1
00:00:00,000 --> 00:00:07,560
on February 27 2023 AMD's ryzen 9

2
00:00:04,400 --> 00:00:10,139
7950x3d was formally declared to be the

3
00:00:07,560 --> 00:00:16,440
best CPU at every price point in the desktop Market at least by everyone else

4
00:00:13,340 --> 00:00:18,960
normally Gamers can count on us to

5
00:00:16,440 --> 00:00:24,300
confirm or refute manufacturer's claims allowing them to get the most for their

6
00:00:20,460 --> 00:00:28,080
money but we said nothing

7
00:00:24,300 --> 00:00:31,619
well about that first our testing was

8
00:00:28,080 --> 00:00:33,780
plagued with so many crashes

9
00:00:31,619 --> 00:00:38,940
that our original script for this video had us speculating whether AMD had

10
00:00:35,579 --> 00:00:40,680
validated this chip at all and once we

11
00:00:38,940 --> 00:00:45,120
worked that out we're going to touch on that later our performance numbers were

12
00:00:43,260 --> 00:00:50,280
so far off what others were reporting that we went back and we retested

13
00:00:47,340 --> 00:00:55,379
everything and what we found was that the deeper we dug the more complex the

14
00:00:52,920 --> 00:00:59,879
story around this chip became fastest on the planet

15
00:00:56,280 --> 00:01:01,920
yes but maybe only on roads that most

16
00:00:59,879 --> 00:01:06,960
people wouldn't drive it on fastest segue on the planet no but you're

17
00:01:04,920 --> 00:01:12,119
watching it anyway to learn about our sponsor any desk experience seamless and

18
00:01:09,780 --> 00:01:15,840
top-notch remote access with any desk you can effortlessly connect to any

19
00:01:14,280 --> 00:01:20,159
device regardless of the operating system and enjoy low latency whether you

20
00:01:18,720 --> 00:01:24,240
need remote support or you're working from home any desk has you covered don't

21
00:01:22,259 --> 00:01:29,299
miss out check out any desk at the link below here are the original graphs that

22
00:01:26,759 --> 00:01:34,740
we were going to publish for our ryzen 7950x3d review you can pause if you want

23
00:01:31,920 --> 00:01:41,159
to look closely but what matters is that at 1080p while we did see performance

24
00:01:37,860 --> 00:01:43,860
improvements over the regular 7950x no

25
00:01:41,159 --> 00:01:48,720
matter how many times we tested our data never quite lined up with AMD's own and

26
00:01:46,920 --> 00:01:54,720
if that wasn't strange enough look at these Modern Warfare 2 results lower

27
00:01:51,299 --> 00:01:57,299
than the 7950x then we turn up the

28
00:01:54,720 --> 00:02:02,759
resolution and I mean how is that even possible normally the performance

29
00:02:00,000 --> 00:02:08,099
differences between CPUs should shrink as we move to 1440p or especially 4K

30
00:02:06,479 --> 00:02:14,120
because we're shifting more of the demand over to the GPU well not here

31
00:02:11,540 --> 00:02:20,280
instead our 7950x3d managed worse performance than

32
00:02:17,640 --> 00:02:28,680
its X denoted cousin in multiple games at 1440p we retested it and retested it

33
00:02:24,720 --> 00:02:30,480
a lot same thing and this flies in the

34
00:02:28,680 --> 00:02:35,220
face of the conventional wisdom that CPUs should be tested at a lower

35
00:02:33,360 --> 00:02:39,660
resolution to better reveal their relative strength I mean it's basically

36
00:02:37,319 --> 00:02:45,420
benchmarking gospel that if it's better for gaming at 1080p it should be better

37
00:02:42,060 --> 00:02:47,519
or at the worst equal for gaming at any

38
00:02:45,420 --> 00:02:52,260
other resolution I mean outside of gaming we expected performance to be

39
00:02:49,560 --> 00:02:58,440
underwhelming due to the 7950x3d's lower clock speeds and TDP the last gen ryzen

40
00:02:55,500 --> 00:03:03,060
7 5800 X 3D behaved much the same way because 3D V cache just doesn't help

