John said:
Do you think that on-chip optical links will ever make sense? A laser
and a photoreceiver seem like a lot of overhead to move signals a cm
or so.
Fast serial inter-chip links seem a little more practical, but they
will still need a fast, low-noise receiver and connectors. Equalized
electrical connections keep getting better... somebody just
demonstrated a 10 GBPS link that works over a couple hundred feet of
unshielded CAT6.
Fiber is great over distances where electrical losses get untenable...
hundreds of meters or hundreds of km.
I think that optics will continue to move in from the long-distance end
toward shorter and shorter links, but how far it gets is anybody's
guess. You're right that the wire guys are smart, and that wire can
still be improved quite a bit. (In fact one of the main problems I see
is that the board-to-board stuff is all 850 nm multimode, whereas the
on-chip stuff is shaping up to be 1.55 um single mode, polarization
preserving, for good reasons of material absorption and
manufacturability in both cases. In reality the gazinta has to connect
with the gazouta.)
The difficulty with equalizing electrical links--pre-emphasis,
decision-feedback equalization, etc.--is that it eats power. Since the
wire attenuation is what it is, that 10 Gb/s Cat6 link has to work by
beating the daylights out of the near end so that something vaguely
recognizable can come out of the DFE at the other end. You can't do
that on chip, because you need thousands of those links (just counting
the long ones), and want to keep the interconnect power below, say, 30W.
Putting a lot of EQ on also requires higher supply voltages.
Wire has some clear limits. The capacitance of a copper wire is going
to be 2 pF/cm forever and ever, and you aren't going to improve the
conductivity of copper either. Damascene processes (where the copper
sits in a liner of refractory metal) make this worse by reducing the
cross sectional area of Cu. You probably can't make reliable electrical
links with less than 200 mV of differential swing. That means that wire
bottoms out in the 1 pJ/bit range, whereas optics can potentially get
below 50 fJ/bit. We think that the crossover comes around the 32-nm node.
If you have the chance, take a look at the current ITRS technology
roadmap--it has charts of the technologies required for each node, and
the boxes are colour-coded green, yellow, or red. The interconnection
section of the 32-nm node is solid red, meaning that nobody has a clue
how to do it, despite having beaten on the wiring problem for decades.
That means that optics is a good contender to be the long-wire
interconnection solution for 32-nm and 22-nm.
As you say, it won't happen with garden variety optical links--which is
why it's a fun area to work in.
Cheers,
Phil Hobbs
