AMD Ryzen, Faster Memory, and the Infinity Fabric

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]