Title: Joint DesignTime and PostSilicon Minimization of Parametric Yield Loss using Adjustable Robust Optim
1Joint Design-Time and Post-Silicon Minimization
of Parametric Yield Loss using Adjustable Robust
Optimization
- Murari Mani, Ashish K Singh, and Michael
Orshansky - University of Texas at Austin
2Outline
- Post silicon tuning and motivation for joint
co-optimization with design-time tuning - Formal metrics of control and measurement
complexity to parameterize design and
optimization trade-offs - Efficient co-optimization using adjustable robust
optimization - Effectiveness of co-optimization for different
patterns of parameter variation and
implementation overhead
3Parametric Yield Loss and Mitigation by
Post-silicon Tuning
- Process variability results in parametric yield
loss - Exponential dependence of leakage on process
spread - Inverse correlation between leakage and
performance - Tuning chips after manufacturing is an effective
yield enhancing strategy - Adaptive body biasing (ABB) tightens Fmax spread
- Improves power limited parametric yield
Source Intel
4Joint Design Time and Post Silicon Co-Optimization
- Design time optimization - Cost is fixed
regardless of actual process conditions for each
chip - Lack ability to react to the actual conditions
on the chip - Both design-time optimization and post-silicon
adaptivity fight against the same variability
budget - Optimal application of two techniques depends on
- Spatial structure of variability and inter- vs.
intra-chip variability ratio - Spatial granularity of adaptivity
- Effectiveness of body biasing for a given
technology - Co-ordination between design time and post
silicon steps can produce a more optimal solution - Derive a strategy to guide post-silicon tuning,
and make the first-phase design decisions
5Contributions of this Work
- Formal framework for joint design time and post
silicon co-optimization - Two-stage optimization under uncertainty with
recourse - Design-time sizing and post-silicon tuning using
adaptive body biasing - Efficient solution using adjustable robust
optimization - Decision variables are affine policy of process
parameters - Introduce the notions of control complexity,
measurement complexity and parameter complexity - Help designers predict the amount of adaptivity
needed
6Adjustable Robust Optimization Paradigm
- What is a formal way of describing our
co-optimization? - Optimization with recourse under uncertainty
- Traditional Two stage stochastic programming
with recourse - Exact solutions difficult for continuous random
parameters - Any non-linearity renders problem NP hard
- Unsuitable for CAD problems
- Adjustable robust optimization (robust
optimization with recourse) - Optimization under uncertainty but formulated for
sets - In some cases, can be used for chance-constrained
(statistical) problems
First stage decisions ? Observation of uncertain
variables ? Tune decision variables
7Robust and Robust Adjustable Optimization
- Robust linear programming
- Adjustable optimization
- Some variables (second-stage) are allowed to
depend on realizations of uncertain parameters - Computationally tractable if adjustable variables
are constrained to be affine functions of the
uncertain variables - Leads to affinely adjustable counterpart
8Choosing Affine Policy
- Joint Minimization of Parametric Yield Loss
- Design-time variable is size (w)
- Post-silicon optimization variable is body bias
(DVSB) - In our case adjustable variable is the amount of
body bias DVSB - Constrain DVSB to be affine function of uncertain
process parameters - p0, p1 and p2 are coefficients determined during
optimization - P (p0,p1,p2) is referred to as body bias policy
9Control and Parameter Complexities
- Effectiveness of tuning depends on spatial
granularity of adaptability - E.g. number of distinct body bias values allowed
- Control complexity n captures flexibility of
adaptation - But also area and routing overhead
- Amount of intra-chip variability influences
effectiveness of tuning - Difficult to compensate large intra-chip
variability - Parameter complexity
10Measurement Complexity
- Measurement infrastructure is an essential part
of tuning methodology - Measurement complexity, k
- Represents the amount of known information about
the structure of variability - Use a single measure k kl kv
- Effectiveness of tuning depends on measurement
complexity - Determines selection of optimal policy
11Incorporating Affine Policy into Objective
- Using first order delay and log-leakage models
- Leakage is log-normal. Expected value
-
- Using affine policy expression
-
- f0, f1, and f2 are linear functions of p0, p1
and p2 - Notice we moved from problem of finding (w and
DVSB) to finding (w and p)
12Formulating Separable Objective Function
- We separate dependence on w and p
- Let Io b w, where b is leakage per unit device
width - Let
- For a vector of gate widths w, expected block
leakage - Minimize expected leakage power under timing
yield - Objective function is linear in size w and
exponential in p
13Problem Formulation and Solution Strategy
- Delay constraints are second order conic path
delays - where
- Final optimization problem
- For efficient solution, solve as a two phase
optimization - w phase sizing at fixed body bias
- Second Order Conic Program
- p phase find body bias at fixed size
- Non-linear, linearize around a fixed point
- Solution is a vector of sizes and optimal policy
P(p0,p1,p2)
14Co-Optimization vs. Heuristic Post-Sizing Tuning
- Joint optimization should also be superior to
disjoint tuning steps - Performed comparison against heuristics that
performs post silicon tuning separately after
sizing - Difficult to pick optimal value of body bias for
cells in design - Performs worse than joint optimization
- Delay spread is higher
- Leakage power consumption is greater
Co-Optimization
Heuristic Disjoint
15Delay Spread Reduction
- Joint optimization is successful in tightening
the delay distribution
16Impact of Changing of Control Complexity
- Increasing number of individually adjustable
circuit clusters allows significant reduction
(for same parameter complexity) - Achieved at the cost of increased implementation
overhead - Proposed optimization allows designers to explore
trade-offs
17Impact of Parameter Complexity (Magnitude of
Intra-Chip Variability)
- As the amount of uncorrelated variability
increases, benefit of post-silicon tuning is
reduced - More difficult to compensate individual regions
- Design time optimization performs better, though
still inferior
18Impact of Increasing Measurement Complexity
- Increasing measurement complexity allows a larger
reduction of leakage power - Improvement depends on parameter complexity
- Larger intra-chip variation requires more sensor
sites
19Conclusions
- Joint co-optimization enables synergy between
design time and post silicon steps - Developed foundation for joint design-time and
post-silicon optimization - Up to 20 savings in leakage power can be
obtained - Introduction of metrics that enable designers to
assess the complexity of biasing circuitry needed - Computationally efficient formulation using
adjustable robust optimization - Good run-time behavior