray tracing could take billions

800
00:31:24,730 --> 00:31:28,900
how far can we go I'm starting to get

801
00:31:27,579 --> 00:31:31,298
quite bullish I've already seen a

802
00:31:28,900 --> 00:31:33,309
PlayStation 5 title that's successfully

803
00:31:31,298 --> 00:31:36,308
using ray tracing based reflections in

804
00:31:33,308 --> 00:31:39,220
complex animated scenes with only modest

805
00:31:36,308 --> 00:31:42,548
costs another set of issues for the GPU

806
00:31:39,220 --> 00:31:44,798
involved size and frequency how big do

807
00:31:42,548 --> 00:31:47,168
we make the GPU and what frequency do we

808
00:31:44,798 --> 00:31:48,700
run it at this is a balancing act the

809
00:31:47,169 --> 00:31:50,950
chip has a cost and there's a cost for

810
00:31:48,700 --> 00:31:53,919
whatever we use to supply that chip with

811
00:31:50,950 --> 00:31:56,080
power and to cool it in general I like

812
00:31:53,919 --> 00:32:00,309
running the GPU at a higher frequency

813
00:31:56,079 --> 00:32:02,349
let me show you why here's two possible

814
00:32:00,308 --> 00:32:04,928
configurations for a GPU roughly of the

815
00:32:02,349 --> 00:32:06,369
level of the PlayStation 4 Pro this is a

816
00:32:04,929 --> 00:32:09,038
thought experiment don't take these

817
00:32:06,369 --> 00:32:11,168
configurations too seriously if you just

818
00:32:09,038 --> 00:32:13,960
calculate teraflops you get the same

819
00:32:11,169 --> 00:32:16,179
number but actually the performance is

820
00:32:13,960 --> 00:32:17,700
noticeably different because teraflops

821
00:32:16,179 --> 00:32:21,130
is defined as the computational

822
00:32:17,700 --> 00:32:23,380
capability of the vector ALU that's just

823
00:32:21,130 --> 00:32:25,809
one part of the GPU there are a lot of

824
00:32:23,380 --> 00:32:28,000
other units and those other units all

825
00:32:25,808 --> 00:32:30,908
run faster when the GPU frequency is

826
00:32:28,000 --> 00:32:34,359
higher at 33% higher frequency

827
00:32:30,909 --> 00:32:37,000
rasterization goes 33% faster processing

828
00:32:34,359 --> 00:32:40,269
the command buffer goes that much faster

829
00:32:37,000 --> 00:32:41,919
the l2 and other caches have that much

830
00:32:40,269 --> 00:32:44,288
higher bandwidth and so on

831
00:32:41,919 --> 00:32:46,870
about the only downside is that system

832
00:32:44,288 --> 00:32:49,690
memory is 33% further away in terms of

833
00:32:46,869 --> 00:32:51,788
cycles but the large number of benefits

834
00:32:49,690 --> 00:32:54,610
more than counterbalanced that as a

835
00:32:51,788 --> 00:32:58,419
friend of mine says a rising tide lifts

836
00:32:54,609 --> 00:33:01,418
all boats also it's easier to fully use

837
00:32:58,419 --> 00:33:04,299
36c use in parallel than it is to fully

838
00:33:01,419 --> 00:33:06,520
use 48c use when triangles are small

839
00:33:04,298 --> 00:33:09,339
it's much harder to fill although sea

840
00:33:06,519 --> 00:33:11,650
use with useful work so there's a lot to

841
00:33:09,339 --> 00:33:12,799
be said for faster assuming you can

842
00:33:11,650 --> 00:33:15,650
handle the resulting

843
00:33:12,799 --> 00:33:18,019
power and heat issues which frankly we

844
00:33:15,650 --> 00:33:19,700
haven't always done the best Java

845
00:33:18,019 --> 00:33:22,279
part of the reason for that is

846
00:33:19,700 --> 00:33:24,289
historically our process for setting CPU

847
00:33:22,279 --> 00:33:26,029
and GPU frequencies has relied on some

848
00:33:24,289 --> 00:33:28,069
heavy duty guesswork with regards to how

849
00:33:26,029 --> 00:33:30,200
much electrical power games will consume

850
00:33:28,069 --> 00:33:33,079
and how much heat will be produced as a

851
00:33:30,200 --> 00:33:35,150
result inside of the console power

852
00:33:33,079 --> 00:33:37,579
consumption varies a lot from game to

853
00:33:35,150 --> 00:33:40,009
game when I play god of war on my ps4

854
00:33:37,579 --> 00:33:42,919
Pro I know the power consumption is high

855
00:33:40,009 --> 00:33:45,349
just by the fan noise but power isn't

856
00:33:42,920 --> 00:33:46,880
simply about engine quality it's about

857
00:33:45,349 --> 00:33:49,730
the minutiae of what's being displayed

858
00:33:46,880 --> 00:33:51,890
and how it's counterintuitive but

859
00:33:49,730 --> 00:33:53,509
processing dense geometry typically

860
00:33:51,890 --> 00:33:56,390
consumes less power than processing

861
00:33:53,509 --> 00:33:58,819
simple geometry which is I suspect why

862
00:33:56,390 --> 00:34:01,910
horizons map screen with its low

863
00:33:58,819 --> 00:34:05,629
triangle count makes my ps4 pro heat up

864
00:34:01,910 --> 00:34:07,070
so much our process on previous consoles

865
00:34:05,630 --> 00:34:09,289
has been to try to guess what the

866
00:34:07,069 --> 00:34:11,569
maximum power consumption during the

