OCuLink Riser for the Supermicro AIOM Slot - clips/highlight.vtt

Overview Files Packages Releases Dependencies Issues Breadcrumbs Pull Requests Activity

main

John Lauer Highlight v2: what the card is for, captions and chapters in the adom-video player 0c50ebd 7d ago

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197

WEBVTT

00:00:00.400 --> 00:00:05.096
A YouTube viewer asked for an OCuLink
riser for the AIOM slot on their

00:00:05.096 --> 00:00:06.304
Supermicro server.

00:00:06.304 --> 00:00:09.256
Before we build it, here is what that
means.

00:00:20.500 --> 00:00:23.854
This is the kind of server the viewer has:

00:00:23.854 --> 00:00:25.850
a Supermicro rack server.

00:00:25.850 --> 00:00:32.318
Inside, besides the normal P C I e slots,
newer Supermicro boards have one or two

00:00:32.318 --> 00:00:33.196
AIOM slots.

00:00:34.600 --> 00:00:37.976
From the back, Supermicro labels them:

00:00:37.976 --> 00:00:40.997
P C I e by sixteen, AIOM, OCP 3.0.

00:00:40.997 --> 00:00:47.128
A full sixteen lane slot, the same
bandwidth as a graphics card slot.

00:00:48.520 --> 00:00:50.584
But only one kind of card fits:

00:00:50.584 --> 00:00:52.914
an AIOM network card like this one.

00:00:52.914 --> 00:00:57.640
If you don't need another network card,
those sixteen lanes sit unused.

00:00:59.040 --> 00:01:00.799
This board changes that.

00:01:00.799 --> 00:01:06.662
It plugs into the AIOM slot in place of
the network card, and splits the sixteen

00:01:06.662 --> 00:01:11.352
lanes into four OCuLink ports, each a full
P C I e by four link.

00:01:12.760 --> 00:01:18.420
Run an OCuLink cable from each port to a U
dot two NVMe drive, and the server gains

00:01:18.420 --> 00:01:23.808
four more fast drives, each running at
full Gen four by four speed, about eight

00:01:23.808 --> 00:01:25.240
gigabytes per second.

00:01:26.640 --> 00:01:31.888
Or cable a port to an OCuLink dock like
this one, and hang a graphics card or

00:01:31.888 --> 00:01:35.568
another P C I e card off the server,
outside the case.

00:01:36.960 --> 00:01:38.269
There is one catch:

00:01:38.269 --> 00:01:42.607
the server's BIOS has to support splitting
that slot four ways.

00:01:42.607 --> 00:01:47.635
When it does, an unused network slot
becomes four general purpose P C I e

00:01:47.635 --> 00:01:48.048
ports.

00:01:49.370 --> 00:01:50.427
Now the build.

00:01:50.427 --> 00:01:54.653
We typed that one request into Claude, in
Adom Hydrogen.

00:01:54.653 --> 00:02:00.314
It reads the OCP 3.0 spec and the OCuLink
standards, pinout tables and all.

00:02:09.370 --> 00:02:14.713
It writes a small field solver and finds
trace widths for 85 ohm differential pairs

00:02:14.713 --> 00:02:16.258
on a six layer stack-up.

00:02:29.370 --> 00:02:34.171
Thirty six high speed pairs are routed
from the gold fingers to four OCuLink

00:02:34.171 --> 00:02:36.066
connectors, with no crossings.

00:02:49.370 --> 00:02:54.283
Every lane in a port is tuned to arrive
within a hundredth of a picosecond of its

00:02:54.283 --> 00:02:54.890
neighbors.

00:03:04.370 --> 00:03:05.548
The design record:

00:03:05.548 --> 00:03:10.850
impedance per layer, skew inside every
pair under two hundredths of a millimeter.

00:03:24.370 --> 00:03:29.822
On the adom-aiflow timeline, the run moves
through design rules:

00:03:29.822 --> 00:03:32.122
JLCPCB's six layer process.

00:03:41.370 --> 00:03:43.801
Then the real board opens in KiCad:

00:03:43.801 --> 00:03:47.274
gold fingers at the bottom, connectors at
the top.

00:03:59.370 --> 00:04:04.818
Forty seven parts land live, one at a
time, each its own undo step.

00:04:19.370 --> 00:04:24.268
Then one hundred sixty nine traces land
the same way, every one checked against

00:04:24.268 --> 00:04:25.322
the design rules.

00:04:39.370 --> 00:04:42.141
The current check needs a real field
solve.

00:04:42.141 --> 00:04:43.816
Claude finds why it fails:

00:04:43.816 --> 00:04:47.554
a tenth of a millimeter via ring the
solver could not see.

00:04:59.370 --> 00:05:02.970
With a finer grid, current and thermal
both pass.

00:05:14.370 --> 00:05:18.617
Clean silkscreen, and the finished board
in KiCad's 3D viewer:

00:05:18.617 --> 00:05:20.946
every test point and port labeled.

00:05:32.370 --> 00:05:37.329
Gerbers for all six copper layers, drill
files, pick and place and the bill of

00:05:37.329 --> 00:05:40.890
materials, each read back and checked
against the board.

00:05:57.370 --> 00:05:58.413
Finish passes:

00:05:58.413 --> 00:06:01.543
zero design rule errors, zero unconnected.

00:06:01.543 --> 00:06:07.280
A manufacturable PCI Express Gen four
card, from a one line comment, designed

00:06:07.280 --> 00:06:10.186
and checked by Claude in Adom Hydrogen.