1
00:00:00,000 --> 00:00:07,259
while AMD endured a relatively lackluster ryzen 7000 launch thanks to

2
00:00:05,279 --> 00:00:13,380
Fierce competition from everyone's favorite budget offering Intel in the

3
00:00:10,080 --> 00:00:15,780
data center AMD is pulling no punches

4
00:00:13,380 --> 00:00:21,300
with today's announcement of Epic Genoa they're bringing double-digit IPC

5
00:00:18,119 --> 00:00:24,720
improvements a maximum of 96 cores per

6
00:00:21,300 --> 00:00:25,980
socket a whopping 12 channels of ddr5

7
00:00:24,720 --> 00:00:31,619
memory I don't even have that many fingers and so much more

8
00:00:29,099 --> 00:00:35,640
we're going to tell you what's new and why Intel's upcoming Sapphire Rapids

9
00:00:33,719 --> 00:00:39,079
probably won't be able to compete with this

10
00:00:36,840 --> 00:00:45,899
spicy salami Mamma Mia oh quasi Picante comic questo

11
00:00:43,140 --> 00:00:48,719
sponsor Segway build Redux build Redux makes it easy to configure your new

12
00:00:47,280 --> 00:00:52,500
build with support guides and competitive pricing compared to building

13
00:00:50,399 --> 00:00:57,059
a PC yourself head to build redux.com Linus and start your new build today for

14
00:00:55,860 --> 00:01:01,320
those that haven't been paying attention let's get you up to speed AMD has been

15
00:00:59,699 --> 00:01:05,280
releasing their epic line of server processors based on the Zen architecture

16
00:01:03,059 --> 00:01:09,600
since 2017 with the first generation dubbed Naples

17
00:01:06,840 --> 00:01:15,780
not until the more recent epic Rome in 2019 in Milan in 2021 however did AMD

18
00:01:12,780 --> 00:01:19,500
actually manage to compete and then beat

19
00:01:15,780 --> 00:01:21,600
Intel taking the performance Crown this

20
00:01:19,500 --> 00:01:27,180
is in part due to core count Intel's monolithic style processors capped out

21
00:01:23,520 --> 00:01:29,580
at a measly 40 chorus per socket

22
00:01:27,180 --> 00:01:34,140
compared to AMD's 64 core chiplet designs back in 2021 Intel hoped to

23
00:01:32,580 --> 00:01:38,820
compete against epic Milan with their own chiplet-based design codenamed

24
00:01:36,299 --> 00:01:43,920
Sapphire Rapids but ended up delaying due to technical difficulties and poor

25
00:01:40,799 --> 00:01:47,700
yields to a now confirmed January 2023

26
00:01:43,920 --> 00:01:49,740
launch and they're still on 10 nanometer

27
00:01:47,700 --> 00:01:54,479
or Intel 7 now they call it Intel seven we don't don't say nanometer come on

28
00:01:51,780 --> 00:01:57,720
things are not looking great for Intel's Enterprise sector right now because

29
00:01:55,860 --> 00:02:02,640
honestly Genoa looks awesome the most obvious

30
00:02:00,479 --> 00:02:07,320
change with the latest generation is the New Zen 4 architecture running on tsmc's

31
00:02:05,340 --> 00:02:12,000
five nanometer process zenfor is more of a refinement of Zen 3

32
00:02:10,200 --> 00:02:16,980
rather than a complete redesign but while much of the micro architecture

33
00:02:14,099 --> 00:02:22,140
layout is the same improvements have been sprinkled all over a big chunk of

34
00:02:19,680 --> 00:02:26,220
the clean 14 IPC uplift comes from optimizations to the architecture front

35
00:02:24,060 --> 00:02:30,239
end with improved Branch prediction that can predict two branches per cycle

36
00:02:27,720 --> 00:02:35,400
rather than just one and a hefty bump to L2 cache they also introduced AVX 512

37
00:02:33,780 --> 00:02:38,520
support which is going to be a huge help for machine learning and AI workloads in

38
00:02:37,379 --> 00:02:42,900
the data center speaking of which the other major

39
00:02:40,560 --> 00:02:48,420
upgrade this time around is core counts across the 18 total skus launching for

40
00:02:46,140 --> 00:02:54,840
Genoa the max cores per socket has increased by a whopping 50 percent from

41
00:02:50,940 --> 00:02:58,260
64 to 96 cores with the minimum core

42
00:02:54,840 --> 00:03:00,360
count seeing a bump from 8 to 16 cores

43
00:02:58,260 --> 00:03:03,239
on the lowest end skus despite the power efficiency

44
00:03:01,800 --> 00:03:08,220
improvements of the new architecture maximum socket power consumption is