867
00:34:09,289 --> 00:34:14,210
entire console lifetime might be which

868
00:34:11,570 --> 00:34:17,269
is to say the worst case scene in the

869
00:34:14,210 --> 00:34:19,429
worst case game and prepare a cooling

870
00:34:17,269 --> 00:34:22,969
solution that we think will be quiet at

871
00:34:19,429 --> 00:34:25,340
that power level if we get it right fan

872
00:34:22,969 --> 00:34:27,109
noise is minimal if we get it wrong the

873
00:34:25,340 --> 00:34:28,880
console will be quite loud for the

874
00:34:27,110 --> 00:34:30,289
higher power games and there's even a

875
00:34:28,880 --> 00:34:33,880
chance that it might overheat and shut

876
00:34:30,289 --> 00:34:36,168
down if we miss estimate power too badly

877
00:34:33,880 --> 00:34:39,829
PlayStation 5 is especially challenging

878
00:34:36,168 --> 00:34:41,889
because the CPU supports 256 bit native

879
00:34:39,829 --> 00:34:44,389
instructions that consume a lot of power

880
00:34:41,889 --> 00:34:47,539
these are great here and there but

881
00:34:44,389 --> 00:34:50,599
presumably only minimally used or are

882
00:34:47,539 --> 00:34:53,090
they if we plan for major 256 bit

883
00:34:50,599 --> 00:34:55,489
instruction usage we need to set the CPU

884
00:34:53,090 --> 00:34:57,410
clock substantially lower or noticeably

885
00:34:55,489 --> 00:35:01,669
increase the size of the power supply

886
00:34:57,409 --> 00:35:03,440
and fan so after much discussion we

887
00:35:01,670 --> 00:35:06,800
decided to go with a very different

888
00:35:03,440 --> 00:35:11,389
direction on PlayStation 5 we built a

889
00:35:06,800 --> 00:35:14,120
GPU with 36c use mind you our DNA to see

890
00:35:11,389 --> 00:35:15,739
use our large each has 62% more

891
00:35:14,119 --> 00:35:18,469
transistors than the CIE use we were

892
00:35:15,739 --> 00:35:21,559
using on PlayStation 4 so if we compare

893
00:35:18,469 --> 00:35:24,409
transistor counts 36 our DNA to see use

894
00:35:21,559 --> 00:35:26,329
equates to roughly 58 PlayStation 4 C

895
00:35:24,409 --> 00:35:29,779
use it is a fairly

896
00:35:26,329 --> 00:35:31,909
sizeable GPU then we went with a

897
00:35:29,780 --> 00:35:33,890
variable frequency strategy for

898
00:35:31,909 --> 00:35:36,799
PlayStation 5 which is to say we

899
00:35:33,889 --> 00:35:39,679
continuously run the GPU and CPU in

900
00:35:36,800 --> 00:35:41,840
boost mode we supply a generous amount

901
00:35:39,679 --> 00:35:44,269
of electrical power and then increase

902
00:35:41,840 --> 00:35:46,100
the frequency of GPU and CPU until they

903
00:35:44,269 --> 00:35:48,679
reach the capabilities of the system's

904
00:35:46,099 --> 00:35:50,719
cooling solution it's a completely

905
00:35:48,679 --> 00:35:52,759
different paradigm rather than running

906
00:35:50,719 --> 00:35:55,129
at constant frequency and letting power

907
00:35:52,760 --> 00:35:57,170
vary based on the workload we run at

908
00:35:55,130 --> 00:35:59,059
essentially constant power and let the

909
00:35:57,170 --> 00:36:02,300
frequency band vary based on the

910
00:35:59,059 --> 00:36:04,009
workload we then tackled the engineering

911
00:36:02,300 --> 00:36:05,630
challenge of a cost-effective and

912
00:36:04,010 --> 00:36:08,300
high-performance cooling solution

913
00:36:05,630 --> 00:36:10,400
designed for that specific power level

914
00:36:08,300 --> 00:36:11,810
in some ways it becomes a simpler

915
00:36:10,400 --> 00:36:14,240
problem because there are no more

916
00:36:11,809 --> 00:36:16,190
unknowns there's no need to guess what

917
00:36:14,239 --> 00:36:18,979
power consumption the worst case game

918
00:36:16,190 --> 00:36:21,019
might have as for the details of the

919
00:36:18,980 --> 00:36:22,699
cooling solution we're saving them for

920
00:36:21,019 --> 00:36:24,409
our teardown I think you'll be quite

921
00:36:22,699 --> 00:36:28,609
happy with what the engineering team

922
00:36:24,409 --> 00:36:31,579
came up with so how fast can we run the

923
00:36:28,610 --> 00:36:33,200
GPU and CPU with this strategy the

924
00:36:31,579 --> 00:36:35,210
simplest approach would be to look at

925
00:36:33,199 --> 00:36:36,980
the actual temperature of the silicon

926
00:36:35,210 --> 00:36:39,409
die and throttle the frequency on that

927
00:36:36,980 --> 00:36:41,210
basis but that won't work it fails to

928
00:36:39,409 --> 00:36:43,579
create a consistent PlayStation 5

929
00:36:41,210 --> 00:36:45,740
experience it wouldn't do to run a

930
00:36:43,579 --> 00:36:48,529
console slower simply because it was in

931
00:36:45,739 --> 00:36:50,750
a hot room so rather than look at the

932
00:36:48,530 --> 00:36:53,000
actual temperature of the silicon die we

933
00:36:50,750 --> 00:36:54,980
look at the activities that the GPU and

934
00:36:53,000 --> 00:36:56,989
CPU are performing and set the

