thinking...

Arlo,

  The question is complex. The SPICE Gummel-Poon model does not scale
well because the scaling assumptions are primitive. The scaling basically assumes
a fixed emitter length with no modeling of end effects. In addition, the model
has no capability for distinguishing between bipolar transistor structures, for
example, modeling the effect of double base contacts on base resistance. I think
that some of the more modern models do provide this capability, HiCUM(?). In
practice, I have seen both scalable and fixed geometry based bipolar models.
For scalable Gummel-Poon modeling, the model parameters are defined as
functions of geometry. This actually works very well and can be combined with
statsitical modeling, that is, scalable bipolar models with statistical dependency.
See the Spectre Monte Carlo modeling application  note for some discussion of
how to include geometric dependency in models. The only problem with this
approach is finding someone who is willing to do the modeling! On the other hand,
using fixed models is a simple and robust methodology that the fab tends to prefer.    
These comments apply to vertical devices. I don't think that scaling of lateral
devices is appropriate other than using multiple devices. Also, AC is correct about
modeling, it is better to use scale factor rather than scaling emitter area if you are
interested in model accuracy, however, you will pay a price in performance. For
the effect of parasitics tends scale down as the device scales up so 10x larger device
will have better speed-power performance than 10 1x devices.  

                                                                        Best Regards,

                                                                            Sheldon

Arlo,

  AC is saying that there are two options for scaling the transistor:

1) Scale the device by stretching the emitter width to create a larger device
or
2) Make many copies of the device and connect them in parallel

Use option 1 and the accuracy of the model will be degraded, use option
2 and the accuracy will be good but the parasitics will be larger.

Technically there is a third option, you could stretch the base and
collector and make multiple copies of the emitter. This is slightly
more accurate than 1 and has less parasitics than 2. See the
attachment for pictures.

                                                                 Best Regards,

                                                                     Sheldon