1 Gbps - state of the art?

[Geronimo Stempovski]

Hi all,

I wonder what is curently state-of-the art in serial high-speed transmission
and what are the prevailing data rates? I know about some SerDes in the
gigabit-per-second range but I cannot imagine if 10 Gbps are really a
challenge or the applied method or if it's 1 Gbps (or something in
between)...?
I recently heard about some 60 GHz in the mobile communication sector and 10
Gbit Ethernet but as far as I know there are those multi-level modulation
methods (like QAM for example) that are able to provide 10 Gbit bandwidth
with a bitrate of some Mbps (is that correct?).
I'm not interested so much in those higher modulation methods (nor in
optical transmission) but in the baseband communication where bitrate =
clockrate, i.e. the line rate. What can be efficiently transmitted today
electrically (over wire or PCB)? What is the prevailing technology of those
circuits, is it CMOS or are there alternatives?
I am a senior electrical engineer and unfortunately did not manage to keep
up-to-date. After googling all night I'm really depressed because I finally
couldn't find an unambiguous answer.
Maybe some guys in the silicon-business or practitioners know the anser and
are willing to share there knoledge with me?

Best regards
Geronimo

 

[Joel Kolstad]

I'm not interested so much in those higher modulation methods (nor in
optical transmission) but in the baseband communication where bitrate =
clockrate, i.e. the line rate. What can be efficiently transmitted today
electrically (over wire or PCB)?

It's around 1Gbps that you really need to start paying attention to your board
materials, transmission lines, etc.: With inexpensive boards (e.g., FR-4),
you're at the point where you're starting to get significant loss, dispersion,
and distance limitations. Check out this file:
http://www.xilinx.com/esp/wired/optical/collateral/xaui_xgmii.pdf -- XAUI
achieves 10Gbps using four 3.125Gbps differential signals (there's overhead on
each one...), and they manage to run it 20" on cheap PCBs -- that's pretty
impressive.

 

[Austin Lesea]

Geronimo,

10 Gb/s electrical, or optical is not unusual, and even has been
standardized for many uses (ie SONET/STH OC-192).

Using wavelength division multiplexing there are commercial optical
systems with n times 10 Gb/s channels (each to its own color).

Nothing even fancy here. ON/OFF keying of the laser diode. Perhaps as
much as ten years old now.

10 Gb/s electrical is challenging, as you need to transmit the signal
the required distance without loss, noise, reflections, phase
distortion. That means perhaps 10 meters maximum with exotic material,
exotic electronics; and only a half a meter or less using regular
printed circuit boards and less exotic circuits.

PCI Express is a new standard that now is in every backplane of every
new PC, and offers up to 16 2.5 Gb/s channels. If that isn't "proof" of
a technology gaining a hold, I don't know what is.

In fact, there are so many applications from 622 Mb/s to ?? Mb/s that it
is difficult to keep track of all of them.

A suggested set of solutions for many of these standards:

http://direct.xilinx.com/bvdocs/userguides/ug196.pdf

and

http://direct.xilinx.com/bvdocs/userguides/ug197.pdf

for PCIe.

You will note all these standards use simple ON/OFF electrical keying,
with no complex signal modulation. Transmit symbol pre-shaping, and
receive signal equalization is used to maintain the eye opening in most
devices running higher than 3 Gb/s. Run length codes are used to
prevent long strings of zeroes or ones from being a challenge to IC
designers (8b10b, 64b66b).

Just one more reason why you tend to find a Xilinx FPGA inside just
about every box nowadays.

Austin

 

[PeteS]

Hi all,

I wonder what is curently state-of-the art in serial high-speed transmission
and what are the prevailing data rates? I know about some SerDes in the
gigabit-per-second range but I cannot imagine if 10 Gbps are really a
challenge or the applied method or if it's 1 Gbps (or something in
between)...?
I recently heard about some 60 GHz in the mobile communication sector and 10
Gbit Ethernet but as far as I know there are those multi-level modulation
methods (like QAM for example) that are able to provide 10 Gbit bandwidth
with a bitrate of some Mbps (is that correct?).
I'm not interested so much in those higher modulation methods (nor in
optical transmission) but in the baseband communication where bitrate =
clockrate, i.e. the line rate. What can be efficiently transmitted today
electrically (over wire or PCB)? What is the prevailing technology of those
circuits, is it CMOS or are there alternatives?
I am a senior electrical engineer and unfortunately did not manage to keep
up-to-date. After googling all night I'm really depressed because I finally
couldn't find an unambiguous answer.
Maybe some guys in the silicon-business or practitioners know the anser and
are willing to share there knoledge with me?

