HICUM evaluation

1
HICUM evaluation

• Cedric Pujol Analog and Mixed Signal Flows
• Jean Remy Analog and Mixed Signal Flows

2
Aim

• The aim of these tests is to verify the HICUM
  implementation on a  real life  design, and
  all the simulators supported by ST.

3
EDA Tools

• The simulators used are the following
• Ams 2004.1 and 2004.2 beta4
• ADS 2003.C
• Aplac 7.92c
• Spectre 5.0.33.031104
• GoldenGate 3.3.14
• Hspice 2004.2-beta2
• Hsim 5.0_2004.20.7

4
Outline

• Evaluation of the implementation of the HICUM
  model in the simulators RF simulations
• Transient simulation experiences
• Comparison between STBJT and HICUM
• Conclusion

5
Outline

• Evaluation of the implementation of the HICUM
  model in the simulators RF simulations
• Mixer
• LNA
• Transient simulation experiences.
• Comparison between STBJT and HICUM
• Conclusion

6
First design under test

• The design is a four bipolars based up-mixer done
  by CRD on 60GHz Bicmos ST process.
• The simulations peformed are
• One 2 tone Harmonic Balance for a low gain
• A swept Harmonic Balance for 13 gain points
• The conditions chosen are 15 harmonics for the LO
  and 8 for the IF. The temperature is set to 25
  C.
• The outputs are the fundamental tone at 1.941GHz
  and the two 3d order intermodulation products at
  1.937 GHz and 1.943 GHz.

7
Non-linear results (fundamental)
STBJT Results (dBm) STBJT Sim time (s) HICUM Results (dBm) HICUM Sim time HICUM SH Results (dBm) HICUM SH Sim time
ADS -6.79 1x 9 -6.80 1.5x -6.82 1.8x
Aplac -6.76 1x 85 -6.75 5.5x -6.78 18.6x
Eldo -6.80 1x 49 -6.80 1.7x -6.82 2.2x
GoldenGate -6.62 1x 22 -6.62 2.4x -6.64 2.4x
Spectre -6.77 1x 34 -6.76 1.6x -6.78 1.8x
8
Non-linear results (3d order)
STBJT Results (dB) STBJT Sim time (s) HICUM Results (dB) HICUM Sim time HICUM SH Results (dB) HICUM SH Sim time
ADS 124.4/75.5 1x 9 107.5/89.3 1.5x 107.0/87.9 1.8x
Aplac 120.7/75.6 1x 85 105.9/68.9 5.5x 105.5/68.6 18.6x
Eldo 119.8/73.8 1x 49 107.0/69.7 1.7x 106.1/69.5 2.2x
GoldenGate 108.8/75.3 1x 22 86.4/74.4 2.4x 86.6/73.8 3.0x
Spectre NS/NS 1x 365 NS/NS NS NS/NS NS
9
Non-linear results (sweep)
STBJT Sim time (s) HICUM Sim time HICUM SH Ratio / STBJT (Ratio / HICUM)
ADS 1x 38 2.7x 3.3x (1.2x)
Aplac 1x 523 85.5x ?
Eldo 1x 524 1.7x 2.6x (1.6x)
GoldenGate 1x 274 7.5x 8.4x (1.1x)
Spectre 1x 381 1.7x 2.0x (1.2x)
10
Outline

• Evaluation of the implementation of the HICUM
  model in the simulators
• Mixer
• LNA
• Comparison between STBJT and HICUM
• Conclusion

11
Second design under test

• The design is a six bipolars based Low Noise
  Amplifier done by CRD on 60GHz Bicmos ST process
• The simulations peformed are
• A DC analysis
• One 2 tone Harmonic Balance for the IP3
  computation (spacing between the tones 4MHz)
• A swept Harmonic Balance for the IP1 computation
• A noise analysis
• The conditions chosen are 10 harmonics for the RF
  tone. The temperature is set to 27 C.

