Comments on Ultra SCSI

Steve Finch TFINCHS%A1.VAX2T.mrouter at vax2t.tus.ssi1.com
Fri Feb 11 09:28:04 PST 1994


Is Ultra SCSI for differential only?  I think it will have to be, as 
single ended has significant problems.  

For Single Ended operation, Ultra SCSI isn't just an easy divide by two 
and go as illustrated by some.  (See Jim McGrath's Email as an example.)  
To help explain, here's a reference diagram:


      0        __________________                       ______________
              /|                |\                     /
REQ*  1 _____/ |                | \___________________/
           ->| |<--T1         ->| |<--T2              |
               |                | |                   |
               |<------T3------>| |<-------T4-------->|


We all know that the period for the fastest "fast SCSI" is a minimum of 
100 ns and is made up of T1+T2+T3+T4.  We have specifications for these:

    T1    5 ns     minimum
    T2    5 ns     minimum
    T3   30 ns     minimum at transmitter
    T4   30 ns     minimum at transmitter
       ---------
         70 ns     of 100 ns budget.  

30 ns slop, fairly easy to do!  This is OK.


Now for Ultra SCSI, the numbers are:

    T1    5 ns     minimum
    T2    5 ns     minimum
    T3   15 ns     minimum at transmitter
    T4   15 ns     minimum at transmitter
       ---------
         40 ns     of 50 ns budget.  

Hmm... 20% slop!  Sounds like maybe we can do it, but close examination 
shows some major problems.

The primary one is slew rate control of the rise and fall times.  A 
design to guarantee a 5 ns slew in the best case results in 15 ns or 
more in the worst case.  So taking this into account, it gives you:

    T1   15 ns     maximum
    T2   15 ns     maximum
    T3   15 ns     minimum
    T4   15 ns     minimum
      ----------
         60 ns     of 50 ns budget!  OOPS!

Tightening down the slew rate control may be possible, but I doubt that 
anyone would be able to do better than: 5ns <= slew <= 10ns, and even 
this would be very expensive.  Even if we IC designers can control the 
maximum variance in slew rate to this range, the result would be:

    T1   10 ns     maximum
    T2   10 ns     maximum
    T3   15 ns     minimum
    T4   15 ns     minimum
      -----------
         50 ns     of 50 ns budget!  

This can't be done, as no one's perfect!

One solution is to scale the 5ns slew rate requirements down to 2.5 ns.  
We should be able to control the maximum slew to around 7.5ns.  But if 
we do this, what happens to the signal quality?  Wasn't this the reason 
for 5ns slew originally?

If we can't scale the slew rate, then room for the process variance must 
be found somewhere.  What this means is that the transmit assertion 
period and transmit negation period of 15 ns must be decreased to at 
least 10 ns, or preferably even smaller. Of course, this creates a whole 
new set of problems.  As an example:

    T1   10 ns     maximum
    T2   10 ns     maximum
    T3   10 ns     minimum
    T4   10 ns     minimum
      -----------
         40 ns     of 50 ns budget!  

We might be able to do this.

If Ultra SCSI is for differential only, are we need to investigate the 
whole picture like was done in SPI Annex C.  This shows skew variance of 
up to 19 ns.

    T1+T2   19 ns    Skew budget
    T3      15 ns    minimum
    T4      15 ns    minimum
         ----------
            49 ns    of 50 ns budget!  

The problems are still here to haunt us....

------------------------------------------------
Out of curiosity, is anyone out there coming to the March 3 Ultra SCSI 
meeting thinking this idea is a done deal?

COMMENTS?

===================================================================
Steve Finch, Silicon Systems, Inc.
Ph.  714 573-6808, Fx. 714 573-6916
email:  5723283 at mcimail.com
===================================================================






More information about the T10 mailing list