Given the compression capabilities of a CDW, it is important to explore its usability for our target goal, namely to execute the atomistic BTI/RTN simulation over strides ofπππ workload. We will be altering a version of the atomistic BTI model [265, 120] and will reformulate as follows: Starting from the premise of first order kinetics, the capture probability of a gate stack defect follows the ODE of Equation 3.7. The general solution returnsππ, afterπ πππ
pulses of specific π andπΌhave been applied (equivalent duration Ξπ‘=π/π).
Parametersπandπare functions of π andπΌ[265, 120].
πππ
ππ‘ =πππ+π β ππ= (οΈ
ππ0+π π
)οΈ
πππβ π
π (3.7)
CDW VERIFICATION FOR BTI MODELING 43
0.2 0 0.6 0.4
1 0.8
0 2
4 6
8 10
β0.02 0 0.02 0.04 0.06
Ρα (p.u.) log10 {Ρf (Hz)}
Β΅{es(t)} (V)
(a)
0.2 0 0.6 0.4
1 0.8
0 2
4 6
8 10
0 0.2 0.4 0.6 0.8
Ξ΅ Ξ± (p.u.) log10 {Ξ΅
f (Hz)}
Ο{es(t)} (V)
(b)
0.2 0 0.6 0.4
1 0.8
0 2
4 6
8 10
0 100 200 300
Ξ΅ Ξ± (p.u.) log10 {Ξ΅
f (Hz)}
Number of CDW Points (p.u.)
(c)
Figure 3.4: Accuracy assessment (signal integrity) of thewavapproxtool that produces CDW representations, using a 1Β΅s bit stream. It should be noted that SPICE requires 17089 points to represent that signal and a VCD representation uses 521 points [234].
Each iteration of the general solution of Equation 3.7 is used as initial condition for the next model evaluation. The wavapprox tool, presented previously, identifies regions of similarπ andπΌin theπππ (i.e. a CDW representation is produced). Then, at each (π, πΌ,Ξπ‘) point, a single iteration of Equation 3.7 is performed. A CDW representation reduces aggressively the number of model iterations required for the calculation ofππ. That way, time-dependent BTI simulation becomes computationally feasible over long time spans (see Figure 3.2). Given the compression capabilities of the CDW representation (see Figure 3.4c), the number of model iterations is aggressively reduced, thus
decreasing the CPU time required for simulation.
In Section 3.3, we have assessed the accuracy of thewavapproxfrom a signal integrity point of view. Given the proposed model of Equation 3.7, we will evaluate the accuracy of a CDW approximation in terms ofππ evaluation. For our experiments we use 100 random bit streams with an average frequency (πππ£π) of 1 GHz and a duration of 1Β΅s. Given each signal, we define an error metric for BTI evaluation, according to Equation 3.8, whereππ,SPICE is the capture probability for a defect at the end of the signal based on SPICE evaluation (i.e. using Equation 3.1). ππ,CDWis a similar capture probability, using a CDW file (for a specific pair ofππ andππΌ) and the model of Equation 3.7. We use 5 different cases, ranging from βfastβ (πππ» = 10β11s) to βslowβ (πππ» = 10β3s) defects. That way, we can calculate the mean, standard deviation and maximum ofππacross the 100 signals for various values of (ππ, ππΌ) and for the five different defects (Figures 3.5a β 3.5f).
ππ=ππ,SPICEβππ,CDW (3.8)
A direct correlation exists between how βfastβ a defect is and the average frequency of theπππ bit stream (πππ£π). More specifically, the defects that are significantly βslowerβ in comparison to the average frequency of the bit stream are accurately modeled, regardless of the compression imposed bywavapprox (Figures 3.5a through 3.5d). Faster defects (Figures 3.5e and 3.5f) have time constants comparable to the bit streamβs period (1/πππ£π). The proposed model is unable to follow theππ of these defects, irrespective ofππ andππΌ values.
The relation betweenπππ£π and the defect time constants has been identified earlier [120], however no comprehensive study has been performed for the accuracy limitations of the AC BTI/RTN model [265, 120]. Given that a version of this model, enhanced with workload memory, is the kernel of our simulation methodology, we should exhaustively explore its accuracy limitations for each model iteration. More specifically, it is important to identifyfor which defects and under which conditions is the accuracy of our model acceptable.
CDW VERIFICATION FOR BTI MODELING 45
0
5
10 0
0.5 5.28521 5.2854 5.2856 5.2858
x 10β6
log{Ξ΅f (Hz)}
ΟcH=1.0eβ003, ΟcL=1.0e+007, ΟeH=1.0eβ001, ΟeL=1.0eβ002
Ρα (p.u.) ¡{ePc} (p.u.)
(a)
0
5
10 0
0.5 1.27761 1.2776 1.2776
x 10β4
log{Ξ΅ f (Hz)}
ΟcH=1.0eβ003, Ο
cL=1.0e+007, Ο
eH=1.0eβ001, Ο eL=1.0eβ002
Ξ΅ Ξ± (p.u.) max{ePc} (p.u.)