12
DC results
STBJT Results (mA) STBJT Sim time (s) HICUM Results (mA) HICUM Sim time HICUM SH Results (mA) HICUM SH Sim time
ADS 4.68 1x 9 4.68 1.1x 4.68 0.9x
Aplac 4.68 1x 7.9 4.68 1.2x 4.68 2.6x
Eldo 4.68 1x 3.3 4.68 1.4x 4.68 1.6x
GoldenGate 4.68 1x 7.8 4.66 1.4x 4.66 2.1x
Spectre 4.68 1x 6.8 4.68 1.0x 4.68 29.4x
13
Noise results
STBJT Results (dB) STBJT Sim time (s) HICUM Results (dB) HICUM Sim time HICUM SH Results (dB) HICUM SH Sim time
ADS 1.1 1x 11 1.2 1.2x 1.1 0.9x
Aplac TBD 1x ? TBD TBD TBD TBD
Eldo 1.07 1x 10 1.07 1.0x 1.07 3.1x
GoldenGate 1.07 1x 12 1.07 1.0x 1.07 1.4x
Spectre 1.1 1x 6 1.2 1.1x 1.1 29.7x
14
IP1 results
STBJT Results (dBm) STBJT Sim time (s) HICUM Results (dBm) HICUM Sim time HICUM SH Results (dBm) HICUM SH Sim time
ADS -16.1 1x 133 -16.1 0.8x -16.0 0.9x
Aplac -15.9 1x 224 No CV NS No CV NS
Eldo -16.2 1x 91 -16.2 0.9x -16.3 1.2x
GoldenGate -16.0 1x 139 -16.0 2.8x -16.0 3.5x
Spectre -17.1 1x 1264 No CV NS No CV NS
15
OIP3 results
STBJT Results (dBm) STBJT Sim time (s) HICUM Results (dBm) HICUM Sim time HICUM SH Results (dBm) HICUM SH Sim time
ADS 14.4 1x 14 14.3 1.0x 14.3 1.1x
Aplac 16.2 1x 1220 No CV NS No CV NS
Eldo 16.1 1x 140 15.9 1.1x 15.9 1.5x
GoldenGate 14.4 1x 33 13.6 1.1x 13.6 1.4x
Spectre No CV 1x ? No CV NS No CV NS
16
ADS results

• The simulation times given for ADS are stopwatch
  times and for others CPU times. So to compare
  thoroughly, ADS simulation times should be
  decreased. Moreover, they are so low that they
  are really impacted by the netlisting process.
• The results seem correct and the simulation times
  are reasonable.
• The self-heating implementation causes an
  estimated 20 simulation time overhead which
  seems quite reasonable.
• The 3d order non-linearities (NL) are quite
  different in ADS from other simulators but seem
  more physical since the slopes for the 3d order
  NL are theoratically correct (3dB/dB).

17
Aplac results

• A bug has been submitted to Aplac Solutions in
  the 7.91c release. It has not been corrected in
  the new release. Therefore, We used a work-around
  that the support gave us for the 7.91c release.
• The results seem correct but the simulation times
  skyrocket (over 12 hours for the mixer sweep).
• The self-heating implementation makes the
  simulation time multiplied by around 3 which is
  much too long.

18
Eldo results

• The results seem correct, and the simulation
  times are correct.
• The results match all the simulators but ADS but
  they seem not physical (slopes are 2dB/dB for a
  3d order NL).
• The self-heating implementation multiplies the
  simulation time by 1.5 to 3x.

19
GoldenGate results

• The results seem most of the time correct but the
  simulation time overhead for HICUM is much too
  high. The mixer testcase will be sent to the
  support.
• The results are quite pessimistic compared with
  other simulators without any convincing
  explanation.
• The self-heating implementation is activated by
  adding a flag in the Spectre model card that is
  not necessary in Spectre. The time overhead is
  pretty high (around 40 for the LNA)

20
Spectre results

• 2 analyses were performed for non-linearities
  PSSPAC and QPSS. The first one is fast but only
  returns the fundamental. The other is longer and
  the non-linearities given are not satisfying for
  the mixer.
• Therefore, these tests is not very relevant for
  HICUM accuracy but the implementation seems
  correct.
• The self-heating implementation causes an
  estimated 20 simulation time for the mixer
  overhead which seems quite reasonable.
• For the LNA, the DC simulation times are
  multiplied by 30 when the self heating is
  activated !

21
Outline

• Evaluation of the implementation of the HICUM
  model in the simulators RF simulations
• Transient simulation experiences.
• Comparison between STBJT and HICUM
• Conclusion

22
Third design under test (1)

• Simulation of an ECL 4GHz frequency divider
• 60GHz Bicmos process, 32 bipolar devices, no self
  heating
• HICUM model introduced as maturity design kit
  update, close to tape-out date
• Many iterations were needed with EDA vendors to
  reach transient simulation convergence

23
Third design under test (2)

• After non convergence corrections, we detected a
  design problem that STBJT could not show-up (too
  high bias and saturation).
• Two versions of the design are simulated
• With the design problem
• With its correction.

