Title: Analytical Thermal Placement for VLSI Lifetime Improvement and Minimum Performance Variation
1Analytical Thermal Placementfor VLSI Lifetime
Improvement and Minimum Performance Variation
Andrew B. Kahng, Sung-Mo Kang, Wei Li, Bao
Liu UC San Diego UC Santa Cruz
2Outline
- Background
- Modeling and Theoretical Results
- Analytical Thermal Placement
- Experiment
- Summary
3VLSI On-Chip Temperature Scaling
Courtesy, Intel
4Temperature Scaling Why and How
- Scaling has led to temperature rise in VLSI
- Higher integration
- Higher clock frequency
- Leakage power
- Cooling techniques are stagnant
- Air ventilation
- Liquid cooling
- Low power design
- Power gating, clock gating, dynamic scheduling
- Placement
5Chip Packaging Structures
- Heat dissipation through bulk silicon in wire
bond packaging
- Devices and interconnects closer to heat sinks in
flip chip packaging
6Heat Dissipation Equations
- From Boltzmanns Equation
- p ( r ) power density
- g ( r ) thermal conductivity
- Electrical analogue RC Circuit
- Thermal conductance G
- Heat capacity C
7Thermal Effects on Performance
- Higher temperature ?
- Superlinear decrease of carrier mobility
- Linear decrease of transistor threshold voltage
- Increase or decrease of transistor output current
depending on transistor threshold voltage, supply
voltage, etc. - Increase of interconnect resistance
8Thermal Effects on Circuit Lifetime
- Circuit lifetime Tf decreases superlinearly with
rising temperature - Hot carriers
- Oxide breakdown
- Electromigration
- where
- J current density
- Q activation energy (1.0eV for copper)
- k Boltzmann constant
- T temperature
- D given by device structure
9Previous Thermal Placers
- Objective
- Total on-chip temperature1
- Maximum on-chip temperature23
- Method
- Simulated annealing34
- Min-cut bi-partition1
- Thermal simulation
- Compute thermal resistance matrix at each
iteration134
- Chao and Wong, Thermal placement for high
performance multichip modules, ICCD, 1995 - Chu and Wong, A matrix synthesis approach to
thermal placement, ISPD, 1997 - Cong, Wei, and Zhang, A thermal-driven
floorplanning algorithm for 3D IC, ICCD, 2004. - Tsai and Kang, Cell-level placement on improving
substrate thermal distribution, IEEE Trans. CAD,
2000
10Outline
- Background
- Modeling and Theoretical Results
- Analytical Thermal Placement
- Experiment
- Summary
11Thermal Modeling
- ? FDM (Finite Difference Method)
- ? MOR (Model Order Reduction)
- ?
Heat source
Boundary thermal resistor
12Objective and Complexity
- Placement for minimum on-chip temperature at a
specific spot is linear - How to locate current sources s.t. Vo is
minimized? - Solved by greedy algorithm
- Locate maximum current source with minimum
resistance
13Objective and Complexity
- Placement for minimum average on-chip temperature
is linear - How to locate current sources s.t. SiVi is
minimized? - Solved by greedy algorithm
- Locate maximum current source with minimum
resistance
14Objective and Complexity
- Placement for minimum maximum on-chip temperature
is NP-hard - Reduces to the bi-partition problem
- Given
-
- we have
i1,2 and i,j on the same side
otherwise
15Outline
- Background
- Modeling and Theoretical Results
- Analytical Thermal Placement
- Experiment
- Summary
16Problem Formulation
- Given
- Chip dimensions 0ltxlta, 0ltyltb, 0ltzltd
- Thermal parameters
- Thermal conductivity k on chip top
- Thermal conductivity kN on chip bottom
- Effective heat transfer coefficient h on chip
bottom - Ambient temperature Tr
- Cells C of power consumption P
- Netlist N
- Find a cell placement which minimizes sum of
total wirelength and maximum temperature
17Analytical Placement
- Approximate the NP-hard placement problem as a
nonlinear optimization problem - Relax the non-overlapping constraint into a cell
density unevenness penalty function - Minimize
relax
legalize
18Cell Density Distribution
- A cell centered at (xc,yc) of width w and height
h distributes its area over a grid of points
(x,y) - where
Cell density
Cell density
1
1
r
-r/2
r/2
x
-r/2
x
r/2
-r
19Smooth Wirelength Function
- Half perimeter wirelength
- Approximate min/max by logarithm of sum of
exponents
20Analytical Thermal Placement
- A, b, g are such that terms are comparative
- G-1 does not change during placement iteration
21Congestion Penalty Function
- If congested sharper increase of penalty ?
stricter enhancement - If not congested no penalty ? more relaxed
22Outline
- Background
- Modeling and Theoretical Results
- Analytical Thermal Placement
- Experiment
- Summary
23Experiment Setting
- We compare analytical thermal placement to
thermal effect oblivious analytical placement
APlace - Two industry design test cases of gate array
logic in 130nm and 180nm technologies
Utilization
Total Power
Technology
rows
blocks
cells
design
0.60
10.0W
130nm
129
0
13397
I
0.43
10.0W
180nm
251
5
7128
II
24Thermal Placement Data Flow
Chip Dimensions Material, Boundary Conditions
Netlist
Thermal Simulation
Power Profile
Thermal Resistances
Analytical Thermal Placement
Temperature Reduction
25A Snapshot of Placement Result
26Analytical Thermal Placement vs. Traditional
Analytical Placement
Test case I 130nm industry design of13K
cells
Placer
HPWL CPU
Max Temp
g
(s)
()
(mm)
()
(K)
APlace
250.15
100.00
504.80
100.00
12.10
0.00
ATP
718.53
92.19
465.35
97.47
11.33
0.00
636.13
92.47
466.78
92.23
11.16
0.02
800.52
95.39
481.54
82.15
9.94
0.10
Test case II 180mm industry design of 7K
cells
Placer
HPWL CPU
Max Temp
g
(s)
()
(mm)
()
(K)
APlace
211.37
100.00
923.55
100.00
2.73
0.00
ATP
204.94
98.06
905.62
109.52
2.99
0.00
496.91
98.03
905.34
99.63
2.72
0.02
507.94
99.55
919.42
69.23
1.89
0.10
27Outline
- Background
- Modeling and Theoretical Results
- Analytical Thermal Placement
- Experiment
- Summary
28Summary
- We propose analytical thermal placement and
achieve 17.85 and 30.77 maximum on-chip
temperature variation reduction and 4.61 and
0.45 wirelength reduction compared with the
existing analytical placement for the two
industry designs, respectively - We present theoretical results on the complexity
of specific spot temperature, average on-chip
temperature, and maximum on-chip temperature
minimum placement as linear, linear, and NP-hard - Future directions
- Thermal effect aware performance optimization
- 3-D thermal placement
29Thanks for your attention!