[Beowulf] materials for air shroud?

Lux, Jim (337C) james.p.lux at jpl.nasa.gov
Fri Sep 16 11:19:18 PDT 2011

On 9/16/11 10:52 AM, "mathog" <mathog at caltech.edu> wrote:

>Now that I have seen the problem for two sockets, it made me wonder
>the manufacturers did for 4 socket motherboard designs.  Easy enough to
>find out.
>A Google Image search for:
>    "quad socket" motherboard
>shows:  some took air flow into account, and some didn't.
>There are quad socket motherboards with the sockets in rectangular
>arrangements, so that
>air cooling all of those equally in a 1U or 2U case is going to be
>really hard, and probably
>it just isn't done.  Water cooling would work, or in a 3U or 4U air
>cooling by snaking in
>ducts, but either way it would be a PITA.
>Then there are other motherboards where air flow was clearly taken into
>account, and the sockets
>are lined up so that they are not in each other's air flow.  Or at
>least not much.  This one
>there might be a tiny bit of overlap right at the edge of the  heat

Having done PWB layout in the past, I can shed some light on this..

First off, doing a design for something this complex is really hard.  Not
only is there the whole thing of just getting all the wires to fit, but
there's a whole tradeoff of how many layers vs manufacturability (e.g. you
could make a 20 layer board, but that means you'd better have all 20 of
those layers aligned really well so that those tiny vias going from layer
10 to layer 17 through layers 11,12,13,14,15,and 16 actually work.

On top of this is all the problems with what's called signal integrity,
that is, do all the signals get to the correct place at the right time
without interference. So there's all these constraints on relative trace
lengths, which layers they can be on, and so forth.

And don't forget thermal management.  Quite a lot of heat is carried out
through the pins of the parts (after all, they have a direct physical
connection to the die), and that heat spreads through the board, carried
by (usually) the power planes, which are fairly solid sheets of copper.
(you can put dedicated heat spreader layers in, but that runs up the cost)

So what most people do is start with some previous design that is known to
work (e.g. Maybe you have a single socket or dual socket board you've been
making for a while) and you sort of cut and paste it.  You drag the
components around, letting the tool try to reroute, and you fix up the 5%
of the traces that just don't autoroute by hand.

So you can see why some boards truly are horrible and others aren't.
Sometimes, the original 1 or 2 CPU layout lends itself to replication and
other times it doesn't. I would think that if you look at the layout of
previous boards from the same mfr, you'll see the commonality.

And since the original design probably wasn't done contemplating tiling
it, it's sort of the luck of the draw whether it's suitable or not.

The CPU & bridge mfrs, by the way, generally provide a "reference design"
for their parts (in a form suitable for ingest into the CAD tool of your
choice), so that people can get up and manufacturing quickly, but that
reference design isn't necessarily optimized. And it almost always needs
to be tweaked for your particular board house's processes.  Every
manufacturer of boards has slightly different design rules and
manufacturing tolerances.  It would be the rare reference design where you
could just generate Gerber files with no changes and go get your board
fabricated and have it actually work.

More information about the Beowulf mailing list