45
00:03:05,700 --> 00:03:14,459
jumping from 280 Watts on the last generation to a whopping 400 watts per

46
00:03:11,760 --> 00:03:18,599
socket a level that AMD's high performance compute customers have

47
00:03:16,440 --> 00:03:21,840
actually been asking for though the company tells us they don't expect the

48
00:03:19,980 --> 00:03:25,379
power envelope to climb much Beyond this point and don't expect to get much

49
00:03:24,000 --> 00:03:30,300
further into this video without a mention of our beanies on ltdstore.com

50
00:03:28,560 --> 00:03:34,080
uh they put LOL in there but you know what it's actually getting colder

51
00:03:31,560 --> 00:03:39,780
unfortunately with great power draw comes great socket demands and Genoa is

52
00:03:37,200 --> 00:03:46,140
no exception gone are the days of rectangular oppression say hello to our

53
00:03:43,440 --> 00:03:50,959
new Square overlords Jokes Aside genoa's new sp5 socket packs

54
00:03:49,440 --> 00:03:57,659
a whopping 6096 pins nice more like a like a nice

55
00:03:55,680 --> 00:04:03,000
backwards it's a similar trade design to the previous sp3 socket but it's now

56
00:04:00,780 --> 00:04:07,200
heatsink actuated rather than socket actuated that's fancy speak for the

57
00:04:05,459 --> 00:04:11,640
cooler providing the necessary pressure to ensure proper contact between the CPU

58
00:04:09,420 --> 00:04:15,120
and the socket's pins it's not dissimilar to the approach Intel has

59
00:04:13,319 --> 00:04:19,139
been taking for a few years now except AMD thoughtfully included the lockdown

60
00:04:17,459 --> 00:04:23,040
plate for the CPU to prevent you from accidentally dropping a chip into the

61
00:04:20,820 --> 00:04:28,380
socket which you wouldn't want to do since

62
00:04:24,020 --> 00:04:31,620
6096 nice pins are tiny head expensive

63
00:04:28,380 --> 00:04:33,780
however they are not as expensive as the

64
00:04:31,620 --> 00:04:40,919
copious amounts of memory this thing supports across 12 channels we're

65
00:04:37,020 --> 00:04:43,800
talking up to six terabytes of ddr5

66
00:04:40,919 --> 00:04:48,080
running at 4 800 Mega transfer per second allowing for a theoretical Peak

67
00:04:45,780 --> 00:04:53,699
throughput it a 480 gigabytes per second across the whole thing six terabytes all

68
00:04:51,419 --> 00:04:58,860
without taking too big of a hit in total latency oh and if that isn't enough fast

69
00:04:56,400 --> 00:05:06,180
memory for you AMD is introducing support for cxl

70
00:05:02,040 --> 00:05:08,699
wait what the hell cxl cxl is short for

71
00:05:06,180 --> 00:05:12,180
compute express link and it could very well be a game changer in the service

72
00:05:10,259 --> 00:05:16,500
space and potentially for Consumer Tech in the future it's a cash coherent

73
00:05:14,759 --> 00:05:21,780
interconnect standard built on top of PCI Express 5 that allows for huge

74
00:05:19,500 --> 00:05:25,259
increases to i o memory and storage flexibility

75
00:05:22,979 --> 00:05:30,060
in the same way that AMD uses PCI Express Lanes as Infinity Fabric in

76
00:05:26,940 --> 00:05:32,100
their dual socket servers cxl uses PCI

77
00:05:30,060 --> 00:05:36,900
Express Lanes as low latency high bandwidth interconnects allowing the

78
00:05:33,840 --> 00:05:38,520
system to pool or expand memory and

79
00:05:36,900 --> 00:05:43,139
improve the performance of various types of Hardware accelerators like gpus or

80
00:05:40,979 --> 00:05:47,759
network cards maybe you need an extra four terabytes of RAM for more

81
00:05:44,820 --> 00:05:52,919
virtualization I know I do slap in a cxl memory card and you are good to go and

82
00:05:50,340 --> 00:05:59,039
cxl lets you allocate that extra RAM to other devices if you want to generos

83
00:05:55,080 --> 00:06:00,900
supports cxl 1.1 plus type 3 which is

84
00:05:59,039 --> 00:06:06,900
the catchiest name I've read all week and is primarily focused on memory

85
00:06:03,479 --> 00:06:08,639
expansion rather than accelerators cxl

86
00:06:06,900 --> 00:06:12,060
is only just coming to the server Market but is primed to be a very important

87
00:06:10,740 --> 00:06:16,139
technology for the future of our computers so keep an eye on it it might

88
00:06:14,280 --> 00:06:19,740
come to Consumer Hardware sooner than you think if you're worried that you

