Re: [SI-LIST] : Clean sheet of paper

About this list Date view Thread view Subject view Author view

From: D. C. Sessions (dc.sessions@vlsi.com)
Date: Wed Dec 22 1999 - 14:27:18 PST


Jim Freeman wrote:
>
> Hi DC,
> You Are stuck with 3.3v technology on the ring , but the interior can be much higher performance with dual gat oxide processes.

Ummm... yes, but that's the point. You've shackled your I/O performance when
much higher performance is possible (and probably needed.) It makes *no* sense
to have two 150 nm devices with core voltages of 1.5 v communicating by demuxing
internal 1.2 GHz signals down to 150 MHz, pumping them up through level shifters
to 3.3v, using low-performance transistors to drive the line, burning lots of
power and radiating like crazy shipping them to another IC which then has to use
low-performance transistors to receive the signal ans shift it down to 1.5 v
signals to be muxed up to 1.2 GHz again.

Rube Goldberg would have blushed.

(Not to belabor the point but I'm an I/O cell designer and rather more
familiar with the details of dual-oxide processes than I'd really like!)

> "D. C. Sessions" wrote:
>
> > Chris.H.Simon@gd-is.com wrote:
> > >
> > > DC,
> > >
> > > I like your suggestion to start with a clean sheet of paper and ask
> > > ourselves what should be done to optimize the I/O design.
> > >
> > > In my experience working with computer system architects, they always want
> > > more and more bandwidth, which in the past has meant more and more pins at
> > > the boundary of an IC and more and more pins through connectors. And also
> > > increased data rates. To combat the higher data rates (and associated
> > > issues of noise, SSO, timing, ...) I have suggested to them to use
> > > differential signalling. The complaint that I get back is that this takes
> > > more pins. "I don't want to use two pins per signal when I can get twice
> > > as many signals using single ended." they whine. Of course this isn't
> > > really true since an increasing number of ground and power pins are
> > > required, but system architects don't consider these since they don't show
> > > up on any block diagram.
> >
> > That's what grouches like me are for :-)
> > We make nuisances of ourselves pointing out that SSO effects are
> > the #1 limiter on available performance. That using DC-balanced
> > signaling cuts SSO effects more than the equivalent number of pins
> > dedicated to power connections. That the balanced signals cause
> > less crosstalk and less EMI. That they can cut the cost of their
> > termination supplies dramatically and reduce the number of PWB layers
> > required.
> >
> > Of course, if those considerations aren't important to them, then
> > we'll be perfectly happy to charge extra for more advanced process
> > technology, packaging, and core area to make up the difference.
> >
> > > If it is really true that "Padrings are some of the most expensive real
> > > estate around, so pin count should be minimized." then why don't we start
> > > using each precious location on the padring to get more than one signal?
> >
> > We do. For busses, we crank up the clock rate and lower the number of
> > lines needed. If we have more bandwidth per pin than a single channel
> > can use, we use packet burst protocols.
> >
> > > I'm suggesting keeping differential signalling to alleviate some of the SI
> > > issues, but putting more than one logical signal on each differential pair.
> >
> > Basically broadband. The desirability of broadband depends a lot on both
> > the nature of your traffic and the limitations of your interconnect. For
> > short interconnects it's not really attractive for quite a while yet (we
> > did a science project of this sort recently. No, I can't discuss it.)
> >
> > > With two logical signals per pair I'm back to the one signal to one wire
> > > ratio that system architects love. See U.S. patent #5,872,813 "Dual
> > > Differential and Binary Data Receiver Arrangement" as an example. Although
> > > that patent refers to bipolar ECL-like circuits, I believe that some
> > > similar concepts could be implemented in CMOS circuits.
> > >
> > > The added complexity of the driver and receiver (and a little more power
> > > due to increased voltage swing) may be worth it if we gain one or more
> > > logical signals for each precious pin on the IC.
> >
> > Don't underestimate the grief that that "little more ... voltage swing"
> > causes. It really messes with your S/N ratio, but worst of all is the
> > fact that it runs up against the voltage-scaling objective. If you use
> > voltages larger than the native supply for a technology, you have to
> > degrade performance to what might as well be the technology appropriate
> > for that voltage. IOW, if you inisist on 3.3v signaling you're going to
> > be stuck with 350 nm CMOS performance.
> >
> > > ***************************************************************
> > >
> > > With the year wrapping up and my inbox filling with
> > > "Out of Office Autoresponse" messages, I thought I'd
> > > kick off something more interesting than the joys of LVDS.
> > >
> > > In particular, what would we use for signaling if we could
> > > start with a totally clean sheet of paper? Rather than
> > > immediately jump to a solution, I'm looking for some criteria:
> > >
> > > * It has to be scalable. Given silicon technology trends, it
> > > should migrate gracefully to lower-voltages and less
> > > voltage-stress-tolerant semiconductors.
> > >
> > > * It has to be SI clean. Output impedance should be matched
> > > (stringency variable) to the line across the switching range.
> > > Inputs switchpoints should be symmetrical and well-defined
> > > (ie differential receivers). Power plane proliferation
> > > leads to bad SI and wasted money, so separate termination
> > > supplies are a Bad Thing.
> > >
> > > * It has to be versatile. Single-ended, balanced single-ended, or
> > > differential; multidrop or point-to-point; uni- or bidirectional;
> > > all should be minor variations on the same system.
> > >
> > > * It should be economical. Wasted power is a Bad Thing, so low
> > > swing is a must. Padrings are some of the most expensive real
> > > estate around, so pincount should be minimized. Line termination
> > > can dominate a PWB so KISS is the rule. Power supplies (esp.
> > > ones that can both sink and source current) are expensive and
> > > nasty to deal with, so do without (both for termination and
> > > funny analog functions in the I/O circuits.)
> > >
> > > What can we add to the list? Remove? Priorities? (This is
> > > engineering, we make tradeoffs.) Where does this take us?
> > >
> > > --
> > > D. C. Sessions
> > > dc.sessions@vlsi.com
> > >
> > > **** To unsubscribe from si-list: send e-mail to majordomo@silab.eng.sun.com. In the BODY of message put: UNSUBSCRIBE si-list, for more help, put HELP.
> > > si-list archives are accessible at http://www.qsl.net/wb6tpu/si-list
> > > ****
> >
> > --
> > D. C. Sessions
> > dc.sessions@vlsi.com
> >
> > **** To unsubscribe from si-list: send e-mail to majordomo@silab.eng.sun.com. In the BODY of message put: UNSUBSCRIBE si-list, for more help, put HELP.
> > si-list archives are accessible at http://www.qsl.net/wb6tpu/si-list
> > ****
>
> **** To unsubscribe from si-list: send e-mail to majordomo@silab.eng.sun.com. In the BODY of message put: UNSUBSCRIBE si-list, for more help, put HELP.
> si-list archives are accessible at http://www.qsl.net/wb6tpu/si-list
> ****

-- 
D. C. Sessions
dc.sessions@vlsi.com

**** To unsubscribe from si-list: send e-mail to majordomo@silab.eng.sun.com. In the BODY of message put: UNSUBSCRIBE si-list, for more help, put HELP. si-list archives are accessible at http://www.qsl.net/wb6tpu/si-list ****


About this list Date view Thread view Subject view Author view

This archive was generated by hypermail 2b29 : Tue Feb 29 2000 - 11:39:12 PST