Title: Analytical Modeling of SRAM Dynamic Stability Bin Zhang, Ari Arapostathis, Sani Nassif and Michael O
1Analytical Modeling of SRAM Dynamic
StabilityBin Zhang, Ari Arapostathis, Sani
Nassif and Michael Orshansky University of
Texas at Austin IBM Austin Research Lab
2Outline
- Motivation
- System modeling setup
- Dynamic state space analysis
- Criterion of dynamic stability of SRAM
- Transient Behavior of SRAM under Noise
- Dynamic noise margin of SRAM
- Experimental Results
3Motivation
- Verification of SRAM stability is an essential
design task - Smaller design windows due to power supply
scaling - Identify the maximum possible unintended
violations - Static vs Dynamic
- Growing number and amplitude of dynamic noise
sources - Single Event Upset (SEU)
- Cross-coupling Noise
- Power supply noise
- Use non-linear system theory to analyze the
stability of cross-coupled inverter system of an
SRAM in the presence of transient noise -
4Prior Work
- Static Noise Margin (SNM) analysis posits a
constant noise signal at circuit nodes Seevinck
et al 1987 Agarwal et al 2006 - Mirror-and-Maximum-Square Approach
- Neglects time-dependent aspect of noise, e.g.
pulse duration. Gives pessimistic noise tolerance
for short noise pulse
Static Noise Margin
5Prior Work
- Simulation-based techniques can capture the
transient behaviors of SRAM and noise, but
provide little insight into problem and requires
iterative search to find critical noise pulses - Device level Fu et al 1984
- Transistor level Hazucha et al 2000
- Mixed mode Mayaram et al 1993
- Breaking the feedback loop analysis ignores the
strong feedback effect in SRAM Shepard 98 - No analytical approach exists for dynamic noise
margin of SRAM
6Inverter-Level SRAM Cell Modeling
Standby mode
in(t), Noise Current
in(t)
In
V1(t)
V2(t)
pw
0
t
C
C
V2(t0)0
State Vector V(t)(V1(t), V2(t))
7State Space Trajectory
- Case 1 Short noise pulse does not result in a
state flip
Noise Duration
Recovery
Noise Duration
Recovery
Noise Duration
Recovery
8State Space Trajectory
Case 2 Long pulse results in state flip
Noise Duration
Failed Recovery
Failed Recovery
Noise Duration
Noise Duration
FailedRecovery
9SRAM Cell Coupled Nonlinear System with Feedback
Solvable?
driving current of the inverter
10Linear Gate I-V Modeling
- To enable analytical solutions, need to use a
linear and separable inverter I-V model - Ignore short-circuit current
- Piece-wise linear model (Horowitz84)
- Assumptions justified by accurate waveforms
(transient behavior!) predicted for logic gates
(Cutoff) (Saturation) (Linear)
11Dynamic State Space Analysis
Time-dependent
Equilibrium states evolve with time
12Dynamic State Space Analysis
1.6
V2 (V)
0.8
0.0
0.0
0.8
1.6
V1 (V)
13Dynamic State Space Analysis
1.6
V2 (V)
0.8
0.0
0.0
0.8
1.6
V1 (V)
14Dynamic State Space AnalysisRecovery Stage
- Equilibrium states have regions of attraction
- Any vector in attraction regions is pulled to the
equilibrium state - If noise pushes state vector beyond the boundary
flip occurs! - Boundary of attraction regions for symmetric
system
15SRAM State Evolution under Noise
- The operation of SRAM can be approximately
classified into two modes before the attraction
boundary is reached - Weak coupling V2ltVdsat
- Strong feedback V2gtVdsat
Strong Feedback
V1
Weak Coupling
V2
16SRAM Modeling Weak Coupling Mode
Linearized equations
17SRAM Modeling Strong Feedback Mode
Linearized cross-coupled equations
18Solving Transient SRAM Equations Critical Pulse
Width
- Critical pulse width, Tcrit, the time when the
state trajectory crosses the attraction boundary.
- Tcrit can be solved from the linearized
equations
Asymptotic behavior, for large noise amplitude,
a simple bound
19Dynamic Noise Margins Experimental Results
- Amplitude (In) vs pulse width (Tcrit) of the
critical pulse - Safe region is below the curves
- Analytical result gives larger and more realistic
DNM than the bounds
20Critical Pulse Width Dependence on Supply Voltage
- Model allows stability-driven cell design
exploration - Dynamic noise margins increase for higher Vdd
- For small noise magnitudes, Tcrit is very
sensitive to Vdd
21Critical Pulse Width Dependence on Cell Area
(Size)
- DNM can be increased by sizing up devices
- But transistor sizing becomes less effective for
large noise amplitude
22Conclusions
- We present an analytical modeling of SRAM dynamic
stability - Model the transient current noise as a
square-pulse - Apply linear gate modeling
- Obtain a closed-form solution of critical pulse
in terms of the noise amplitude. - Excellent match is achieved compared to SPICE