(b)
0
5
10 0
0.5 1 1 1.5 2 2.5 3
x 10β3
log{Ξ΅f (Hz)}
ΟcH=1.0eβ007, Ο
cL=1.0e+007, Ο
eH=1.0eβ005, Ο eL=1.0eβ006
Ρα (p.u.) ¡{ePc} (p.u.)
(c)
0
5
10 0
0.5 1 0.03 0.04 0.05 0.06
log{Ξ΅ f (Hz)}
Ο
cH=1.0eβ007, Ο
cL=1.0e+007, Ο
eH=1.0eβ005, Ο eL=1.0eβ006
Ξ΅ Ξ± (p.u.) max{ePc} (p.u.)
(d)
0
5
10 0
0.5 0.41 0.5 0.6 0.7 0.8
log{Ξ΅ f (Hz)}
ΟcH=1.0eβ011, ΟcL=1.0e+007, ΟeH=1.0eβ009, ΟeL=1.0eβ010
Ρα (p.u.) ¡{ePc} (p.u.)
(e)
0
5
10 0
0.5 0.99011 0.9901 0.9901 0.9901 0.9901
log{Ξ΅ f (Hz)}
Ο
cH=1.0eβ011, Ο
cL=1.0e+007, Ο
eH=1.0eβ009, Ο eL=1.0eβ010
Ξ΅ Ξ± (p.u.) max{ePc} (p.u.)
(f)
Figure 3.5: Accuracy assessment of the atomistic BTI model of Equation 3.7 for various degrees of CDW-compression [234].
In order to formally conclude on the accuracy limitations of our approach, we first create a set of 400 different defects, with time constants (πππ», πππΏ, πππ», πππΏ) uniformly distributed in the logarithmic scale. The defects are arranged in a two dimensional plane according to min{πππ», πππΏ} and min{πππ», πππΏ}.
We use three sets of bit streams with πππ£π equal to 1 GHz, 100 MHz and 10 MHz for each set. Each set contains 100 signals, each one with a duration of 1 Β΅s. Each signal is CDW-approximated with maximum compression (ππ β+β
andππΌβ1) andππ is calculated according to Equation 3.8. That way, we can calculate π{ππ}, π{ππ} and max{ππ} for each defect and for each πππ£π case.
The results are shown in Figures 3.6a through 3.6f. It is important to note that, these results represent the accuracy of the proposed modelfor a single iteration of Equation 3.7. However they provide useful insight on the accuracy limitations of the proposed model, when using it for an arbitrary number of iterations.
For allπππ£πvalues, we can always find a subset of defects that we are unable to model accurately. As πππ£π is increasing, this subset is reduced and an acceptable accuracy of our model spreads to more defect time scales. For the case ofπππ£π= 1 GHz, our model can accurately simulate all defects that satisfy min{πππ», πππΏ} β₯10β7 sand min{πππ», πππΏ} β₯10β7s.
CDW VERIFICATION FOR BTI MODELING 47
β10
β5 5 0
10
β10 0
10 0
0.2 0.4 0.6 0.8 1
min{logΟe (s)}
favg=1GHz
min{logΟ c (s)}
Β΅{ep} (p.u.)
(a)
β10
β5 5 0
10
β10 0
10 0
0.2 0.4 0.6 0.8 1
min{logΟ e (s)}
favg=1GHz
min{logΟ c (s)}
max{ep} (p.u.)
(b)
β10
β5 0 5 10
β10 0
10 0
0.2 0.4 0.6 0.8 1
min{logΟe (s)}
favg=100MHz
min{logΟc (s)}
Β΅{ep} (p.u.)
(c)
β10
β5 0 5 10
β10 0
10 0
0.2 0.4 0.6 0.8 1
min{logΟ e (s)}
favg=100MHz
min{logΟ c (s)}
max{ep} (p.u.)
(d)
β10 0 β5
5 10
β10 0
10 0
0.2 0.4 0.6 0.8 1
min{logΟ e (s)}
favg=100MHz
min{logΟc (s)}
Β΅{ep} (p.u.)
(e)
β10 0 β5
5 10
β10 0
10 0
0.2 0.4 0.6 0.8 1
min{logΟ e (s)}
favg=10MHz
min{logΟ c (s)}
max{ep} (p.u.)
(f)
Figure 3.6: Errors from a single iteration of Equation 3.7 on bit streams of different πππ£π. The error of the proposed model is restricted to small defect subsets, asπππ£πis increased [234].
β15 β10 β5 0 5 10 15 0.8
1 1.2 1.4 1.6
β20
β10 0 10 20
Ο0 (s) Vdd (V)
Ο (s)
(a) Scaling defect time constants for various bias (πππ) conditions, according to Equation 3.9
β15 β10 β5 0 5 10 15
0 50 100 150 200
β20
β10 0 10 20
Ο0 (s) T (oC)
Ο (s)
(b) Scaling defect time constants for various temperature (π) conditions, according to Equation 3.10
Figure 3.7: Equations 3.9 and 3.10 enable full support of bias and temperature variations in the CDW-based BTI simulation framework [234].