Title: Architectural-Level Prediction of Interconnect Wirelength and Fanout
1Architectural-Level Prediction of Interconnect
Wirelength and Fanout
- Kwangok Jeong, Andrew B. Kahng and Kambiz Samadi
- UCSD VLSI CAD Laboratory
- abk_at_cs.ucsd.edu
- CSE and ECE Department
- University of California, San Diego
2Motivation
- Early prediction of design characteristics
- Interconnect wirelength
- Interconnect fanout
- Clock frequency
- Area, etc.
- Enable early-stage design space exploration
- Abstractions of physically achievable system
implementations - Models to drive efficient system-level
optimizations - Existing models fail to capture the impact of (1)
architectural and (2) implementation parameters - Significant deviation against layout data
3Existing Models
- Wirelength statistics
- Christie et al. 2000
- Point-to-point wirelength distribution based on
Rents rule - Extends Davis et al. wirelength distribution
model - Significant deviation against layout data
- Fanout statistics
- Zarkesh-Ha et al. 2000
- Error in counting the number of m-terminal nets
per gate - Significant deviation against layout data
- Existing models fail to take into account
combined impacts of architectural and
implementation parameters - ? Question What is the impact of considering
architectural parameters in early prediction of
physical implementation?
4Implementation Flow and Tools
Architectural Parameters
Router / DFT RTL (Netmaker / SPIRAL)
Synthesis (Design Compiler)
Implementation Parameters
Wirelength and Fanout Models
Place Route (SOC Encounter)
Model Generation (Multiple Adaptive Regression
Splines)
Wiring Reports
- Timing-driven synthesis, place and route flow
- Consider both architectural and implementation
parameters for more complete modeling of design
space - Rent parameter extraction through internal
RentCon scripts
5Design of Experiments
- Netmaker ? generation of fully synthesizable
router RTL code - SPIRAL ? generation of fully synthesizable DFT
RTL code - Libraries TSMC (1) 130G, (2) 90G, and (3) 65GP
- Tools Netmaker (University of Cambridge), SPIRAL
(CMU), Synopsys Design Compiler and PrimeTime,
Cadence SOC Encounter, Salford MARS 3.0 - Experimental axes
- Technology nodes 130nm, 90nm, 65nm
- Clock frequency
- Aspect ratio
- Row utilization
- Architectural parameters fw, nvc, nport, lbuf
for routers and - n, width, t, nfifo for DFT cores
6Modeling Problem
?
- Accurately predict y given vector of parameters x
- Difficulties (1) which variables x to use, and
(2) how different variables combine to generate y - Parametric regression requires a functional form
- Nonparametric regression learns about the best
model from the data itself - ? For our purpose, allows decoupling of
underlying architecture / implementation from
modeling effort - We use nonparametric regression to model
interconnect wirelength (WL) and fanout (FO)
?
?
7Multivariate Adaptive Regression Splines (MARS)
- MARS is nonparametric regression technique
- MARS builds model of form
- Each basis function Bi(x) takes the following
form - (1) a constant, (2) a hinge function, and (3) a
product of two or more hinge functions - There are two steps in the modeling
- (1) forward pass obtains model with defined
maximum number of terms - (2) backward pass improves generality by
avoiding an overfit model
?
?
?
8Example Proposed Model
Wirelength Model
B1 max(0, nDFT - 16) B2 max(0, 16 nDFT)
B4 max(0, 16 - width)B1 B5 max(0, util
0.5) B35 max(0,t - 2)B31 WLavg 22.487
0.056B1 - 0.328B2 - 0.003B27 - 0.013B34
Fanout Model
B1 max(0, nDFT - 16) B2 max(0, 16 nDFT)
B4 max(0, nfifo - 2) B30 max(0, width -
16)B9 B33 max(0, 16 - nDFT)B18 FOavg
3.707 0.003B1 - 0.034B2 - - 8.567e-6B30 -
1.225e-5B33
- 2 Models (1) interconnect wirelength, and (2)
interconnect fanout - Closed-form nonlinear equations with respect to
architectural and implementation parameters - Suitable to drive early-stage architectural-level
design exploration
9Impact of Architectural and Implementation
Parameters
- Prop.
- WL and FO are directly modeled from architectural
/ implementation parameters - Model 1
- Rents parameters are modeled from architectural
/ implementation parameters - Model 2
- Rents parameters are modeled from architectural
parameters - ?, ? for impacts from implementation
- Model 3
- Rents parameters are extracted from implemented
layout
10Model Validation
Prop. (FO)
Prop. (WL)
Chr
ZH
- WL estimation error reductions
- DFT max. error 73 (79.5 ? 21.3), avg. error
81 (18.1 ? 3.1) - Router max. error 70 (59.9 ? 17.9), avg.
error 91 (27.2 ? 2.3) - FO estimation error reductions
- DFT max. error 74 (22.7 ? 5.7), avg. error
92 (10.1 ? 0.8) - Router max. error 92 (18.2 ? 1.4), avg. error
96 (5.6 ? 0.2)
11Recent Extensions
- Used described methodology to develop models for
- (1) chip area and (2) total power
- Area model
- Sum of standard cell area whitespace
- On average within 5 of the layout data
- Power model
- Includes both dynamic and leakage components
- On average within 6 of the layout data
12Conclusions and Future Directions
- Proposed a reproducible modeling methodology
based on RTL to layout implementation - Developed accurate DFT and router interconnect
wirelength (WL) and fanout (FO) models - Improvement over Model 3
- WL up to 81 (91) error reduction on average
for DFT (router) cores - FO up to 92 (96) error reduction on average
for DFT (router) cores - Improvement over Model 2
- WL up to 85 (85) error reduction on average
for DFT (router) cores - FO up to 89 (96) error reduction on average
for DFT (router) cores - Future Directions
- Model maximum fclk w.r.t. architectural and
implementation parameters - Estimators of achievable power/performance/area
envelope - Enable efficient system-level design space
exploration