935
00:36:54,980 --> 00:36:59,380
frequencies on that basis which makes

936
00:36:56,989 --> 00:37:02,269
everything deterministic and repeatable

937
00:36:59,380 --> 00:37:04,369
while we're at it we also use AMD's

938
00:37:02,269 --> 00:37:07,610
smart shift technology and send any

939
00:37:04,369 --> 00:37:10,699
unused power from the CPU to the GPU so

940
00:37:07,610 --> 00:37:13,370
it can squeeze out a few more pixels the

941
00:37:10,699 --> 00:37:16,849
benefits of this strategy are quite

942
00:37:13,369 --> 00:37:18,889
large running a GPU at two gigahertz was

943
00:37:16,849 --> 00:37:21,380
looking like an unreachable target with

944
00:37:18,889 --> 00:37:23,900
the old fixed frequency strategy with

945
00:37:21,380 --> 00:37:26,510
this new paradigm we're able to run way

946
00:37:23,900 --> 00:37:28,340
over that in fact we have to cap the GPU

947
00:37:26,510 --> 00:37:30,230
frequency at two point two three

948
00:37:28,340 --> 00:37:35,000
gigahertz so that we can guarantee that

949
00:37:30,230 --> 00:37:37,190
the on chip logic operates properly 36c

950
00:37:35,000 --> 00:37:38,889
use at two point two three gigahertz is

951
00:37:37,190 --> 00:37:41,380
ten point three teraflops

952
00:37:38,889 --> 00:37:44,108
and we expect the GPU to spend most of

953
00:37:41,380 --> 00:37:45,450
its time at or close to that frequency

954
00:37:44,108 --> 00:37:48,219
and performance

955
00:37:45,449 --> 00:37:49,989
similarly running the CPU at 3 gigahertz

956
00:37:48,219 --> 00:37:52,689
was causing headaches with the old

957
00:37:49,989 --> 00:37:55,750
strategy but now we can run it as high

958
00:37:52,690 --> 00:37:59,170
as 3.5 gigahertz in fact it spends most

959
00:37:55,750 --> 00:38:00,909
of its time at that frequency that

960
00:37:59,170 --> 00:38:03,430
doesn't mean all games will be running

961
00:38:00,909 --> 00:38:06,308
in 2.2 3 gigahertz and 3.5 gigahertz

962
00:38:03,429 --> 00:38:09,068
when that worst case game arrives it

963
00:38:06,309 --> 00:38:11,650
will run at a lower clock speed but not

964
00:38:09,068 --> 00:38:13,420
too much lower to reduce power by 10% it

965
00:38:11,650 --> 00:38:15,250
only takes a couple of percent reduction

966
00:38:13,420 --> 00:38:17,889
in frequency so I'd expect any down

967
00:38:15,250 --> 00:38:19,539
clocking to be pretty minor all things

968
00:38:17,889 --> 00:38:21,818
considered the change to a variable

969
00:38:19,539 --> 00:38:24,579
frequency approach will show significant

970
00:38:21,818 --> 00:38:26,048
gains for PlayStation gamers the final

971
00:38:24,579 --> 00:38:28,720
of our three principals was about

972
00:38:26,048 --> 00:38:30,969
finding new dreams it's important for us

973
00:38:28,719 --> 00:38:33,818
on the hardware team to find new ways to

974
00:38:30,969 --> 00:38:37,959
expand or deepen gaming and that's what

975
00:38:33,818 --> 00:38:39,489
led us to a focus on 3d audio as players

976
00:38:37,960 --> 00:38:41,318
we experience the game through the

977
00:38:39,489 --> 00:38:43,000
visuals through audio and through the

978
00:38:41,318 --> 00:38:46,028
feedback we received from the controller

979
00:38:43,000 --> 00:38:48,099
such as Rumble or haptics personally I

980
00:38:46,028 --> 00:38:50,858
feel a game is just dead without audio

981
00:38:48,099 --> 00:38:54,160
visuals are of course important but the

982
00:38:50,858 --> 00:38:55,869
impact of audio is huge as well at the

983
00:38:54,159 --> 00:38:57,578
same time the audio team on a game

984
00:38:55,869 --> 00:38:59,890
project has to do a lot with a little

985
00:38:57,579 --> 00:39:02,200
for example on PlayStation 4 there's

986
00:38:59,889 --> 00:39:04,538
fierce competition for the Jaguar CPU

987
00:39:02,199 --> 00:39:07,358
cores audio typically ends up getting

988
00:39:04,539 --> 00:39:09,220
just a fraction of a core that's not

989
00:39:07,358 --> 00:39:11,170
much of a computational resource

990
00:39:09,219 --> 00:39:13,750
particularly when you consider that the

991
00:39:11,170 --> 00:39:16,000
visuals run at 30 or 60 frames a second

992
00:39:13,750 --> 00:39:20,469
but audio processing needs to happen at

993
00:39:16,000 --> 00:39:22,358
almost 200 times a second so it's been

994
00:39:20,469 --> 00:39:23,768
tough going making forward progress on

995
00:39:22,358 --> 00:39:26,170
audio with Playstation 4

996
00:39:23,768 --> 00:39:29,169
particularly when PlayStation 3 was such

997
00:39:26,170 --> 00:39:31,210
a beast when it came to audio the SPU's

998
00:39:29,170 --> 00:39:33,430
and cell were almost a perfect device

999
00:39:31,210 --> 00:39:35,318
for audio rendering simple pipeline

1000
00:39:33,429 --> 00:39:37,838
algorithms could really take advantage

1001
00:39:35,318 --> 00:39:39,818
of asynchronous DMA and frequently

1002