24
Transient simulation times
STBJT Pb HICUM Pb HICUM corr. STBJT M2 corr. Comment
ADS 22s1x 3.9x 1.8x No model Approxlevel1
Aplac 73s 3 hours - - Wrong results
Eldo 12s1x 8.3x 2.3x 1.3x Amsver flag
GoldenGate 200s1x 4x 2x Not significant Harm. Bal. Meth.
Hspice 5s1x 15.4x - - No Pb. detected
Spectre 8s1x 54.5x 2.5x 0.7x Updatelevel flag
25
Transient simulation results

• Hsim (semi-electrical) does not detect the design
  problem
• For nanosim, wait Q2 2005 !

26
Transient simulation conclusions (1)

• Incomplete partial derivatives highly impact
  transient (and assimilated) simulation time, and
  cause simulation crashes
• Many iterations with EDA vendors for partial
  derivatives implementation
• A lot of energy spent to convince designers of
  HICUM interest after first bad contact.
• Partial derivatives corrections modify AC
  behavior with bias gt fTmax

27
Transient simulation conclusion (2)

• Need to re-qualify all AC validation simulations
• Porting of excess phase and non quasi static
  effects was difficult on some simulators
• Transient simulation still not completely
  reliable on every simulator, with random effects

28
Outline

• Evaluation of the implementation of the HICUM
  model in the simulators RF simulations
• Transient simulation experiences.
• Comparison between STBJT and HICUM
• Conclusion

29
Aim

• The aim of these tests is to compare the STBJT M2
  (DK 2.1) with the HICUM (DK 3.0) model.
• The simulation (1 dB compression point) has been
  performed on ADS and Spectre on all the
  simulators.

30
Simulation times (LNA)
STBJT M2 Sim time (s) HICUM Sim time
ADS 1x 108 1.1x
Aplac 1x 373 ?
Eldo 1x 101 2.5x
GoldenGate TBD TBD
Spectre 1x 1541 1.0x
31
Simulation times (Mixer)
STBJT M2 Sim time (s) HICUM Sim time
ADS 1x 43 2.4x
Aplac 1x 696 64.2x
Eldo 1x 803 1.1x
GoldenGate 1x 381 5.4x
Spectre 1x 444 1.5x
32
Mixer results
HICUM is in blue and STBJT in red. The curves
are very different. But HICUM curves seem more
physical since the slopes are 3dB/dB for the
third harmonics. The compression of the
fundamental arrives ealier too in HICUM.
33
Outline

• Evaluation of the implementation of the HICUM
  model in the simulators
• Comparison between STBJT and HICUM
• Conclusion

34
Results sum up (HICUM no SH)

• The small signal implementation is fast and
  accurate in every simulator, at least before
  transient simulation corrections.
• The non linear simulations exhibit problems in
  terms of simulation speed and convergence.
• The differences computed by ADS between the STBJT
  and HICUM are quite large on one of the third
  harmonics (35 dB) but other simulators disagree.
• STBJT was always optimistic compared with HICUM
  which could be annoying.

35
Results sum up (HICUM with SH)

• The self-heating implementation depends a lot on
  the simulator in terms of speed.
• Moreover we met some problems on the LNA (DC with
  Spectre) that we dont have on the mixer.
• Therefore, it should be wise not to activate by
  default the self-heating.

36
Actions and recommendations (1)

• We have to push on some EDA vendors to improve
  the convergence for both standard and
  self-heating implementations.
• DC for SPECTRE
• Non convergence for APLAC
• Simulation speed for GOLDENGATE
• Self heating speed improvement for ELDO and ADS
• We should warn the designers that the overhead on
  simulation time could be from 2x to 6x when
  switching from STBJT to the HICUM models for non
  linear analyses.

37
Actions and recommendations (2)

• The ADS curves showing a 3dB/dB physical slope
  are in contradiction with what is predicted by
  all its competitors. It has to be explained to
  see which simulator is right.
• Try other testcases to refine this evaluation

38
Conclusion

• We are still waiting a stabilized implementation
  of self-heating in all supported simulators
  before releasing this new feature
• First designer contact was very negative
• A lot of energy/credibility lost in partial
  derivatives implementation
• Multiplication of DEVICE compatibility flags, and
  absence of a reference source code
• No decision to drop-out of STBJT model