AMD, ​​​​​​​AMD, ​​​​​​​AMD, ​​​​​​​AMD, ​​​​​​​AMD, ​​​​​​​AMD, ​​​​​​​AMD, ​​​​​​​AMD, ​​​​​​​AMD, ​​​​​​​AMD
Ryzen AMD

AMD Ryzen, Faster Memory, and the Infinity Fabric

AMD Ryzen AMD Ryzen AMD Ryzen AMD Ryzen AMD Ryzen AMD Ryzen AMD Ryzen AMD Ryzen 

1
00:00:00,030 --> 00:00:04,470
following the initial release of AMD

2
00:00:02,760 --> 00:00:07,109
Rison there was a ton of discussion

3
00:00:04,470 --> 00:00:09,540
regards to using higher clocked system

4
00:00:07,109 --> 00:00:10,950
ram particularly in games the

5
00:00:09,540 --> 00:00:12,599
performance benefits of having higher

6
00:00:10,950 --> 00:00:14,400
clock memory was something that we

7
00:00:12,599 --> 00:00:16,350
didn't necessarily see an Intel side

8
00:00:14,400 --> 00:00:18,420
it's just not now the architecture is

9
00:00:16,350 --> 00:00:20,340
made it's not made to benefit from

10
00:00:18,420 --> 00:00:21,689
higher clock memory but rise ins

11
00:00:20,340 --> 00:00:23,400
architecture is and that's what I want

12
00:00:21,689 --> 00:00:26,580
to talk about in this video if it's

13
00:00:23,400 --> 00:00:28,980
worth buying really expensive high clock

14
00:00:26,580 --> 00:00:30,929
RAM out of the box for the sake of a

15
00:00:28,980 --> 00:00:33,120
rise in CPU and a lot of the people who

16
00:00:30,929 --> 00:00:34,860
spend you know 400 bucks on an 1800 X

17
00:00:33,120 --> 00:00:36,420
are probably gonna buy a higher clock to

18
00:00:34,860 --> 00:00:39,360
rise in memory anyway at least a memory

19
00:00:36,420 --> 00:00:40,500
that's capable of reaching 3000 or 3,200

20
00:00:39,360 --> 00:00:41,670
maker it's on the rise in platform which

21
00:00:40,500 --> 00:00:43,469
is difficult to do we'll talk about that

22
00:00:41,670 --> 00:00:46,079
as well but for those who are in the

23
00:00:43,469 --> 00:00:48,030
mid-range Rison five-tier should you

24
00:00:46,079 --> 00:00:50,219
still consider fast memory and why does

25
00:00:48,030 --> 00:00:57,629
it even matter in the first place I talk

26
00:00:50,219 --> 00:01:00,149
about all that in this video now when it

27
00:00:57,629 --> 00:01:02,640
comes down to it all rise in CPUs on a

28
00:01:00,149 --> 00:01:04,350
fabrication level are basically the same

29
00:01:02,640 --> 00:01:05,580
there are a few imperfections in some of

30
00:01:04,350 --> 00:01:07,229
the cores and that is ultimately

31
00:01:05,580 --> 00:01:09,780
differentiates a rise in three cpu from

32
00:01:07,229 --> 00:01:11,580
a rise in seven cpu so of the eight

33
00:01:09,780 --> 00:01:12,600
cores that are packed into every Rison

34
00:01:11,580 --> 00:01:15,540
chip no matter if you have a rise in

35
00:01:12,600 --> 00:01:17,310
three up to Horizon seven CPU if four of

36
00:01:15,540 --> 00:01:18,509
those aren't doing too well then AMD is

37
00:01:17,310 --> 00:01:19,830
gonna just disable and they're just

38
00:01:18,509 --> 00:01:21,840
gonna turn them off it's better than

39
00:01:19,830 --> 00:01:23,880
making an entirely new fabrication

40
00:01:21,840 --> 00:01:25,680
process for a four core chip this way

41
00:01:23,880 --> 00:01:27,299
they can save money makes perfect sense

42
00:01:25,680 --> 00:01:28,890
I don't blame them for doing that so a

43
00:01:27,299 --> 00:01:30,720
before remaining course if these can

44
00:01:28,890 --> 00:01:33,210
handle power loads pretty well and the

45
00:01:30,720 --> 00:01:35,400
the performance degradation isn't too

46
00:01:33,210 --> 00:01:39,090
substantial then these might be arisin

47
00:01:35,400 --> 00:01:40,619
five 1400s or 1500 X's if they aren't

48
00:01:39,090 --> 00:01:43,110
doing too well with handling a lot of

49
00:01:40,619 --> 00:01:44,850
power and the schedulers aren't doing

50
