|
# 7 - Bottom up calibration for multilevel truthfulness and
secured designs |
|
Three calibrations are at stake:
- with measurements on silicon for transistor models
and parameters
- with extractions from layout for electrical netlist
back-annotations
- with simulation results from any level for a higher
level model
TEST CASE: Read Margin of a memory
| First calibration: HORIZONTAL
from silicon to transistor models |
 |
IDS Current Drain
Source
VDS Voltage Drain Source
VGS Voltage Gate Source |
BSIM4 model gives the equations which describe the behavior
of the transistors
Calibration provides the parameters for adjusting
those equations
| Second calibration: TOPOLOGICAL
from layout to netlist upwards |

Based on experience the designer extracts manually some
parameters from the layout such as subtract injection, coupling between
bit-lines… to introduce in the electrical netlist.
Then a next step should be to perform the same operation with silicon
measurements:
back to step 1.
Once the Spice models and the electrical netlist are
calibrated vs layout and silicon, the simulation results can be representative
of the physical reality.
Analyses, such as Monte-Carlo, can be performed in SMASH with the guarantee
of the results on silicon.
Moreover, using Monte Carlo with this calibration technique, yield of
the circuit can be predicted with a higher accuracy than with a single
test chip!

| Third calibration: MODELING upwards, from
process descriptions on to electrical -> structural -> functional
-> behavioral -> system |

The electrical simulation is used to calibrate a behavioral
simulation.
This behavioral simulation will be faster and will allow SoC level simulation.

Beyond calibration
While multi-level zooming requires calibration for a memory “Read
margin” or for ensuring that complex chains of transistor models
truthfully represent silicon, some multi-level issues raise problems
beyond zooming:
- Laplace or s-transforms are powerful structural representations
for continuous-time or analog implementations of filters and other
devices with a linear functionality: but they can be embedded within
the electrical models in SPICE, as well as within the behavioral models
in VHDL,
- Their discrete-time counterparts, the celebrated
z-transforms serve for digital filters: but they can be implemented
at the functional level, synthesizable into fixed logic at the gate
level, as well as programmed in C to be executed by a DSP (Digital
Signal Processor).Worse then: probably, because these circuits process
analog inputs, the z-transform absurdly is only accessible as part
of VERILOG-A or VHDL-AMS.This is in complete contradiction with the
VSIA objective of Black-Box modeling, whereby even mixed signal circuits
processing mixed-signal inputs should be modeled in pure VERILOG!
The ultimate calibration issue.
- Worse then, the celebrated Schmitt Trigger: look for its
model in a SPICE library of devices!It is the paradigm of a
mixed signal device thought by many to be a logic component,
but dealing with analog inputs. As a matter of fact, it acts
like a sense amplifier with the ReadMargin issues!
|
 |
< SMASH Differentiators
|
|