41
00:03:00,840 --> 00:03:07,680
that much with most productivity tasks but it was supposed to make up for this

42
00:03:05,700 --> 00:03:12,720
with its outstanding gaming performance something we simply weren't seeing the

43
00:03:10,800 --> 00:03:18,420
bottom line is that the review we never published was scathing performance was

44
00:03:15,900 --> 00:03:22,200
not in line with AMD's claims especially when we increased the resolution and

45
00:03:20,400 --> 00:03:26,879
from what we'd heard and what we're now seeing Supply was expected to be

46
00:03:24,239 --> 00:03:30,780
basically non-existent this combined with our experience with the early

47
00:03:28,140 --> 00:03:36,000
drivers our BIOS issues and the crashes we saw all pointed to a rushed possibly

48
00:03:33,540 --> 00:03:40,440
even paper launch AMD had every opportunity to contest our findings but

49
00:03:38,459 --> 00:03:45,060
when we showed them our results they basically went yeah seems okay so with

50
00:03:42,780 --> 00:03:49,260
nothing else to go on if we had launched that review we would have been the only

51
00:03:47,040 --> 00:03:52,860
outlet with a broadly negative take on this chip to their credit AMD did

52
00:03:51,180 --> 00:03:57,060
eventually acknowledge that our chip was probably defective saving us that

53
00:03:54,720 --> 00:04:01,440
embarrassment and giving us us the time to get the story right now we had some

54
00:03:59,400 --> 00:04:06,299
hints that our review sample was more yellow than golden with the integrated

55
00:04:04,200 --> 00:04:11,519
GPU enabled Windows would immediately crash once our GeForce gpus drivers

56
00:04:08,519 --> 00:04:13,200
loaded and across four different Expo

57
00:04:11,519 --> 00:04:17,459
certified memory kits we experienced constant crashes and memory intensive

58
00:04:15,000 --> 00:04:22,620
applications like blender and 7-Zip both of these behaviors were absent with

59
00:04:19,440 --> 00:04:24,660
regular non-x3d chips and this pattern

60
00:04:22,620 --> 00:04:31,320
was observed with multiple motherboards meaning our x3d CPU was the only common

61
00:04:28,139 --> 00:04:33,540
denominator so you might think why is

62
00:04:31,320 --> 00:04:37,199
that only a hint at a bad chip well the thing is these kinds of teething panes

63
00:04:35,759 --> 00:04:41,639
are a lot more common than you'd probably think on brand new platforms if

64
00:04:39,600 --> 00:04:46,139
you recall we had very similar stability issues with first gen ryzen and with

65
00:04:44,040 --> 00:04:49,620
12th gen core especially at higher memory speeds usually though they end up

66
00:04:48,000 --> 00:04:54,120
being footnotes in the review process because everything is pre-release and we

67
00:04:52,320 --> 00:04:57,660
often see these things resolved by the time they're actually generally

68
00:04:55,500 --> 00:05:01,860
available but but they're important here because all of this troubleshooting and

69
00:05:00,180 --> 00:05:06,960
Reporting all of our findings to AMD including the anomalous performance data

70
00:05:03,900 --> 00:05:10,320
ate up two days out of before that we

71
00:05:06,960 --> 00:05:12,600
had for testing wait only four days to

72
00:05:10,320 --> 00:05:16,680
evaluate this thing ah I'm glad you noticed that the unusually crunched

73
00:05:14,699 --> 00:05:21,360
review window was another reason for us to believe hey maybe our chip is

74
00:05:19,320 --> 00:05:25,620
perfectly normal and AMD is just trying to limit our time with it so that we

75
00:05:22,919 --> 00:05:31,199
don't dig too deep I know that's a pretty cynical take but the truth is we

76
00:05:28,139 --> 00:05:32,759
have seen the strategy before and aside

77
00:05:31,199 --> 00:05:37,020
from the crashing our performance numbers were at least close enough to

78
00:05:34,620 --> 00:05:42,360
AMD's expectations that even they initially told us yep everything looks

79
00:05:39,600 --> 00:05:47,940
fine but here's the thing silicon is notoriously fickle especially with these

