Title: LOW-LEAKAGE REPEATERS FOR NETWORK-ON-CHIP INTERCONNECTS
1LOW-LEAKAGE REPEATERS FOR NETWORK-ON-CHIP
INTERCONNECTS
Technion Israel Institute of Technology
- Arkadiy Morgenshtein, Israel Cidon, Avinoam
Kolodny, Ran Ginosar
QNoC Research Group Electrical Engineering
Department Technion Israel Institute of
Technology Haifa, Israel
2Highlights
- Leakage in NoC links with repeaters
- Selecting the Repeater Type
- Optimizing Repeater Insertion
- Utilization-Oriented Analysis
3Networks-on-Chip (NoC)
- NoC characteristics
- Packet-based data routing
- Multiple Quality-of-Service levels
- Physical layer of NoC
- Low link utilization
- Most links idle most of the time!
- Leakage power is important
4Leakage Reduction in Logic
Subthreshold leakage is dominant at high
temperatures
Solutions
5Leakage Reduction in Repeaters
Solutions
?
specific solutions needed
6Existing Repeater Types
LVT Low-Vt Repeaters
HVT High-Vt Repeaters
SVT - Staggered-Vt
- fast (In 0?1)
- slow (In 1?0)
- low leakage (idle)
16 Sylvester et al.
7Research Outline
Network-on-Chip
Low Varying Utilization
Selecting the Repeater Type
Optimizing Repeater Insertion
Utilization-Dependant Optimal Number of Repeaters
SR Sleep Repeaters
DTD Dual-Vt Domino Repeaters
Utilization-Oriented Analysis
8Dual-Threshold Domino (DTD) Repeaters
High-Vt Evaluation Transistors
Low-Vt Pre-charge Transistors
synchronized Clk link
9DTD Repeaters Operation
- Precharge transistors disconnected
- CLK line is synchronized with Data
- Evaluation by HVT transistors slower but
tolerant to Vt fluctuations - Each Evaluation transistor drives only one
transistor at next stage faster and can be
down-sized
10DTD Repeaters Operation
- Evaluation transistors disconnected
- Precharge to low-leakage mode
- Precharge by LVT transistors - fast
11DTD Repeaters Operation
X
X
0
0
1
1
X
X
X
X
1
0
1
0
X
X
- HVT transistors are off low leakage
0
1
0
1
12DTD Repeaters Operation
X
X
X
X
0
0
1
1
X
X
X
X
X
X
X
X
1
0
1
0
X
X
X
X
- For Data0 - no transition occurs
13DTD Repeaters Operation
14DTD Highlights
- Application of domino and double-Vt techniques
to low-leakage repeaters
Benefits Effective leakage reduction during
standby Reduced load on each repeater allowing
downscaling and area reduction Tolerance to VT
fluctuations by using HVT evaluation transistors
Drawbacks - Increased dynamic power consumption
due to signaling in domino protocol - Overhead of
clock line and pre-charge wiring
15Sleep Transistors in Repeaters
MTCMOS
Logic
Evolution
SR
Repeaters
16MTCMOS in Repeaters
- Common sleep transistors insertion
-
- Both NMOS and PMOS are used
- All stages enter and exit sleep mode
simultaneously - LARGE sleep transistors
- High routing complexity and wiring overhead
17Repeaters with Per-Stage Sleep Transistor
- Distributed sleep transistors along the link
-
- Each stage of repeaters has a separate pair of
sleep transistors - One stage of repeaters is active
- - others are in low-leakage
- standby
- - Sleep Transistor is heavily loaded and has to
be scaled with link width
18Repeaters with Per-Stage Sleep Transistor
- Distributed sleep transistors along the link
-
- Each stage of repeaters has a separate pair of
sleep transistors - One stage of repeaters is active
- - others are in low-leakage
- standby
- - Sleep Transistor is heavily loaded and has to
be scaled with link width
active
sleep
sleep
19Repeaters with Per-Stage Sleep Transistor
- Distributed sleep transistors along the link
-
- Each stage of repeaters has a separate pair of
sleep transistors - One stage of repeaters is active
- - others are in low-leakage
- standby
- - Sleep Transistor is heavily loaded and has to
be scaled with link width
active
sleep
sleep
20Repeaters with Per-Stage Sleep Transistor
- Distributed sleep transistors along the link
-
- Each stage of repeaters has a separate pair of
sleep transistors - One stage of repeaters is active
- - others are in low-leakage
- standby
- - Sleep Transistor is heavily loaded and has to
be scaled with link width
active
sleep
sleep
21SR Sleep Repeaters
Parallel link using individual zigzag sleep
transistors
- One sleep transistor per repeater
-
- Smaller sleep transistors
- Simpler routing
- Zigzag connection
- Only to transistors that are off during sleep
- Number of sleep transistors is reduced by 50
1
0
0
1
0
0
22Sleep Repeater Highlights
- Novel Efficient sleep transistors for repeaters
Benefits Effective leakage reduction during
standby Optimized structure according to
specifics of repeater insertion
Drawbacks - Area overhead - Increased dynamic
power consumption due to additional transistors
23Simulation Setup
- 65nm BPTM models for transistors and
interconnect - 32-bit link operating at 105C temperature
- LVT design was used as baseline for repeater
insertion - Scaling factor was adjusted for SVT, DTD and SR
to meet the delay target equal to LVT - Area, delay and energy were obtained for each
of the compared techniques
24Total Repeater Area
DTD smallest area - SR largest area
25Energy vs. Utilization
- SVT Least energy at high utilization
- SR Least energy at low utilization
8mm link
26Utilization-Dependant Optimal Number of Repeaters
Set Target Delay (DltDopt)
Repeaters Sizing for (1ltKltn)
Optimal K for minimal Leakage Power - Kleak
Optimal K for minimal Dynamic Power - Kdyn
!
Calculate Ratio of Total Power for Kdyn and Kleak
vs. Utilization
Find for which utilization rates k_leak or k_dyn
is optimal
27Optimal Number of Repeaters
example
- For each k a suitable sizing factor h is found
to meet the target delay - Optimal k for minimal leakage is kleak4
- Optimal k for minimal dynamic power is kdyn6
-
kleak
kdyn
Power vs. k for target D309ps (instead of
Dmin280ps), L10mm
28Number of Repeaters vs. Utilization
example
- Total power as function of utilization for Kdyn
and Kleak - Power ratio is calculated for Kdyn and Kleak
- Break-even point is at 40 utilization
- The results of Kleak are up-to 17 better at
low utilization rates
Prefer Kleak
Prefer Kdyn
Power ratio of Kdyn vs. Kleak
29Summary
- DTD (Dynamic Dual-Threshold) Repeaters
- SR (Sleep Repeaters)
- Zig-zag structure
- SR least power at low utilization
- Thanks to low leakage
- Optimal number of repeaters depends on link
utilization
30Questions?