00:39:37,838 --> 00:39:41,588
reached a hundred percent utilization of

1003
00:39:39,818 --> 00:39:43,420
the floating-point unit there's

1004
00:39:41,588 --> 00:39:46,838
unfortunately nothing comparable on

1005
00:39:43,420 --> 00:39:49,119
PlayStation 4 probably the most dramatic

1006
00:39:46,838 --> 00:39:51,788
progress in the PlayStation 4 generation

1007
00:39:49,119 --> 00:39:52,559
has been with Virtual Reality the PSB

1008
00:39:51,789 --> 00:39:54,990
are hard

1009
00:39:52,559 --> 00:39:57,539
has its own audio unit it supports about

1010
00:39:54,989 --> 00:40:00,029
fifty pretty decent 3d sound sources and

1011
00:39:57,539 --> 00:40:02,070
this provided a hint as to where we

1012
00:40:00,030 --> 00:40:05,400
could go with audio as well as some

1013
00:40:02,070 --> 00:40:08,280
valuable experience not to oversimplify

1014
00:40:05,400 --> 00:40:11,309
but here were our goals for audio on

1015
00:40:08,280 --> 00:40:14,220
PlayStation 5 the first goal was great

1016
00:40:11,309 --> 00:40:16,529
audio for everyone not just VR users or

1017
00:40:14,219 --> 00:40:18,000
sound bar owners or headphone users that

1018
00:40:16,530 --> 00:40:18,570
meant audio had to be part of the

1019
00:40:18,000 --> 00:40:21,599
console

1020
00:40:18,570 --> 00:40:23,490
it couldn't be a peripheral the second

1021
00:40:21,599 --> 00:40:25,440
goal was to support hundreds of sound

1022
00:40:23,489 --> 00:40:27,449
sources we didn't want developers to

1023
00:40:25,440 --> 00:40:29,220
have to pick and choose what sounds

1024
00:40:27,449 --> 00:40:31,529
would get 3d effects and which wouldn't

1025
00:40:29,219 --> 00:40:34,709
we wanted every sound in the game to

1026
00:40:31,530 --> 00:40:37,050
have dimensionality and finally we

1027
00:40:34,710 --> 00:40:40,650
wanted to really take on the challenges

1028
00:40:37,050 --> 00:40:42,930
of presence and locality now when we say

1029
00:40:40,650 --> 00:40:45,660
presence we mean the feeling that you're

1030
00:40:42,929 --> 00:40:47,849
actually there you've entered the matrix

1031
00:40:45,659 --> 00:40:49,799
it's not of course something we thought

1032
00:40:47,849 --> 00:40:51,750
we could perfectly achieve but the idea

1033
00:40:49,800 --> 00:40:54,750
was that if we stopped using just a rain

1034
00:40:51,750 --> 00:40:56,579
sound and instead use lots of 3d audio

1035
00:40:54,750 --> 00:40:58,530
sources for raindrops hitting the ground

1036
00:40:56,579 --> 00:41:00,809
at all sorts of locations around you

1037
00:40:58,530 --> 00:41:02,430
then at some point your brain would take

1038
00:41:00,809 --> 00:41:05,579
a leap and you'd begin to have this

1039
00:41:02,429 --> 00:41:07,259
feeling this feeling of real presence

1040
00:41:05,579 --> 00:41:09,869
inside the virtual world of the game

1041
00:41:07,260 --> 00:41:12,570
this has the capacity to affect your

1042
00:41:09,869 --> 00:41:16,349
appreciation of the game just like music

1043
00:41:12,570 --> 00:41:18,300
in a game does the concept of locality

1044
00:41:16,349 --> 00:41:20,400
is simpler it's just your sense of where

1045
00:41:18,300 --> 00:41:22,710
the audio is coming from to the right of

1046
00:41:20,400 --> 00:41:24,809
you behind you above you this can

1047
00:41:22,710 --> 00:41:27,619
immerse you further in the game and it

1048
00:41:24,809 --> 00:41:30,989
can also directly enhance the gameplay

1049
00:41:27,619 --> 00:41:33,690
to use dead space as an example I know

1050
00:41:30,989 --> 00:41:37,109
old school you're fighting enemies in

1051
00:41:33,690 --> 00:41:38,970
fairly dark spooky locations back in the

1052
00:41:37,110 --> 00:41:40,500
day if you played the game using the TV

1053
00:41:38,969 --> 00:41:42,539
speakers you could tell that there was

1054
00:41:40,500 --> 00:41:44,699
one last enemy growling and hunting you

1055
00:41:42,539 --> 00:41:47,550
down but it was difficult to tell quite

1056
00:41:44,699 --> 00:41:49,379
where that enemy was with headphones you

1057
00:41:47,550 --> 00:41:51,210
can tell that the enemy was somewhere on

1058
00:41:49,380 --> 00:41:52,710
the right which lets you deduce if you

1059
00:41:51,210 --> 00:41:55,769
couldn't see it that it must be

1060
00:41:52,710 --> 00:41:58,019
somewhere behind and to your right but

1061
00:41:55,769 --> 00:42:00,179
with 3d audio with good locality the

1062
00:41:58,019 --> 00:42:03,980
idea is you know the enemy is precisely

1063
00:42:00,179 --> 00:42:06,839
there and you turn and you take it out

1064
00:42:03,980 --> 00:42:09,929
so how do we know where a sound is

1065
00:42:06,840 --> 00:42:12,120
coming from in the first place well all

1066
00:42:09,929 --> 00:42:14,699
those bumps and folds in the ear have a

1067
00:42:12,119 --> 00:42:16,109
meaning evolutionarily speaking based on

1068
00:42:14,699 --> 00:42:18,299
