Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
12/1/2012 1
Outline
Reduction Techniques
12/1/2012 2
Trends in TFT Power Consumption
600.00
575.00
Max power
550.00 consumption (W)
Frequency(In Mhz)
525.00
LOC - Fmax
500.00
LOS - Fmax
475.00
Fmax frequency difference
Test Module1 Module2
450.00 suspected to be due to high
switching activity in LoS
425.00 LOS 1.05 1.21
400.00
1 175 349 523 697 871 1045 1219 1393 1567 1741 1915 2089 2263 2437 2611 2785 2959 3133
12/1/2012 4
Most Power Consuming Pattern
Comparisons
LOC LOC LOS LOS LOS launch LOS capture LOS capture
launch capture launch capture Vs Vs Vs
(mW) (mW) (mW) (mW) LOC launch LOC capture LOC launch
1400
1200
200
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Patterns
12/1/2012 6
Root Cause Analysis
12/1/2012 7
Clock Gate Hookup in Current Designs
Combo cloud
Q EN Q
CLK
FF TE FF
Scan enable
Clock gates LOC Launch : Clock to the FF through
L the functional path
Scan enable
Q
Q Combo cloud EN CLK
TE
L FF
FF
Scan enable
Clock gates
LOS Launch : Clock to the FF through
O the Test path
Scan enable
FF will definitely get the
S launch Clock. Only capture
CLK
clock will depend on the
CLK functional enable (EN) of
12/1/2012 the clock gates 8
Reduction Techniques
Can be classified into two broad categories
ATPG techniques
Fill options
Low effort ATPG.
Design techniques
Clock gates TE control
Clock gates FE control
Low compression
12/1/2012 9
1. Reduction Techniques : Fill
Ckt1a : Netlist without compression
LOS launch power ~44% greater than LOC
launch power for Adjacent fill and 0 fill
techniques
1400
1200
1000
adjacent fill
800
Power
random fill
600 0 fill
1 fill
400
200
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Patterns
12/1/2012 10
1. Reduction Techniques : Fill
Ckt1b : Netlist with compression
LOS launch power for the most power consuming
pattern ~90% higher than LOC launch
Effect of Fill techniques on LOS launch for compression inserted
netlist
1200
1000
800
Adjacent fill
Power
Random fill
600
0 fill
400 1 fill
200
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
patterns
12/1/2012 11
2. Reduction Techniques : Modified
LOS
100% flops get clocked during
LOS launch TEN_CG TEN CG
38% flops get clocked during
LOC launch EN
Changing the LOS clocking for
TEN CG
Launch (make it similar to LOC
clocking) EN
12/1/2012 12
Reduction Techniques : TE control of
clock gates and adjacent fill
Ckt1a : Netlist without compression
For the most power consuming pattern LOS
launch is ~18% higher than LOC launch
789 mW 900
800
18%
700
600
665 mW
Power
500
LOS Launch
400 LOC launch
300
Max Power for LOC launch
200
100
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
800
704 mW 700
30% 600
542 mW 500
Power
LOS launch
400
LOC launch
300
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
TEN_CG
TE
Clk_out
ICG cell modified to include extra EN
Latch
-velevel
gate and input TE_new FE Sensitive
latch
D
FE control of clock gates controlled
ICG
during at-speed cycles
Data
If TE_new = 1 only then functional
Q COMBO D Q
D LOGIC
value propagates Clock
12/1/2012 15
3. Reduction Techniques : FE control
of clock gates
TEN_CG
TE
Clk_out
ICG cell modified to include extra EN
Latch
-velevel
TE_new
gate and input TE_new FE Sensitive
latch
D
FE control of clock gates controlled
ICG
during at-speed cycles
Data
If TE_new = 1 only then functional
Q COMBO D Q
D LOGIC
value propagates Clock
12/1/2012 16
3. Reduction Techniques : FE control of
clock gates and adjacent fill (contd..)
Peak
Pattern switching
Activation Count activity
All four groups allowed to toggle 4400 6700
One group at a time followed by all four 5483 6364
Two groups at a time followed by all four 7608 6072
One group at a time followed by two followed
by four 8590 5905
Two groups at a time followed by three
followed by four 7890 5363
Progressive Group Activation ATPG
pass
G1
Suitable for regenerating high switching patterns,
G2 the other low power DFT/ATPG techniques
after
have
G3 been applied.
12/1/2012 17
G4
4. Low compression and low effort
ATPG
7000
7000
6000
6000
switching activity
activity
5000
5000
Peak switching
4000
4000
3000
3000
Peak
2000
2000
1000
1000
0 0
0 0 500 5001000 1000
1500 2000
1500 250020003000 2500
3500 4000
3000 4500
3500 5000 4000 4500 5000
Patterns Patterns
12/1/2012 19
Retarget with Low compression and low
effort ATPG
6000
Peak switching activity
5000
4000
3000
2000
1000
0
0 500 1000 1500 2000 2500
Patterns
12/1/2012 20
Consolidated Flow
Input Design
Pattern re
-generation:
Insert dual Low compression and
compression codec low ATPG effort
Provide TE and FE
Power Estimation
controls to clock gates
(Toggle Count)
Design
Preparation
Threshold
Prune high switching
patterns
Run ATPG
(High Compression, Adj Fill, Updated Fault DB
TE Control Enabled)
Pattern re
-generation:
Progressive enabling of
clock gate FEs
Power Estimation
(Toggle Count)
Gate-Level
Power Estimation
Threshold Prune high switching
patterns
Final Power
Profile
Updated Fault DB
12/1/2012 21
Conclusions
LOS launch power ~95% (for the most power consuming
pattern) higher than LOC launch power.
Several techniques proposed for TFT power reduction
Adjacent Fill + TE Controls
LOS launch power ~18% higher than LOC launch
Pattern volume inflation for Ckt1a and Ckt1b was
21.8% and 52.8%, respectively.
FE controls
LOS launch power comparable to LOC launch power.
The final pattern volume inflation after applying this
technique on the reduced fault set (for the high
switching patterns) for Ckt1b was ~59%.
12/1/2012 22