00:01:43,110 --> 00:01:46,409
that great a job at increasing the

51
00:01:44,850 --> 00:01:48,509
performance level then they might be

52
00:01:46,409 --> 00:01:50,130
rising three CPUs with multi-threading

53
00:01:48,509 --> 00:01:52,290
and disabled so it really comes down to

54
00:01:50,130 --> 00:01:55,200
what AMD deems as acceptable on both

55
00:01:52,290 --> 00:01:57,000
power delivery and execution levels now

56
00:01:55,200 --> 00:01:58,710
the difference between AMD and Intel

57
00:01:57,000 --> 00:02:01,799
because Intel does this too with bidding

58
00:01:58,710 --> 00:02:03,420
is that AMD decides well if these four

59
00:02:01,799 --> 00:02:06,180
cores aren't going to cut it we're gonna

60
00:02:03,420 --> 00:02:08,369
disable two of the cores in one cc X

61
00:02:06,180 --> 00:02:11,250
which is a four core cluster in a Rison

62
00:02:08,369 --> 00:02:13,280
CPU and two cores in the other CC X

63
00:02:11,250 --> 00:02:16,069
which is the other cluster of

64
00:02:13,280 --> 00:02:18,560
horizon CPU every rise in CPU has to CC

65
00:02:16,069 --> 00:02:20,209
x's per die the red Ripper has four

66
00:02:18,560 --> 00:02:22,670
which is why you can get up to 16 cores

67
00:02:20,209 --> 00:02:24,470
in those CPUs maybe this diary shot will

68
00:02:22,670 --> 00:02:26,180
help explain things a bit better so if

69
00:02:24,470 --> 00:02:27,740
two cores are activated in one cc X and

70
00:02:26,180 --> 00:02:29,630
two are activated in the other there has

71
00:02:27,740 --> 00:02:31,040
to be some sort of efficient way for

72
00:02:29,630 --> 00:02:32,930
them both to communicate because these

73
00:02:31,040 --> 00:02:35,000
are essentially two separate dies they

74
00:02:32,930 --> 00:02:36,890
might not be that physically but each

75
00:02:35,000 --> 00:02:39,620
core cluster has an own set of resources

76
00:02:36,890 --> 00:02:41,330
including l3 cache which means that when

77
00:02:39,620 --> 00:02:43,040
you're in especially intense situations

78
00:02:41,330 --> 00:02:44,480
on a computational level and they have

79
00:02:43,040 --> 00:02:47,330
to share information and process

80
00:02:44,480 --> 00:02:49,970
information simultaneously and exchange

81
00:02:47,330 --> 00:02:52,190
resources there has to be an efficient

82
00:02:49,970 --> 00:02:54,380
Highway a way for data to be transferred

83
00:02:52,190 --> 00:02:57,050
super quick between the cores to

84
00:02:54,380 --> 00:02:59,660
eliminate that latency this highway is

85
00:02:57,050 --> 00:03:01,489
called the Infinity fabric as Tom's

86
00:02:59,660 --> 00:03:02,750
Hardware puts it the large amount of

87
00:03:01,489 --> 00:03:05,060
data flowing through this pathway

88
00:03:02,750 --> 00:03:07,430
requires a lot of scheduling magic to

89
00:03:05,060 --> 00:03:09,080
ensure a high quality of service it's

90
00:03:07,430 --> 00:03:11,330
also logical to assume that these six

91
00:03:09,080 --> 00:03:14,090
and four core models benefit from less

92
00:03:11,330 --> 00:03:15,140
cross CCX traffic compared to the eight

93
00:03:14,090 --> 00:03:16,430
core models

94
00:03:15,140 --> 00:03:19,579
so they're actually hinting at here is

95
00:03:16,430 --> 00:03:22,040
the dual CCX design is a blessing in a

96
00:03:19,579 --> 00:03:23,630
way and a curse it's a blessing for AMD

97
00:03:22,040 --> 00:03:25,519
from a financial standpoint because all

98
00:03:23,630 --> 00:03:27,709
they have to do is slap multiple CC exes

99
00:03:25,519 --> 00:03:29,720
into the same chip and just run infinity

100
00:03:27,709 --> 00:03:31,100
fabric all between them which saves them

101
00:03:29,720 --> 00:03:32,600
money because they can control how many

102
00:03:31,100 --> 00:03:34,100
CC exes are in each dial without