what direction the sound is coming from

1069
00:42:16,110 --> 00:42:19,860
sound waves bounce around inside the ear

1070
00:42:18,300 --> 00:42:21,720
there's some constructive and

1071
00:42:19,860 --> 00:42:24,599
destructive interference and the result

1072
00:42:21,719 --> 00:42:26,669
is a change in volume the phase of the

1073
00:42:24,599 --> 00:42:28,559
sound also shifts depending on what path

1074
00:42:26,670 --> 00:42:31,500
the sound wave took to reach the ear

1075
00:42:28,559 --> 00:42:33,269
canal these volume changes and phase

1076
00:42:31,500 --> 00:42:35,940
shifts are different for each direction

1077
00:42:33,269 --> 00:42:38,670
and also vary depending on the frequency

1078
00:42:35,940 --> 00:42:40,740
of the sound head size and head shape

1079
00:42:38,670 --> 00:42:43,289
also impact the sound in a similar

1080
00:42:40,739 --> 00:42:45,719
fashion the way that the sound changes

1081
00:42:43,289 --> 00:42:47,279
based on direction and frequency can be

1082
00:42:45,719 --> 00:42:50,129
encoded in a table called the head

1083
00:42:47,280 --> 00:42:53,010
related transfer function or HR TF

1084
00:42:50,130 --> 00:42:55,680
here's part of one the vertical axis is

1085
00:42:53,010 --> 00:42:57,750
the frequency the horizontal axis is the

1086
00:42:55,679 --> 00:43:00,149
direction front back left right and the

1087
00:42:57,750 --> 00:43:03,539
color gives the degree of attenuation of

1088
00:43:00,150 --> 00:43:05,760
the sound at that frequency the HRT F is

1089
00:43:03,539 --> 00:43:07,980
as unique to an individual as a

1090
00:43:05,760 --> 00:43:10,920
fingerprint is in fact you're looking at

1091
00:43:07,980 --> 00:43:13,199
mine right now here is how we measure in

1092
00:43:10,920 --> 00:43:15,809
h RTF we've taken hundreds of people

1093
00:43:13,199 --> 00:43:17,969
through this process we put a microphone

1094
00:43:15,809 --> 00:43:21,269
in the subjects left and right ear

1095
00:43:17,969 --> 00:43:24,389
canals and then sit the subject down in

1096
00:43:21,269 --> 00:43:26,639
the middle of an array of 22 speakers we

1097
00:43:24,389 --> 00:43:29,400
then play an audio sweep from each

1098
00:43:26,639 --> 00:43:31,139
speaker as we rotate the subject in the

1099
00:43:29,400 --> 00:43:34,019
course of ten or twenty minutes we're

1100
00:43:31,139 --> 00:43:37,049
able to sample the H RTF at over 1,000

1101
00:43:34,019 --> 00:43:39,480
locations using an H RTF when rendering

1102
00:43:37,050 --> 00:43:42,330
audio creates unparalleled quality but

1103
00:43:39,480 --> 00:43:44,340
it's computationally expensive the

1104
00:43:42,329 --> 00:43:46,409
simplest way to use an H RTF is to

1105
00:43:44,340 --> 00:43:47,700
process a sound source to make it appear

1106
00:43:46,409 --> 00:43:49,909
as if it's coming from one of those

1107
00:43:47,699 --> 00:43:51,989
thousand locations we sample

1108
00:43:49,909 --> 00:43:53,699
unfortunately the processing has to be

1109
00:43:51,989 --> 00:43:55,709
done in frequency domain rather than

1110
00:43:53,699 --> 00:43:57,359
time domain so there's multiple fast

1111
00:43:55,710 --> 00:43:59,670
Fourier transforms needed for every

1112
00:43:57,360 --> 00:44:03,210
sound source for every audio tech that's

1113
00:43:59,670 --> 00:44:04,950
a lot of multiplies this computational

1114
00:44:03,210 --> 00:44:06,960
complexity was the determining factor

1115
00:44:04,949 --> 00:44:08,969
for our strategy Ament we had to bite

1116
00:44:06,960 --> 00:44:11,820
the bullet and design and build a custom

1117
00:44:08,969 --> 00:44:13,529
hardware unit for 3d audio collectively

1118
00:44:11,820 --> 00:44:16,230
we're referring to the hardware unit and

1119
00:44:13,530 --> 00:44:20,220
the proprietary algorithms we run on it

1120
00:44:16,230 --> 00:44:22,170
tempest 3d audio Tech the meaning of 3d

1121
00:44:20,219 --> 00:44:23,009
audio and technology should be pretty

1122
00:44:22,170 --> 00:44:25,860
obvious here

1123
00:44:23,010 --> 00:44:28,020
as for tempest I feel it really reflects

1124
00:44:25,860 --> 00:44:30,599
our goals with audio it suggests a

1125
00:44:28,019 --> 00:44:32,849
certain intensity of experience and also

1126
00:44:30,599 --> 00:44:34,589
hints at your presence within it we're

1127
00:44:32,849 --> 00:44:37,319
calling the hardware unit that we built

1128
00:44:34,590 --> 00:44:40,079
the tempest engine it's based on AMD's

1129
00:44:37,320 --> 00:44:42,510
GPU technology we modified a compute

1130
00:44:40,079 --> 00:44:44,789
unit in such a way as to make it very

1131
00:44:42,510 --> 00:44:46,320
close to the SPU is in PlayStation 3

1132
00:44:44,789 --> 00:44:49,440
remember when I said that they were

1133
00:44:46,320 --> 00:44:52,590
ideal for audio so the tempest engine

1134
00:44:49,440 --> 00:44:56,490
has no caches just like an SPU all data