89
00:06:18,000 --> 00:06:26,580
won't have enough PCIe lanes for all this stuff then don't be a single CPU

90
00:06:22,319 --> 00:06:30,900
provides you with 128 PCI Express Gen 5

91
00:06:26,580 --> 00:06:33,120
lanes and up to 160 in two CPU systems

92
00:06:30,900 --> 00:06:38,039
providing ample flexibility in terms of expansion and if you want to see just

93
00:06:35,880 --> 00:06:42,720
how much you can fit into these things super micro let us unveil their new h13

94
00:06:40,680 --> 00:06:47,160
Genoa Suite check out the link in the description to see some Hands-On

95
00:06:44,340 --> 00:06:51,300
Hardware content and last but certainly not least AMD introduced a few

96
00:06:49,620 --> 00:06:56,580
improvements to their security such as support for over a thousand encrypted

97
00:06:53,819 --> 00:07:00,479
virtualizations that is pretty neat especially for vps's what kind of

98
00:06:58,860 --> 00:07:04,979
perform from all of these slick new feature

99
00:07:02,039 --> 00:07:09,419
enhancements though for starters more than double the performance in integer

100
00:07:06,720 --> 00:07:14,160
arithmetic over Intel's 40 core Xeon 8380 Flagship which is expected

101
00:07:12,300 --> 00:07:19,800
considering AMD has more than double the cores and ends up being just an eight

102
00:07:16,800 --> 00:07:22,380
percent increase per core but the

103
00:07:19,800 --> 00:07:25,919
customer is buying 96 core CPUs like this one

104
00:07:23,819 --> 00:07:32,400
they would probably actually just want the more cores rather than faster cores

105
00:07:28,860 --> 00:07:34,380
so does AMD's dominance then carry over

106
00:07:32,400 --> 00:07:39,840
to the rest of the product stack according to AMD yes very yes

107
00:07:37,500 --> 00:07:44,460
but before we get too far please note that AMD has been pretty liberal with

108
00:07:42,780 --> 00:07:48,900
the scaling in their performance graphs so try to focus on the numbers

109
00:07:46,259 --> 00:07:55,919
themselves rather than the pictures AMD claims that their 24 core epic 9274f

110
00:07:53,520 --> 00:08:01,740
will compete with that same 40 core Intel CPU at least in integer operations

111
00:07:59,000 --> 00:08:06,479
calling integer operations general purpose computing

112
00:08:03,479 --> 00:08:08,099
is a bit of a stretch but

113
00:08:06,479 --> 00:08:13,560
technically assume that all of these CPUs will perform 25 percent worse than

114
00:08:11,039 --> 00:08:18,900
AMD promises you'll still see comparably cord AMD skus competing with Intel's

115
00:08:16,400 --> 00:08:23,280
flagships maybe you shouldn't be quite so cynical because the way AMD tested

116
00:08:21,660 --> 00:08:28,199
their CPUs for this could be leaving performance on the table let me explain

117
00:08:25,520 --> 00:08:32,399
AMD has embraced the Silicon Lottery with their new power management feature

118
00:08:29,759 --> 00:08:36,959
called determinism wow just like gonna take a philosophy there

119
00:08:34,380 --> 00:08:41,520
well it comes in two flavors performance and power oh it is philosophy it's kind

120
00:08:39,659 --> 00:08:45,420
of like AMD's Precision boost found on their ryzen chips and allows you to bump

121
00:08:43,620 --> 00:08:50,880
up your performance within spec if you have the thermal or power Headroom to do

122
00:08:47,279 --> 00:08:53,399
so in the performance determinism mode

123
00:08:50,880 --> 00:08:57,660
the CPU will Target the maximum performance at a given TDP with power

124
00:08:55,740 --> 00:09:01,860
varying to maintain stable performance power determinism is flipped with the

125
00:09:00,240 --> 00:09:06,240
CPU targeting the highest frequencies while maintaining stable power drop

126
00:09:03,779 --> 00:09:10,980
while a consumer might want more performance at all times Well if you run

127
00:09:08,640 --> 00:09:15,180
servers that vary and load having a constant power draw can actually improve

128
00:09:13,380 --> 00:09:19,980
efficiency and reduce stress on the components increasing the lifetime of

129
00:09:17,160 --> 00:09:24,360
your server now back to the graphs for these reported benchmarks AMD didn't use

130
00:09:22,320 --> 00:09:27,300
their determinism modes which means that customers could end up seeing extra

131
00:09:25,800 --> 00:09:31,500
performance on their chips should they decide to take advantage of this feature

132
00:09:28,860 --> 00:09:36,600
and even without that extra Edge AMD still claims massive per core increases

133
00:09:33,959 --> 00:09:41,100
over Intel's lineup promising as much as twice the performance per core in high