Best regards
Geronimo

In terms of bitrates, then I designed a board with serial links at 5Gb/s
*per pair* a couple of years ago. Before that I designed some switches
and gateways with 2.5Gb/s pairs (lots and lots of them). PCI Express has
just released the 5Gb/s signalling revision (within the last month or
so, I believe). The things I designed a few years ago were Infiniband
(my name is actually one of many on the latest spec).

Over FR4 (or other materials less than totally exotic) 5Gb/s is about
the most you'll get except for _very_ short runs, and as Joel notes
everything's a transmission line at those rates. I've seen 2.5Gb/s
Infiniband on a 4x cable (4 pairs in each direction) achieve 10 metres
within the signalling budget. At 5Gb/s things are more difficult.

10G ethernet is actually 4 signals, incidentally.

So state of the art in terms of practical, shipping and costs less than
a trip to the moon is currently in the 5Gb/s per pair range.

Cheers

PeteS

 

[[email protected]]

You will note all these standards use simple ON/OFF electrical keying,

with no complex signal modulation. Transmit symbol pre-shaping, and
receive signal equalization is used to maintain the eye opening in most
devices running higher than 3 Gb/s. Run length codes are used to
prevent long strings of zeroes or ones from being a challenge to IC
designers (8b10b, 64b66b).

What's the thumb rule limit of electrical "on-off" signaling with current
fpgas? (length, FR4/Cat.5, speed etc..)

I did some calculations on ordinary S-ATA, should be able to run through
5 meters of cat.5 just barely (maybe I missed something .

Just one more reason why you tend to find a Xilinx FPGA inside just
about every box nowadays.

Doesn't the competition have anything substantial to come with?

 

[Frithiof Andreas Jensen]

It's around 1Gbps that you really need to start paying attention to your board
materials, transmission lines, etc.: With inexpensive boards (e.g., FR-4),
you're at the point where you're starting to get significant loss, dispersion,
and distance limitations.

If you can get the physical size of the circuit well below one wavelength then
things becomes "simply" DC again (Well worth doing if you, say, happen to be
building a RADAR front-end).

 

[Austin Lesea]

"> Doesn't the competition have anything substantial to come with?"

Competition is always there.

Just visit the each vendor's websites to see their offerings. Not every
implementation is the same: some use more power, some less. Some have
more built in hardware function, some less. Xilinx may have been the
first to place the gigabit transceivers on a FPGA, but that was years
ago now, and offerings of FPGA with transceivers is common today.

I would like to think that PCIe+MGT+65nm+lowest power (Virtex 5 LX, LXT
shipping NOW) is a substantial offering, and the soonest any serious
competition is expected (from their hasty press releases) is late this
year (end of 2007).

So the simple answer for 65nm?

No, Xilinx is now one full year ahead of its competition. There is
absolutely no competition for a 65nm FPGA socket. We are shipping, and
no one else is, nor will they be for a very long time.

If you mean does anyone have gigabit transceivers, then the answer is
yes, and there is plenty of competition.

Austin

 

[Joel Kolstad]

If you can get the physical size of the circuit well below one wavelength
then
things becomes "simply" DC again (Well worth doing if you, say, happen to
be
building a RADAR front-end).

Yes, certainly... this is why Jim sometimes seems to have an easier job than
those of us doing board-level design. I was quite humbled once to learn
that a ~4"x10" PCB was having *significant* distributed behavior even down
at HF (30MHz), but after adding up all the trace lengths it became clear
that I had hit something over a quarter wavelength... oops!

I do remember reading how IC design generally uses lossy enough metals that
after the "lumped" approximation no longer becomes valid, switching to an RC
"transmission line" model is generally used, whereas with board-level design
usually LC is often most appropriate at first. (Eventually everything is
RLC, of course, or just true transmission lines if you have a simulator that
can deal with them -- SPICE sometimes has difficulty with lossy transmission
lines).

---Joel

 

[glen herrmannsfeldt]

Joel Kolstad wrote:

(snip)

I do remember reading how IC design generally uses lossy enough metals that
after the "lumped" approximation no longer becomes valid, switching to an RC
"transmission line" model is generally used,

Not necessarily lossy, it is part of scaling. As the wires shrink
in width, they also shrink in height, but often not in length.
(Chips are getting larger, instead of smaller.) Resistance per
unit length increases as the square of the shrinkage factor.
Capacitance per unit length decreases linearly with shrinkage,
so RC increases.

whereas with board-level design
usually LC is often most appropriate at first. (Eventually everything is
RLC, of course, or just true transmission lines if you have a simulator that
can deal with them -- SPICE sometimes has difficulty with lossy transmission
lines).

Board metal thickness doesn't usually decrease as the wires
get narrower. (Well, eventually maybe it will.)

-- glen