1135
00:44:52,590 --> 00:44:58,680
access is via DMA just like an SPU our

1136
00:44:56,489 --> 00:45:01,079
target was that it would have more power

1137
00:44:58,679 --> 00:45:03,449
than a CPU thanks to the parallelism

1138
00:45:01,079 --> 00:45:05,639
that a GPU can achieve and then it would

1139
00:45:03,449 --> 00:45:08,309
be more efficient than our GPU thanks to

1140
00:45:05,639 --> 00:45:10,799
the SPU like architecture the goal being

1141
00:45:08,309 --> 00:45:15,029
to make possible near 100% utilization

1142
00:45:10,800 --> 00:45:17,730
of the C use vector units where we ended

1143
00:45:15,030 --> 00:45:20,220
up is a unit with roughly the same sim D

1144
00:45:17,730 --> 00:45:22,889
power and bandwidth as all eight Jaguar

1145
00:45:20,219 --> 00:45:24,750
cores in the PlayStation 4 combined if

1146
00:45:22,889 --> 00:45:27,049
we were to use the same algorithms as

1147
00:45:24,750 --> 00:45:30,030
psvr that's enough for something like

1148
00:45:27,050 --> 00:45:32,010
5,000 sound sources but of course we

1149
00:45:30,030 --> 00:45:34,260
want to use more complex algorithms and

1150
00:45:32,010 --> 00:45:37,650
we don't need anything like that number

1151
00:45:34,260 --> 00:45:40,200
of sounds it would have been wonderful

1152
00:45:37,650 --> 00:45:42,360
if a simpler strategy such as using

1153
00:45:40,199 --> 00:45:44,129
Dolby Atmos peripherals could have

1154
00:45:42,360 --> 00:45:46,170
achieved our goals but we wanted 3d

1155
00:45:44,130 --> 00:45:49,170
audio for all not just those with

1156
00:45:46,170 --> 00:45:50,849
licensed sound bars or the like also we

1157
00:45:49,170 --> 00:45:52,800
wanted many hundreds of sound sources

1158
00:45:50,849 --> 00:45:55,079
not just the 32 that at most supports

1159
00:45:52,800 --> 00:45:57,180
and finally we wanted to be able to

1160
00:45:55,079 --> 00:45:59,039
throw an overwhelming amount of

1161
00:45:57,179 --> 00:46:00,899
processing power at the problem and it

1162
00:45:59,039 --> 00:46:03,869
wasn't clear what any peripheral might

1163
00:46:00,900 --> 00:46:04,710
have inside of it in fact with the

1164
00:46:03,869 --> 00:46:06,630
tempest engine

1165
00:46:04,710 --> 00:46:08,940
we've even got enough power that we can

1166
00:46:06,630 --> 00:46:10,490
allocate some to the games to the extent

1167
00:46:08,940 --> 00:46:13,110
that games want to make use of

1168
00:46:10,489 --> 00:46:14,429
convolution reverb and other algorithms

1169
00:46:13,110 --> 00:46:17,370
that are either computationally

1170
00:46:14,429 --> 00:46:19,679
expensive or need high bandwidth but the

1171
00:46:17,369 --> 00:46:24,539
primary purpose of the tempest engine

1172
00:46:19,679 --> 00:46:27,089
remains 3d audio now 3d audio is a major

1173
00:46:24,539 --> 00:46:28,670
academic research topic it's safe to say

1174
00:46:27,090 --> 00:46:30,440
that no one in the world

1175
00:46:28,670 --> 00:46:32,480
all of the answers and the set of

1176
00:46:30,440 --> 00:46:34,400
algorithms that has to be invented tuned

1177
00:46:32,480 --> 00:46:38,780
or implemented to realize our vision for

1178
00:46:34,400 --> 00:46:42,108
3d audio is immense for example use of

1179
00:46:38,780 --> 00:46:43,910
HR TFS in games is quite complex before

1180
00:46:42,108 --> 00:46:46,519
I talked about making a sound source

1181
00:46:43,909 --> 00:46:48,889
appear as if it's coming from one of

1182
00:46:46,519 --> 00:46:52,309
those thousand HR TF sample locations

1183
00:46:48,889 --> 00:46:54,650
but for high quality 3d game audio we

1184
00:46:52,309 --> 00:46:56,450
have to handle other possibilities the

1185
00:46:54,650 --> 00:46:58,460
sound source might not be at one of the

1186
00:46:56,449 --> 00:47:01,039
thousand HR TF sample locations so we

1187
00:46:58,460 --> 00:47:02,990
have to blend the HR TF data from the

1188
00:47:01,039 --> 00:47:05,150
closest locations that we have sampled

1189
00:47:02,989 --> 00:47:07,098
the sound source might be moving which

1190
00:47:05,150 --> 00:47:09,139
needs very special handling as that

1191
00:47:07,099 --> 00:47:11,510
blend keeps changing and that can cause

1192
00:47:09,139 --> 00:47:14,389
phase artifacts in the resulting audio

1193
00:47:11,510 --> 00:47:16,069
or the sound source might have a size to

1194
00:47:14,389 --> 00:47:18,618
it meaning it should feel as if it's

1195
00:47:16,068 --> 00:47:22,159
coming from an area rather than a single

1196
00:47:18,619 --> 00:47:24,440
point there's also whole categories of

1197
00:47:22,159 --> 00:47:26,779
approaches to be handled 3d audio can be

1198
00:47:24,440 --> 00:47:29,510
implemented using individual processing

1199
00:47:26,780 --> 00:47:31,670
of 3d sound sources but alternatively

1200
00:47:29,510 --> 00:47:33,890
ambisonics can be used for 3d audio

1201
00:47:31,670 --> 00:47:36,170