80
00:05:44,699 --> 00:05:50,400
bleeding edge process nodes and good or

81
00:05:47,940 --> 00:05:54,600
bad might not be as binary as you might think for any given wafer fresh off the

82
00:05:53,160 --> 00:05:59,880
production line there are going to be multiple unusable eyes it's just the way

83
00:05:57,780 --> 00:06:04,800
it is as exact as the manufacturing process is it is truly ridiculous how

84
00:06:02,699 --> 00:06:08,820
small each transistor is these days smaller than even a virus and when

85
00:06:07,320 --> 00:06:13,560
you're working at that level you're at the mercy of literal quantum physics

86
00:06:11,520 --> 00:06:16,919
get the gate pitch off by a fraction of an animator and maybe that transistor

87
00:06:15,300 --> 00:06:20,460
doesn't flip as quickly as its neighbors and you can't ramp up clock speed or

88
00:06:19,020 --> 00:06:25,020
maybe some of them are bridged together and entirely non-functional unlike our

89
00:06:22,979 --> 00:06:29,340
Tech accessory holder from ltdstore.com it's got lots of compartments and it's

90
00:06:27,539 --> 00:06:34,319
very functional now these kinds of defects are normal and these flawed

91
00:06:32,100 --> 00:06:40,020
lower performing processor dies are often used in lower end products which

92
00:06:36,840 --> 00:06:42,840
is fine except that from time to time a

93
00:06:40,020 --> 00:06:47,940
chip can be so marginal that it barely passes those usual quality control

94
00:06:44,819 --> 00:06:50,100
checks and then ends up falling short in

95
00:06:47,940 --> 00:06:53,880
the field this is what we call losing the Silicon Lottery winning the Silicon

96
00:06:52,080 --> 00:06:58,440
Lottery on the other hand isn't simply getting a working chip it's having an

97
00:06:56,340 --> 00:07:02,400
exceptional one one whose production had very few defects giving it extra

98
00:07:00,539 --> 00:07:06,600
Headroom compared to the average for a given design in other words a good

99
00:07:04,680 --> 00:07:09,960
overclocker this is the part where I tell you that once we got our

100
00:07:08,039 --> 00:07:14,880
replacement chip all of our problems disappeared or at least would be if I

101
00:07:12,780 --> 00:07:19,319
were a big liar the stability problems are gone but the performance problems

102
00:07:17,520 --> 00:07:23,419
the ones that led to our negative conclusion in the first place are

103
00:07:21,419 --> 00:07:27,599
actually no better these new numbers then that I'm showing

104
00:07:25,800 --> 00:07:31,919
you now are the culmination of what is now weeks of testing and troubleshooting

105
00:07:29,940 --> 00:07:37,919
and they leave us with some really tough questions like are AMD's yields bad are

106
00:07:35,819 --> 00:07:42,240
we so unlucky as to get two poor performers in a row or is our testing

107
00:07:40,020 --> 00:07:46,800
methodology out of whack somehow when we ran our original tests we made sure to

108
00:07:44,699 --> 00:07:50,460
use AMD's recommended settings with fresh Windows installs and most

109
00:07:48,599 --> 00:07:54,539
importantly fresh chipset driver installs that last part is much more

110
00:07:53,039 --> 00:07:59,580
important than usual because the pre-release chipset driver only installs

111
00:07:57,240 --> 00:08:05,039
the new performance Optimizer and the awkwardly named PPM provisioning file

112
00:08:01,919 --> 00:08:09,180
driver when it detects a dual die x3d

113
00:08:05,039 --> 00:08:11,160
processor like the 7950x3d or 7900 x3d

114
00:08:09,180 --> 00:08:15,060
the release driver might behave the same way by the way so if you buy one of

115
00:08:13,199 --> 00:08:18,780
these as an upgrade double check to be sure we're gonna have purchase links

116
00:08:16,319 --> 00:08:22,800
down below these two tools are essential for getting the best performance out of

117
00:08:20,400 --> 00:08:28,379
these chips because only one of its two dies has 3D V cache and runs at a lower

