HardwareBanter - View Single Post

**Paul[_28_]** · June 30th 19, 03:13 AM posted to alt.comp.hardware.pc-homebuilt

wrote:
I was checking out an old (LGA1366) server, and noticed that it has 9 memory
slots - triple channel, 3 each.
Also have a desktop (LGA2011) with 8 slots - quad channel ,2 each.
So I presume then that each memory channel requires hundreds of pins, but
extra slots per channel only needs a few pins?
But then Ryzen AM4 needs 1331 pins, but only scores 2 memory channels.
Pins goes up to 4094 (Epyc 8 channel) or 3647 (Xeon 6 channel) but some of
those pins are for multi-socket interconnects.

That's pretty accurate.

The second DIMM on a bus, shares a lot of resources, and
only needs a few Chip Selects and clock signals to finish
the job. But you can't operate too many DIMMs on one channel,
because the loading screws up the impedance and reflections.

On busses with FB (fully buffered) DIMMs, you can run
three sticks per bus (just going from memory), and it's
because the memory controller only sees the "hot chip"
in the center of the DIMM, and cannot "feel" the loading
from all the memory chips. That's why the various schemes
have different limits. Consumer computing systems, practically
nothing is buffered.

The CPU chips also have a lot of "power pins". If a bond
wire has a 300mA limit, and your CPU uses up to 400W
of power (1V @ 400A), you can divide one into the
other, and see you need 1200+ VCC and 1200+ GND contacts.
So that's 2400 contacts right there. There are various
ways to do interconnect, and gold bond wires is the old way.

This is one of the reasons, that S478 went to S775. The
"bus math" did not change. There was no significant change
in logic signals. But "the Prescott team" figured the power
needed was about to become astronomical, so they added some
extra contacts to allow higher power chips. Which in the
end, probably wasn't necessary.

Theory says that 4094 is a bit too high, but I have
to tip my hat to people who make stuff like this work.
Imagine how many pounds of force it takes to do that!

I thought the record was around 5000 or 6000, using
a technology called "fuzz buttons". 4094 is pretty
high, for "conventional" technology, and conventional
techniques require a *lot* of contact pressure to make
them work, and make them "vibration/shock insensitive".
The "springs" in these things, have to be reliable
and not lose their spring force over years of usage.

Paul