ambisonics speak somewhat like the

1202
00:47:33,889 --> 00:47:38,420
spherical harmonics used in computer

1203
00:47:36,170 --> 00:47:40,548
graphics and finally there's audio

1204
00:47:38,420 --> 00:47:42,619
devices the player might be using

1205
00:47:40,548 --> 00:47:44,358
headphones or TV speakers or have a

1206
00:47:42,619 --> 00:47:46,490
higher end surround sound set up with

1207
00:47:44,358 --> 00:47:49,338
six or more speakers all of which need

1208
00:47:46,489 --> 00:47:51,259
different approaches that's a lot of

1209
00:47:49,338 --> 00:47:53,269
variations it's nice to have the

1210
00:47:51,260 --> 00:47:55,460
computational resources of the tempest

1211
00:47:53,269 --> 00:47:57,440
engine but it's clear that achieving our

1212
00:47:55,460 --> 00:48:00,220
ultimate goals with 3d audio is going to

1213
00:47:57,440 --> 00:48:02,389
be a multi-year step-by-step process

1214
00:48:00,219 --> 00:48:04,519
having said that headphone audio

1215
00:48:02,389 --> 00:48:06,769
implementation is largely complete at

1216
00:48:04,519 --> 00:48:08,420
this time it was a natural place for us

1217
00:48:06,769 --> 00:48:10,219
to start with headphones we control

1218
00:48:08,420 --> 00:48:12,260
exactly what each ear hears and

1219
00:48:10,219 --> 00:48:13,699
therefore the algorithmic development

1220
00:48:12,260 --> 00:48:16,849
and implementation are more

1221
00:48:13,699 --> 00:48:18,679
straightforward for TV speakers and

1222
00:48:16,849 --> 00:48:21,230
stereo speakers were in the process of

1223
00:48:18,679 --> 00:48:22,969
implementing virtual surround sound the

1224
00:48:21,230 --> 00:48:25,250
idea being that if you're sitting in a

1225
00:48:22,969 --> 00:48:26,899
sweet spot in front of the TV then the

1226
00:48:25,250 --> 00:48:28,880
sound can be made to feel as if it's

1227
00:48:26,900 --> 00:48:29,650
coming from any direction even behind

1228
00:48:28,880 --> 00:48:32,000
you

1229
00:48:29,650 --> 00:48:34,130
virtual surround sound has a lot in

1230
00:48:32,000 --> 00:48:37,130
common with 3d audio on headphones but

1231
00:48:34,130 --> 00:48:38,630
it's more complex because the left ear

1232
00:48:37,130 --> 00:48:40,880
can hear the right speaker and

1233
00:48:38,630 --> 00:48:42,619
vice-versa we have a basic

1234
00:48:40,880 --> 00:48:44,150
implementation of virtual surround

1235
00:48:42,619 --> 00:48:46,160
up and running we're now looking at

1236
00:48:44,150 --> 00:48:48,588
increasing the size of that sweet spot

1237
00:48:46,159 --> 00:48:51,259
which is to say making the area you need

1238
00:48:48,588 --> 00:48:53,119
to be in to feel the 3d effect larger

1239
00:48:51,259 --> 00:48:56,358
and we're also working to boost the

1240
00:48:53,119 --> 00:48:58,309
sense of locality headphone audio is the

1241
00:48:56,358 --> 00:49:00,139
current gold standard for 3d audio on

1242
00:48:58,309 --> 00:49:01,910
playstation 5 but we're going to see

1243
00:49:00,139 --> 00:49:04,608
what we can do to bring virtual surround

1244
00:49:01,909 --> 00:49:06,440
sound to a similar level after which

1245
00:49:04,608 --> 00:49:08,150
we'll start in on setups with more

1246
00:49:06,440 --> 00:49:10,729
speakers such as 6 channel surround

1247
00:49:08,150 --> 00:49:12,139
sound it's now to the point where some

1248
00:49:10,728 --> 00:49:13,998
of the playstation 5 games in

1249
00:49:12,139 --> 00:49:16,308
development are extensively using these

1250
00:49:13,998 --> 00:49:17,718
systems one of the game demos allows you

1251
00:49:16,309 --> 00:49:20,690
to toggle between conventional

1252
00:49:17,719 --> 00:49:23,539
PlayStation 4 style stereo audio and our

1253
00:49:20,690 --> 00:49:25,309
new 3d audio I listened with just an

1254
00:49:23,539 --> 00:49:27,710
ordinary pair of over the Year

1255
00:49:25,309 --> 00:49:30,289
headphones and wow I could feel a

1256
00:49:27,710 --> 00:49:32,210
difference 3d audio has that dimensional

1257
00:49:30,289 --> 00:49:35,329
feel to it conventional stereo audio

1258
00:49:32,210 --> 00:49:38,749
feels smashed flat by comparison the

1259
00:49:35,329 --> 00:49:40,910
improvement is obvious so a big

1260
00:49:38,748 --> 00:49:43,308
advancement but have I entered the

1261
00:49:40,909 --> 00:49:45,139
matrix does my brain believe I'm really

1262
00:49:43,309 --> 00:49:47,778
there like I was talking about earlier

1263
00:49:45,139 --> 00:49:50,690
when I explained our target of presence

1264
00:49:47,778 --> 00:49:52,489
well the answer is no but you've

1265
00:49:50,690 --> 00:49:55,548
probably caught on to what's missing

1266
00:49:52,489 --> 00:49:58,880
here namely whose HRT F was being used

1267
00:49:55,548 --> 00:50:01,159
it wasn't mine it was the default HRT F

1268
00:49:58,880 --> 00:50:02,539
the audio team analyzed the hundreds

1269