118
00:08:25,800 --> 00:08:32,099
clock speed while the other is a standard Cash die that's optimized for

119
00:08:30,419 --> 00:08:36,539
higher clock speeds performance Optimizer then automatically promotes

120
00:08:34,260 --> 00:08:41,120
the appropriate dies course as the preferred cores in Windows scheduler

121
00:08:38,880 --> 00:08:46,920
while the PPM provisioning file driver Parks or puts to sleep the cores on the

122
00:08:44,399 --> 00:08:51,540
unused Die games will then assuming everything's working properly typically

123
00:08:48,839 --> 00:08:55,560
gets sent to the 3D V cache die for best performance and as far as we can tell

124
00:08:53,760 --> 00:08:59,880
this is down to a simple integration with Windows game mode and Xbox game bar

125
00:08:58,200 --> 00:09:03,060
rather than any special sauce that's been cooked Up by Andy themselves an

126
00:09:01,800 --> 00:09:07,019
important thing to note though the driver does not disable the other die

127
00:09:05,220 --> 00:09:11,220
entirely and if enough threads are needed it will wake them up even if

128
00:09:09,360 --> 00:09:16,800
you're in game so it's worth noting that a 7900x 3D would wake up more often

129
00:09:14,640 --> 00:09:21,000
since each die has only six cores instead of eight unsurprisingly then if

130
00:09:19,500 --> 00:09:25,680
these features aren't working properly the Windows scheduler is just flying

131
00:09:23,220 --> 00:09:30,240
blind unaware of the differences between the two processor dies and it'll end up

132
00:09:28,080 --> 00:09:34,380
ranking the best course based on the fastest available across the whole CPU

133
00:09:32,580 --> 00:09:39,180
something it can't rightly determine this can lead to situations where an app

134
00:09:36,600 --> 00:09:43,560
might run across multiple dies at once despite only needing one die's worth of

135
00:09:41,519 --> 00:09:47,100
CPU threads which can incur latency penalty since the Dives will have to

136
00:09:45,000 --> 00:09:50,519
transfer data from cash to cash rather than accessing it directly now we

137
00:09:48,839 --> 00:09:56,100
confirmed that ours is definitely working correctly by comparing our data

138
00:09:52,920 --> 00:09:58,380
running stock versus having only one die

139
00:09:56,100 --> 00:10:02,940
manually enabled in the BIOS so you can see here that in games our performance

140
00:10:00,779 --> 00:10:08,459
numbers are roughly the same between stock and forcing the V Cash die while

141
00:10:05,880 --> 00:10:12,899
forcing the frequency die often ran much slower all of which sounds totally

142
00:10:10,740 --> 00:10:17,940
normal and fine right with the right drivers x3d performs better for gaming

143
00:10:15,779 --> 00:10:23,519
what's the problem the problem is that it doesn't look at f122 at 4K with Ray

144
00:10:21,060 --> 00:10:29,120
tracing enabled this is a really heavy GPU load so our CPU should easily be

145
00:10:26,339 --> 00:10:35,760
able to hit Max boost but not the 7950x3d I mean we expect it to run at a

146
00:10:32,519 --> 00:10:39,060
lower clock speed than the 7950x but we

147
00:10:35,760 --> 00:10:41,220
didn't expect is this inconsistency I

148
00:10:39,060 --> 00:10:48,660
mean with adequate cooling both the 3900k and the 7950x are Rock Solid at

149
00:10:45,660 --> 00:10:51,120
5.5 gigahertz but the x3d it's all over

150
00:10:48,660 --> 00:10:56,579
the place that ended up leading to lower GPU power and therefore lower GPU

151
00:10:54,480 --> 00:11:03,120
performance and when we looked at the CPU package power we can see why the

152
00:11:00,200 --> 00:11:09,180
7950x3d in this gaming load is drawing less than half of its rated 120 watt TDP

153
00:11:06,959 --> 00:11:13,820
such low power draws certainly impressive but on a chip that advertises

154
00:11:11,540 --> 00:11:18,480
pinnacle-tier raw gaming performance it's not a behavior that we we want to

155
00:11:16,500 --> 00:11:23,399
see and our replacement chip behaved like this across three motherboards from

