Re: [10GBASE-CX4] Working Paper Available on IEEE site
Hi, Kamal,
I support the general model of TX emphasis, and designing the
spec so that adaptive RX eq is *not* required. (It can be a value-add
for any vendor who thinks it's worth the cost.) So I think we're
basically in agreement about how implementation should shake out.
I still don't think the XAUI experience is a good test of the
wild & wooly interop world we face in CX4 because so often the
backplane and line cards are co-engineered. Architecturally,
fixed TX and vendor-specific RX is more robust in the long run,
when the channel is going to be quite variable.
Given that we both want a system based on TX pre-emphasis,
a well-specified channel, and passive RX eq, the only difference
is whether the signal is standardized at the TX end or the RX
end of the link. (I don't think we can standardize both places
because the channel varies.) With the question rephrased that
way, what do you see as the relative benefits of TX vs RX signal
standardization?
Side note: If we standardize at one end, we can always include
an *informative* section about behavior at the other end. And
the compliance part will probably specify testing at certain
limit conditions, which is almost like normative text.
Cheers,
Chuck
kdalmia@marvell.com wrote:
>
> Chuck,
>
> XAUI was primarily targeted at backplanes. Cable environment is not much
> different from the backplane environment - backplane trace lengths can and
> do vary a lot. Most intuitive approach is to solve the length variation
> problem by "adaptive RX equalization". However, it has been well proven in
> backplane environment that adaptive RX equ is not the most optimum
> technique from a practical implementation point of view - specifically,
> power and performance. Instead, it is better to "over de-emphasize" - it
> works well for short and long links (cable or traces).
>
> Regards
> Kamal
>
> ===============================
> KAMAL DALMIA
> Technical Marketing Manager
> Marvell Semiconductor
> 700 First Av.
> Sunnyvale, CA 94089
> Phone: (408) 222 8979
> ===============================
>
>
> Chuck Harrison
> <cfharr@erols.com> To: kdalmia@marvell.com
> Sent by: cc: stds-802-3-10GBCX4@ieee.org
> owner-stds-802-3-10gbcx4@majordo Subject: Re: [10GBASE-CX4] Working Paper Available on IEEE site
> mo.ieee.org
>
>
> 01/02/2003 10:15 AM
>
>
>
> Kamal,
>
> kdalmia@marvell.com wrote:
> >
> [...]
> > Instead, I believe, the RX and the Channel should be specified (with TX
> > implied). Current XAUI specs have RX sensitivity and jitter tolerance
> > numbers. This implies that the burden is on the TX to deliver a "clean"
> eye
> > at the RX.
>
> The big difference between CX4 and XAUI is that we now need to
> tolerate a dramatically larger range of channel-loss behaviors.
> An architecture which uses fixed TX preemphasis and a single-
> parameter cable model gives RX designers a well-defined problem
> to solve. It allows (but does not require) RX eq based on
> estimated cable length. If you invert this -- fixed RX behavior
> and vendors trying to put their smarts into transmit EQ -- you
> can never get as good a system result because the TX side has
> no way of knowing about the channel characteristic. And if
> both TX and RX sides (not to mention the cable vendors!) are
> blindly trying to be "smart" then interoperability is at
> serious risk.
>
> In the variable-channel CX4 environment, I would *strongly*
> lean towards Howard's proposal -- fixed TX step response, and
> a narrow tolerance on the channel's deviation from the nominal
> cable model.
>
> Cheers,
> Chuck Harrison
> Far Field Associates, LLC