00:50:01,159 --> 00:50:04,248
that they measured and chose the one

1270
00:50:02,539 --> 00:50:06,920
they felt with the closest fit to the

1271
00:50:04,248 --> 00:50:09,439
total game playing audience this shows a

1272
00:50:06,920 --> 00:50:11,838
piece of the default HRT F on the left

1273
00:50:09,440 --> 00:50:13,338
and my HRT F on the right you can see

1274
00:50:11,838 --> 00:50:16,239
that the general features are much the

1275
00:50:13,338 --> 00:50:19,369
same but the details are quite different

1276
00:50:16,239 --> 00:50:21,739
with the default HRT F as I said the 3d

1277
00:50:19,369 --> 00:50:24,619
audio sounds pretty great when I use my

1278
00:50:21,739 --> 00:50:27,289
HRT F though the audio reaches a higher

1279
00:50:24,619 --> 00:50:30,160
level of realism which is to say that

1280
00:50:27,289 --> 00:50:32,690
when using headphones and my HRT F I

1281
00:50:30,159 --> 00:50:34,909
occasionally get fooled and even think a

1282
00:50:32,690 --> 00:50:38,019
sound is coming from the real world when

1283
00:50:34,909 --> 00:50:40,159
it's actually coming from the game a

1284
00:50:38,018 --> 00:50:42,498
corollary to this is that there are a

1285
00:50:40,159 --> 00:50:45,170
few people whose HRT FS are sufficiently

1286
00:50:42,498 --> 00:50:47,209
far from the default HRT F that's the

1287
00:50:45,170 --> 00:50:49,909
red dot here that they can toggle

1288
00:50:47,210 --> 00:50:52,849
between ps4 style and PS 5 style audio

1289
00:50:49,909 --> 00:50:55,518
and not sense much difference I've had a

1290
00:50:52,849 --> 00:50:56,450
few people describe the playstation 5 3d

1291
00:50:55,518 --> 00:50:58,719
audio is sounding

1292
00:50:56,449 --> 00:51:00,889
like a bit better stereo audio

1293
00:50:58,719 --> 00:51:04,789
presumably they're the ones at the very

1294
00:51:00,889 --> 00:51:07,368
edges of this diagram which means what

1295
00:51:04,789 --> 00:51:08,989
HRT F you're using is key I'd like

1296
00:51:07,369 --> 00:51:11,150
everyone to be able to experience what

1297
00:51:08,989 --> 00:51:13,549
I'm experiencing but obviously it's not

1298
00:51:11,150 --> 00:51:16,369
possible to measure the HRT F of every

1299
00:51:13,550 --> 00:51:18,740
PlayStation user that means HRT F

1300
00:51:16,369 --> 00:51:21,170
selection and synthesis are going to be

1301
00:51:18,739 --> 00:51:24,139
big topics going forward as the tempest

1302
00:51:21,170 --> 00:51:26,150
technology matures at PlayStation 5

1303
00:51:24,139 --> 00:51:28,730
launch will be offering a choice of five

1304
00:51:26,150 --> 00:51:29,869
HRT FS there's a simple test where you

1305
00:51:28,730 --> 00:51:32,179
pick the one that gives you the best

1306
00:51:29,869 --> 00:51:34,430
locality that's just the first step

1307
00:51:32,179 --> 00:51:37,129
though this is an open-ended research

1308
00:51:34,429 --> 00:51:39,139
topic maybe you'll be sending us a photo

1309
00:51:37,130 --> 00:51:41,630
of your ear and we'll choose a neural

1310
00:51:39,139 --> 00:51:43,670
network to pick the closest HRT F in our

1311
00:51:41,630 --> 00:51:45,950
library maybe you'll be sending us a

1312
00:51:43,670 --> 00:51:47,690
video of your ears and your head and

1313
00:51:45,949 --> 00:51:50,480
we'll make a 3d model of them and

1314
00:51:47,690 --> 00:51:53,179
synthesize the HRT F maybe you'll play

1315
00:51:50,480 --> 00:51:55,639
an audio game to tune your HRT F will be

1316
00:51:53,179 --> 00:51:57,529
subtly changing it as you play and home

1317
00:51:55,639 --> 00:51:59,509
in on the HRT F that gives you the

1318
00:51:57,530 --> 00:52:02,900
highest score meaning that it matches

1319
00:51:59,510 --> 00:52:04,579
you the best this is a journey we'll all

1320
00:52:02,900 --> 00:52:04,940
be taking together over the next few

1321
00:52:04,579 --> 00:52:06,950
years

1322
00:52:04,940 --> 00:52:09,260
ultimately we're committing to enabling

1323
00:52:06,949 --> 00:52:13,549
everyone to experience that next level

1324
00:52:09,260 --> 00:52:15,230
of realism hopefully I've been able to

1325
00:52:13,550 --> 00:52:17,599
illustrate a bit about our design and

1326
00:52:15,230 --> 00:52:19,550
decision making process today and why

1327
00:52:17,599 --> 00:52:21,920
PlayStation 5 has the feature set that

1328
00:52:19,550 --> 00:52:23,780
it does now comes the fun part we get to

1329
00:52:21,920 --> 00:52:26,420
see how the development community takes

1330
00:52:23,780 --> 00:52:28,579
advantage of that feature set I'm hoping

1331
00:52:26,420 --> 00:52:30,800
for the completely unexpected will it

1332
00:52:28,579 --> 00:52:33,050
come from audio ray-tracing the

1333
00:52:30,800 --> 00:52:36,130
capabilities of the SSD or something

1334
00:52:33,050 --> 00:52:40,510
else I guess we'll find out soon enough

1335
00:52:36,130 --> 00:52:40,510
thank you for your time today