timing on GTL+ bus

To All:

Be aware that when simulating a GTL+ bus using some processor
models, you must adjust the receiver input thresholds when the
driver has different slow and fast models and assumes different
termination voltages in them.

Some new processors have internal pull-up resistors for the GTL bus,
which reduces parts significantly, and result in extremely
attractive signal quality when the processor receives. The
resistors are connected to separate termination voltage (Vtt) supply
pins.

When simulating slow corner for timing, one uses the low termination
voltage and high internal resistor values. The fast corner uses
high Vtt and low resistance. For the components with internal pull-
ups, this can be done with a model whose [Voltage Range] is set to
the termination voltage desired, and the V-I curves take care of the
resistor values.

Remember that Vinl and Vinh for the GTL bus are based on some
reference voltage, Vref, which is typically developed via a
resistive divider between Vtt and ground. For a widely used GTL
variant called "GTL+", Vtt = +1.5V and Vref is nominally 2/3*Vtt =
+1.00V. Often, Vinl and Vinh are specified as offsets from Vref,
such as:

        Vinl(max) = Vref - 200mV
        Vinh(min) = Vref + 200mV

In the nominal case listed above, then, Vinl would be +0.800V, and
Vinh would be +1.200V. Indeed, a popular GTL+ bridge component has
these values listed in its IBIS model.

However, if Vtt is at its low margin level, for example +1.35V, then
Vref will be +0.900V, and the receiver thresholds should "reflect"
this. Using the example, Vinl(max) = +0.700V, Vinh(min) = +1.100V.

In the case where a slow corner model includes the low Vtt value,
the output voltage step may not be able to incident wave switch
on the low to high transition, if one uses the Vinh value derived
from the nominal Vtt as the timing point on the receiver waveform.

In a specific instance, the existing model for a GTL+ bridge
component, with Vinl=+0.8V, Vinh=+1.2V, showed timing violations on
the slow corner. The slow corner drivers on the processor showed a
shelf just below +1.2V, when the PCB trace impedance was at its low
value. The voltage didn't cross the high threshold until after the
slow RC charging from the step up to Vtt, which violated timing.

Because Vtt and Vref were common to the processor and bridge, it is
reasonable to assume that the switching threshold at the bridge
would track the Vtt. Setting Vinh to +1.1V showed that the timing
was acceptable. (One just might think that they designed the
processor output buffers just big enough to handle the low trace
impedance case.)

This discussion extends to the fast corner, where the nominal
thresholds would give an overly conservative (smaller) hold time
estimate, because the thresholds would be lower than actual.

The unfortuate part of this is that you now need "slow" and "fast"
receiver models, as there is no "typ/min/max" for Vinl and Vinh.

The other point to keep in mind is that you need to be very careful
if you have a system design with separate Vtt supplies to the loads
on a GTL bus, or if you have separate Vref's to those loads.

        Note: for margin, one could take into account the
        tolerance of the divider resistors, and adust Vinl
        and Vinh up or down appropriately.

Regards,

-wis

-- 
Wis Macomson
SiQual
c/o Wis_Macomson@mentor.com