103
00:03:32,600 --> 00:03:36,680
completely redoing the fabrication

104
00:03:34,100 --> 00:03:39,380
process but the curse involved is the

105
00:03:36,680 --> 00:03:41,870
latency involved between the CC X data

106
00:03:39,380 --> 00:03:44,239
transfers so the Infinity fabric itself

107
00:03:41,870 --> 00:03:46,820
is not the most efficient means by which

108
00:03:44,239 --> 00:03:49,250
data is transferred from one CC X to

109
00:03:46,820 --> 00:03:50,480
another this is where the faster memory

110
00:03:49,250 --> 00:03:52,549
comes into play it's but this whole

111
00:03:50,480 --> 00:03:54,170
video is about why rise and benefits

112
00:03:52,549 --> 00:03:55,310
from faster memory we hear people say

113
00:03:54,170 --> 00:03:57,859
that maybe they don't know what they're

114
00:03:55,310 --> 00:03:59,690
talking about but they are correct in a

115
00:03:57,859 --> 00:04:01,850
sense it really depends on the number of

116
00:03:59,690 --> 00:04:04,489
cores enabled per cc X but ultimately

117
00:04:01,850 --> 00:04:06,470
Rison will benefit more from faster

118
00:04:04,489 --> 00:04:07,940
memory than Intel will and the reason

119
00:04:06,470 --> 00:04:09,769
why is because the Infinity fabrics

120
00:04:07,940 --> 00:04:12,049
speed this the rate at which it can

121
00:04:09,769 --> 00:04:14,600
transfer data is directly tied to the

122
00:04:12,049 --> 00:04:16,640
frequency of your system memory Intel is

123
00:04:14,600 --> 00:04:19,400
independent on the same variable because

124
00:04:16,640 --> 00:04:22,580
their consumer-grade CPUs are reliant on

125
00:04:19,400 --> 00:04:24,960
the ring bus design two rings share

126
00:04:22,580 --> 00:04:26,820
information between let's say 4 cores

127
00:04:24,960 --> 00:04:29,789
and basically you have less traffic

128
00:04:26,820 --> 00:04:32,009
being dispatched between cores so the

129
00:04:29,789 --> 00:04:34,410
the congestion isn't as great as it

130
00:04:32,009 --> 00:04:36,120
would be on an 8-core Rison cpu because

131
00:04:34,410 --> 00:04:37,620
then you have four cores trying to share

132
00:04:36,120 --> 00:04:40,020
information down let's say a single

133
00:04:37,620 --> 00:04:41,130
pathway with the Infinity fabric so

134
00:04:40,020 --> 00:04:43,949
there can be quite a bit of congestion

135
00:04:41,130 --> 00:04:46,320
and lag I shouldn't say lag it's more or

136
00:04:43,949 --> 00:04:49,169
less a gamer term but latency it's the

137
00:04:46,320 --> 00:04:50,880
delay in data transfer and we see a huge

138
00:04:49,169 --> 00:04:53,190
difference between the rain bus design

139
00:04:50,880 --> 00:04:56,250
with Intel a comparable Intel CPU and

140
00:04:53,190 --> 00:04:59,550
the Infinity fabric whereas cores on an

141
00:04:56,250 --> 00:05:03,120
i7 7 700 K rely on the latency roughly

142
00:04:59,550 --> 00:05:06,330
between 30 and 40 nanoseconds cross

143
00:05:03,120 --> 00:05:09,570
quartic or latency on a cc x-ray for

144
00:05:06,330 --> 00:05:12,630
Horizon is somewhere in the realm of 200

145
00:05:09,570 --> 00:05:14,940
nanoseconds almost 10 times the latency

146
00:05:12,630 --> 00:05:17,039
just because data has to be transferred

147
00:05:14,940 --> 00:05:19,320
across an infinity fabric not to be

148
00:05:17,039 --> 00:05:21,419
frank we're talking nanoseconds here not

149
00:05:19,320 --> 00:05:22,650
even milliseconds but it does add up

150
00:05:21,419 --> 00:05:24,990
over time as huge computational

151
00:05:22,650 --> 00:05:26,940
workloads bottleneck that infinity

152
00:05:24,990 --> 00:05:29,010
fabric you can imagine how things get

153
00:05:26,940 --> 00:05:30,330
pretty backed up in the long run and you

154
00:05:29,010 --> 00:05:33,510