156
00:11:21,180 --> 00:11:27,779
three different vendors with multiple Expo certified RAM kits with fresh

157
00:11:25,500 --> 00:11:33,240
Windows installs and even with different chipset drivers so there were two things

158
00:11:30,480 --> 00:11:38,040
wrong with AMD's review kit to us the defective chip and the bad product

159
00:11:35,640 --> 00:11:44,760
positioning I mean if you like to drag race your 700 CPU at 1080p High hits the

160
00:11:42,420 --> 00:11:50,760
bee's knees but I think it's fair to say that most people with this kind of money

161
00:11:46,740 --> 00:11:52,320
are running 1440p monitors at least and

162
00:11:50,760 --> 00:11:56,640
probably don't mind turning on some in-game eye candy and for them our data

163
00:11:54,779 --> 00:12:01,260
simply doesn't support the conclusion that the ryzen 9 7950x3d is

164
00:11:59,000 --> 00:12:07,200
unquestionably the gaming performance Monarch that isn't to say it's a bad

165
00:12:03,600 --> 00:12:09,000
chip AMD in our mind just made the wrong

166
00:12:07,200 --> 00:12:14,279
call by calling out their less impressive and less defensible claims of

167
00:12:11,880 --> 00:12:17,279
performance leadership and instead we feel they should have put far more

168
00:12:15,660 --> 00:12:24,360
services on their true strength performance per watt I mean if we step

169
00:12:20,399 --> 00:12:26,760
back you are getting core I9 3900k ish

170
00:12:24,360 --> 00:12:31,700
performance for a substantial reduction in power consumption even in a GPU bound

171
00:12:29,459 --> 00:12:38,459
scenario like this one that's astounding mind-boggling even under a full Prime 95

172
00:12:34,800 --> 00:12:42,899
load the 13900k consistently pulls over

173
00:12:38,459 --> 00:12:45,600
320 Watts with the 7950x3d at less than

174
00:12:42,899 --> 00:12:49,980
half of that AMD should be shouting this from the rooftops but instead they were

175
00:12:48,060 --> 00:12:53,820
so focused on the gaming performance crowd that we followed them down that

176
00:12:51,899 --> 00:12:58,440
rabbit hole expecting to find the same thing but as it turns out the Dual die

177
00:12:56,040 --> 00:13:03,660
x3d chips work differently from most other CPUs and it looks like the cores

178
00:13:01,079 --> 00:13:08,160
will aggressively try to clock down or even sleep whenever possible when

179
00:13:05,760 --> 00:13:12,660
they're not fully loaded and that would explain what we observed with our GPU

180
00:13:10,139 --> 00:13:18,000
bound scenario and with our 1440p results it would also explain where some

181
00:13:14,820 --> 00:13:19,920
of that lost TDP went our best guess

182
00:13:18,000 --> 00:13:23,399
here is that AMD is airing on the side of caution when it comes to their x3d

183
00:13:21,779 --> 00:13:28,260
chips in order to keep them from redlining thanks to the 3D V cache being

184
00:13:25,740 --> 00:13:32,399
much more sensitive to voltage and to heat than a typical ryzen die this

185
00:13:30,540 --> 00:13:36,300
behavior of course means that there is some performance being left on the table

186
00:13:34,019 --> 00:13:41,040
and we can actually see some of that being clawed back with ryzen Master's

187
00:13:38,760 --> 00:13:45,000
curve tuner function while it isn't the fully unlocked overclocking that we were

188
00:13:42,779 --> 00:13:49,139
promised for x3d going forward it does do a decent job of undervolting the chip

189
00:13:47,339 --> 00:13:52,860
allowing it to push higher frequencies more often the performance difference

190
00:13:50,820 --> 00:13:56,339
isn't night and day but you are more likely to see that performance

191
00:13:54,180 --> 00:14:00,240
leadership that AMD is after assuming that your sample is more golden than

192
00:13:57,600 --> 00:14:05,279
yellow if you enable this function and in our GPU bound F1 run at 4K we can see

193
00:14:03,480 --> 00:14:10,019
that where the stock performance is all over the place it is much flatter with

