1. FlexiBuffer: Reducing Leakage Power in On-Chip Network Routers
Gwangsun Kim, John Kim
Dept. of Computer Science
Korea Advance Institute of Science and Technology
Sungjoo Yoo
Dept. of Electronic and Electrical Engineering
Pohang University of
Science and Technology

2. Motivation Use power-gating and turn off unused entries!
[Kumar et al., ICCD’07]
Allocator 3%
Router Power Breakdown
Crossbar
Switch
Input
buffer
Clock
46%
16%
35%
On-chip network is becoming more critical.
Buffer size has a huge impact on performance.
Buffers take a large portion of router power.
However, not all of the buffers are fully utilized even at a high load.
Buffer size
Use power-gating and turn off unused entries!

3. Our Approach Dynamically adjust the active window size.
Active window: set of ON (or active) entries of a buffer.
At a low traffic load
Active window F ON
OFF
At a high traffic load F

4. Issue 1: Flow Control Need to communicate the availability of buffers
Case 1: Increase the active window size using early credit
Router 0
Router 1
Router 2
flit
flit F ON
OFF
credit
credit CR 2 1
When?
There is an incoming flit.
There is an OFF buffer entry.
There is congestion in both upstream and local router.

5. Issue 1: Flow Control (cont’d)
Case 2: decrease the active window size by withholding credit.
Router 0
Router 1
Router 2
flit
flit F F
credit
credit CR 2
When?
There is an outgoing flit.
There is more than the minimum # of ON entries.

6. Issue2: Circular Queue Problem
When utilization is low, each incoming flit turns on an entry.
→ Each activation of an entry incurs power overhead!
Problematic circular buffer
Each flit activates an entry.
Ideal buffer management
The same entry is reused.
OFF ON
OFF
FLIT 0
OFF
FLIT 2
FLIT 4
FLIT 1
FLIT 3
FLIT 0 ON
FLIT 1
OFF ON
OFF
FLIT 2
OFF ON
OFF
FLIT 3
OFF ON
OFF
FLIT 4
OFF ON
OFF
Large power overhead
No power overhead

7. Unified mode Split Queue A buffer is separated into two regions.
Use the primary region only (as long as possible).
Adjust the active window size dynamically.
Operate like a circular queue
Unified mode
FLIT 0 ON
Primary region ON
FLIT 1
FLIT 2 ON
OFF
Not used ON
OFF
Secondary region
OFF
OFF

8. Split Queue (cont’d) Cannot stay in the unified mode indefinitely.
Switch to split mode.
When the primary region is empty,
Switch back to unified mode.
Primary region is empty!
Flits are read out from here.
Primary region
Secondary region
Primary region
Secondary region
FLIT 3 ON
Next flit’s place is
NOT available. ON
FLIT 1
OFF
FLIT 2 ON
Yet, there are unused entries.
FLIT 4 ON
Flits are written to here.
OFF
FLIT 5 ON
OFF ON
OFF
OFF
Split queue
Unified mode

9. Summary of Evaluation 13% 39%
Simulator : Cycle-accurate OCN simulator - Booksim
Power Measurement - Orion 2.0
Parameter
Topology
8x8 2D mesh
Technology node
32nm
# of VCs 4
Clock frequency
1.5GHz
VC buffer depth 8
Vdd
1.0 V
Performance
Power consumption
13%
39%

10. Conclusions
There’s a huge opportunity of power-saving with fine- grained power gating when buffers are large.
Proposed modified credit-based flow control.
Split queue is proposed to minimize activation power overhead.
Our simulation results show that, with minimal performance loss, FlexiBuffer + SQ can save
39% of router power at low traffic load
13% of router power at high traffic load

11. For more discussion, please come to my poster!
Thank you!
Questions?
For more discussion, please come to my poster!