134
00:09:39,240 --> 00:09:46,860
performance Computing which makes them an obvious choice for simulation

135
00:09:43,140 --> 00:09:50,040
workloads AMD has far faster cores which

136
00:09:46,860 --> 00:09:51,959
helps them win an HPC and they have way

137
00:09:50,040 --> 00:09:57,540
more cores which gives them the upper hand virtualization and hyperscaling but

138
00:09:54,420 --> 00:09:59,760
here is the most damning comparison this

139
00:09:57,540 --> 00:10:07,680
performance per server chart AMD can fit the same number of cores in

140
00:10:03,540 --> 00:10:09,920
60 percent fewer servers using just half

141
00:10:07,680 --> 00:10:16,320
the power despite the extra socket draw at this point Intel isn't even competing

142
00:10:13,500 --> 00:10:20,040
they're sitting on the bench but all this performance of course means nothing

143
00:10:17,820 --> 00:10:24,060
if the price isn't right is AMD leveraging their performance leadership

144
00:10:21,660 --> 00:10:29,339
and charging through the nose well their Flagship

145
00:10:26,540 --> 00:10:35,580
9654 that I'm holding in my hand right here is coming in at effectively 12 000

146
00:10:33,000 --> 00:10:40,260
US Dollars which means the price per core remains virtually identical to

147
00:10:37,320 --> 00:10:44,519
their Milan Flagship but they are charging around a 10 premium on most of

148
00:10:42,720 --> 00:10:48,540
the product stack compared to Zen 3 which is a bit of a bump considering

149
00:10:46,800 --> 00:10:52,320
customers are likely going to need all new servers in addition to all new Chips

150
00:10:50,579 --> 00:10:56,459
thanks to the socket change and probably also all new RAM yeah it's gonna get

151
00:10:54,480 --> 00:10:59,399
pricey but with the ever increasing amount of computing being done in the

152
00:10:58,440 --> 00:11:04,440
cloud I'm sure that major players like AWS and

153
00:11:01,560 --> 00:11:08,100
Google will happily pay the premium or you know just have their own CPUs like

154
00:11:06,480 --> 00:11:12,959
you do and what's coming in the future well we know that we'll be seeing 3D V

155
00:11:11,279 --> 00:11:18,600
cash verses of these chips with the upcoming Genoa X lineup and that the

156
00:11:15,180 --> 00:11:21,660
Bergamo lineup will max out at 128 cores

157
00:11:18,600 --> 00:11:23,760
on a single die while we've yet to see

158
00:11:21,660 --> 00:11:28,260
what kind of upgrades Intel will bring with sapphire Rapids

159
00:11:25,740 --> 00:11:32,160
it's clear that AMD isn't resting on its Laurels employing every trick in the

160
00:11:30,360 --> 00:11:35,579
book to maintain the massive growth they've seen in the past couple of years

161
00:11:33,480 --> 00:11:39,420
in the server space and with the ever-growing presence and threat of

162
00:11:37,200 --> 00:11:45,480
ARM-based servers it's going to be tough for Intel to keep up now just make me

163
00:11:42,360 --> 00:11:48,060
some non-pro Zen 4 threadrippers and

164
00:11:45,480 --> 00:11:52,860
I'll be really happy almost as happy as I am to tell you about our sponsor MSI

165
00:11:50,339 --> 00:11:56,579
the holidays are coming quickly but MSI has you covered for gifts this year

166
00:11:54,360 --> 00:12:00,720
MSI's holiday deals are on now where you can save on their 12th gen laptops

167
00:11:58,260 --> 00:12:04,680
desktops peripherals and components do the kids need a new laptop for school or

168
00:12:03,000 --> 00:12:08,820
maybe your partner needs an upgrade to their outdated gaming rig there's

169
00:12:06,480 --> 00:12:14,220
something for everyone on sale now and for those of you in the US there's the

170
00:12:10,500 --> 00:12:16,019
hashtag timsi holiday giveaway tell MSI

171
00:12:14,220 --> 00:12:20,880
what you're thankful for and enter to win gift cards apparel and more so what

172
00:12:19,200 --> 00:12:25,380
are you waiting for check out MSI's deals and the hashtag Ty MSI holiday

173
00:12:23,399 --> 00:12:29,700
giveaway at the link below to start saving today thanks for watching guys

174
00:12:27,839 --> 00:12:34,620
obviously we haven't had a chance to put this guy through its Paces yet but maybe

175
00:12:32,100 --> 00:12:39,060
if you want to see some epic action we have our holy on the Epic Milan

176
00:12:37,920 --> 00:12:44,480
which was wow 2021

177
00:12:41,519 --> 00:12:44,480
how time flies