194
00:14:07,740 --> 00:14:13,860
curved tuner enabled while also drawing even less power than it did before it's

195
00:14:11,940 --> 00:14:18,240
actually scary how efficient this chip is its performance characteristics like

196
00:14:15,720 --> 00:14:23,820
these which we first noticed to a lesser degree with the 65 watt non-x ryzen 7000

197
00:14:21,420 --> 00:14:27,899
CPUs that pushed us to start testing at multiple resolutions and that's the

198
00:14:26,339 --> 00:14:31,800
reason why we will continue to do so going forward it's also why we insist on

199
00:14:30,300 --> 00:14:36,959
re-running our tests with the latest drivers latest BIOS revisions and the

200
00:14:34,560 --> 00:14:40,860
same OS date as of the date that we start testing a new product we've been

201
00:14:38,820 --> 00:14:44,699
doing that for years now and while it means we need to rush more to make these

202
00:14:42,720 --> 00:14:49,320
deadlines because we are retesting every time we're committed to giving you guys

203
00:14:46,440 --> 00:14:54,120
a lesser quantity maybe of truly up-to-date data versus a greater amount

204
00:14:51,600 --> 00:14:57,660
of outdated data the answer to the question then of whether the Dual die

205
00:14:55,800 --> 00:15:03,600
x3d chips are worth the extra hundred dollars two weeks and 2300 words later

206
00:14:59,839 --> 00:15:05,339
is probably not if you need all of the

207
00:15:03,600 --> 00:15:09,000
strengths it brings to the table good gaming performance double-digit core

208
00:15:07,019 --> 00:15:15,360
counts and industry-leading power efficiency that's the big one then maybe

209
00:15:12,899 --> 00:15:19,019
but it comes at the cost of worse productivity performance relying on

210
00:15:17,220 --> 00:15:23,940
chipset drivers to manage the dissimilar dies and the inability to overclock in

211
00:15:21,779 --> 00:15:28,740
the traditional sense so I suspect that most people would be better off with the

212
00:15:25,920 --> 00:15:34,800
regular or X variants of the 7950 and 7900 or better off waiting for the 7800

213
00:15:32,279 --> 00:15:38,279
X 3D but should be able to clock Higher by putting more of its TDP towards its

214
00:15:37,079 --> 00:15:42,180
single die we're going to have a full review of that chip when it launches but

215
00:15:40,079 --> 00:15:47,579
for now we have this message from our sponsor backblaze World backup day is

216
00:15:45,180 --> 00:15:51,540
March 31st and backlays is here to help you back up your data with their easy to

217
00:15:49,380 --> 00:15:55,860
use Cloud solution that starts at just seven dollars a month they make it

218
00:15:54,060 --> 00:15:59,760
simple by allowing you to back up almost anything from your Mac or PC and access

219
00:15:58,079 --> 00:16:04,019
it anywhere in the world with their web and mobile apps backblaze also lets you

220
00:16:01,860 --> 00:16:07,440
restore your data by mail they will ship a hard drive with your data right to

221
00:16:05,579 --> 00:16:10,920
your door and after you're done you can return the hard drive for a refund and

222
00:16:09,540 --> 00:16:15,240
if you're worried about accidentally deleting files you can increase your

223
00:16:12,660 --> 00:16:20,760
data retention history to a year for an extra two dollars a month two dollars a

224
00:16:18,000 --> 00:16:24,660
month with over 55 billion files restored and two exabytes of data under

225
00:16:22,920 --> 00:16:28,320
their management backblaze has got you covered so don't be that person who

226
00:16:26,699 --> 00:16:33,180
forgets to back up their important files sign up and get a 15-day free trial with

227
00:16:31,380 --> 00:16:38,040
no credit card required today at backblaze.com LTT thanks for watching

228
00:16:36,120 --> 00:16:43,019
guys this was a very different format from our usual CPU reviews so maybe go

229
00:16:40,380 --> 00:16:46,920
check out our review of the ryzen 70065 watt chips for a little bit more about

230
00:16:45,060 --> 00:16:49,699
the other options available to you on socket am5