might even be able to tell a difference

155
00:05:30,330 --> 00:05:35,550
in things like games and also in render

156
00:05:33,510 --> 00:05:37,470
times if you do have heavy workloads

157
00:05:35,550 --> 00:05:39,419
being pushed to rise in CPUs they might

158
00:05:37,470 --> 00:05:42,270
not be able to handle them as well with

159
00:05:39,419 --> 00:05:44,789
that lower clocked system Ram so there

160
00:05:42,270 --> 00:05:46,470
you have it yes AMD rice and CPUs do

161
00:05:44,789 --> 00:05:47,880
benefit from higher clock memory it's

162
00:05:46,470 --> 00:05:50,340
not a myth and the reason why is because

163
00:05:47,880 --> 00:05:51,900
AMD employs infinity fabric which is

164
00:05:50,340 --> 00:05:53,940
directly dependent on the speed of your

165
00:05:51,900 --> 00:05:55,289
RAM now getting too much into detail

166
00:05:53,940 --> 00:05:57,060
that's really all you need to know

167
00:05:55,289 --> 00:05:58,620
something else worth noting is that rise

168
00:05:57,060 --> 00:06:00,360
in 3cp is because they only have two

169
00:05:58,620 --> 00:06:01,919
cores activated per CCX aren't going to

170
00:06:00,360 --> 00:06:04,740
benefit as much from the higher clocked

171
00:06:01,919 --> 00:06:07,020
Ram only because the data transfer rates

172
00:06:04,740 --> 00:06:09,870
between CC X's are going to be as high

173
00:06:07,020 --> 00:06:11,849
because only two cores per CC X are

174
00:06:09,870 --> 00:06:15,720
actually sending information whereas in

175
00:06:11,849 --> 00:06:17,550
horizon 7 1700 700 X or 800 X CPU get

176
00:06:15,720 --> 00:06:19,020
four cores on each side sending a ton of

177
00:06:17,550 --> 00:06:21,270
information things can get pretty

178
00:06:19,020 --> 00:06:22,289
congested in there for more info on this

179
00:06:21,270 --> 00:06:24,060
maybe you'd like to read an article

180
00:06:22,289 --> 00:06:25,169
about the Infinity fabric I invite you

181
00:06:24,060 --> 00:06:26,430
to check out the link at the top this

182
00:06:25,169 --> 00:06:27,900
video's description sometimes it helps

183
00:06:26,430 --> 00:06:29,639
to just read something over and over

184
00:06:27,900 --> 00:06:31,320
until it clicks each it takes me 5 or 6

185
00:06:29,639 --> 00:06:33,330
times I'm like I'm like you're bringing

186
00:06:31,320 --> 00:06:35,250
the sentence until till something clicks

187
00:06:33,330 --> 00:06:36,330
you get that light bulb go off now ok I

188
00:06:35,250 --> 00:06:37,199
get it now

189
00:06:36,330 --> 00:06:38,669
it's harder to do that with a

190
00:06:37,199 --> 00:06:39,930
videocassette to keep hearing me say it

191
00:06:38,669 --> 00:06:41,490
over and over again it's less annoying

192
00:06:39,930 --> 00:06:42,990
when you're doing it yourself in your

193
00:06:41,490 --> 00:06:44,069
head and invite you to check that link

194
00:06:42,990 --> 00:06:46,110
out by the way it's linked to Tom's

195
00:06:44,069 --> 00:06:47,669
hardware and they have a great article

196
00:06:46,110 --> 00:06:49,080
breaking down the Infinity fabric and

197
00:06:47,669 --> 00:06:50,460
why it behaves the way it does also is a

198
00:06:49,080 --> 00:06:52,800
pretty cool benchmarks in there to back

199
00:06:50,460 --> 00:06:53,610
up the claims made in this video if you

200
00:06:52,800 --> 00:06:54,750
like this video be sure to give it a

201
00:06:53,610 --> 00:06:55,830
thumbs up I appreciate it thumbs down

202
00:06:54,750 --> 00:06:56,819
for the opposite click to subscribe but

203
00:06:55,830 --> 00:06:58,020
if you have any stay tuned for more

204
00:06:56,819 --> 00:07:01,460
content like this this is science studio

205
00:06:58,020 --> 00:07:01,460
thanks for learning with us

206
00:07:03,290 --> 00:07:10,989
[Music]