Evaluation and Design TradeOffs Between CircuitSwitched and PacketSwitched NOCs for ApplicationSpeci

1
Evaluation and Design TradeOffs
BetweenCircuitSwitched and PacketSwitched
NOCs for ApplicationSpecific SOCs

• Kuei-Chung Chang, J.-S. Shen, T.-F. Chen
• DAC 2006, San Francisco
• July 25, 2006

Computer Science Information Engineering Nationa
l Chung Cheng University Chia-Yi, Taiwan, R.O.C.
2
Motivation

• Future interconnection design trends
• Parallel chip architecture On-chip network
  (NOC)
• Packet-switched NOC
• However, two interesting questions
• How much design costs?
• What is suitable for application-specific?
• Discussions
• Shared bus is just suitable for small scale
• Packet-switched NOC is too expensive for
  non-large scale
• How about circuit-switched NOC?

How to choose a suitable interconnection
architecture?
3
Objectives

• Detailed comparative evaluations of
  interconnection architectures for
  application-specific designs
• Circuit-switched interconnection
• Segmented bus
• Crossbar
• Packet-switched NOC
• Propose the Crossroad Interconnection
  Architecture as a solution for circuit-switched
  NOC
• Claim It can be a good choice for
  application-specific SOC designs under few tens
  of cores.

4
Outline

• Motivation Objectives
• Interconnection techniques
• Circuit-switched v.s. Packet-switched
• What are the problems with packet-switched NOC?
• Application-specific NOC features
• The proposed circuit-switched interconnection
  architecture Crossroad NOC
• Evaluations for packet-switched and
  circuit-switched NOCs
• How to choose a suitable interconnection?
• Conclusion

5
Outline

• Motivation Objectives
• Interconnection techniques
• Circuit-switched v.s. Packet-switched
• What are the problems with packet-switched NOC?
• Application-specific NOC features
• The proposed circuit-switched interconnection
  architecture Crossroad NOC
• Evaluations for packet-switched and
  circuit-switched NOCs
• How to choose a suitable interconnection?
• Conclusion

6
Circuit-switched InterconnectionSegmented Bus
Chen 1999

• Partition a shared bus into isolated bus segments

• Advantages
• Small load per bus line
• Small delay
• Low energy consumption
• More parallel

• Disadvantages
• Intra-segment
• Shared bus
• Inter-segment
• Cross drivers
• Low performance

7
Circuit-switched Interconnection
(contd)Router-based Lee 2003

• Modules are connected by a N x N router
• Router could be a fully connected crossbar
• Advantages
• Isolation of modules from each other
• Low energy consumption
• High parallel communications
• Disadvantages
• Low extensibility
• High complexity

8
Network-centric Interconnection Packet-switched
NOC Pande 2005

• Modules connected by routers
• Communicate by packets
• Components in the router
• Ingress/Egress packet process units, Arbiter,
  Switch fabrics, Queuing buffer
• Advantages
• High scalability
• High parallelism
• High throughput
• Disadvantages
• Too much design costs

9
What are the problems with packet-switched NOC?
Problem 1 Latency
Ref. Pande 2005

• Message passing latency
• Protocol stack
• Queuing buffer
• When injection load approaches the throughput
  saturation limit, latency starts to increase
  exponentially
• The latency will be high between few hot modules

10
What are the problems with packet-switched NOC?
Problem 2 Power consumption

• Leakage power distribution
• Queuing buffer consumes approximately 64 leakage
  power of the router
• Leakage power becomes increasingly significant of
  the total power

Ref. Chen 2003
11
Application-specific NOC design features

• Customized communication topology
• Well-defined system specifications
• Design and mapping depend on the specific
  communication requirements
• Communication localization
• Traffic will exhibit highly localized patterns in
  custom-built topologies
• Localization makes shorter communication distance

12
Observations

• Packet-switched NOC consumes too much power,
  especially so in the queuing buffer
• For application-specific designs
• Designing energy efficient SOC is more important
  rather than looking for the fastest
  implementation
• High throughput and parallelism could be achieved
  by proper mapping algorithms.
• Circuit-switched NOC proposed

13
Outline

• Motivation Objectives
• Interconnection techniques
• Circuit-switched v.s. Packet-switched
• Whats the problem with packet-switched NOC?
• Application-specific NOC features
• The proposed circuit-switched interconnection
  architecture Crossroad NOC
• Evaluations for packet-switched and
  circuit-switched NOCs
• How to choose a suitable interconnection?
• Conclusion

14
The proposed interconnection Architecture
Crossroad NOC Chang 2005

• Crossroad communication block (CCB)
• A crossroad switch to control communications
  between four directional paths
• circuit switching mechanism, no buffer,
  self-routing
• Overpass mechanism
• Two channels can be built simultaneously
• Easily construct Irregular topologies according
  to application characteristics by mapping tools

15
Crossroad NOC Bus control

• Data transmission transaction
• Step 1 Build the path
• Step 2 Transmit data along the path

16
Crossroad NOC Characteristics

• Features for application-specific designs
• Fully configurable irregular topologies
• Fit the specific traffic requirements by
  specific-purpose mapping algorithms
• Low power, low cost, high performance

• Power optimization by localization
• Profiling the application
• High communicative CCB provides local
  communications

• Better parallel communications
• Overpass mechanism
• Different parallel working regions

17
Outline

• Motivation Objectives
• Interconnection techniques
• Circuit-switched v.s. Packet-switched
• Whats the problem with packet-switched NOC?
• Application-specific NOC features
• The proposed circuit-switched interconnection
  architecture Crossroad NOC
• Evaluations for packet-switched and
  circuit-switched NOCs
• How to choose a suitable interconnection?
• Conclusion

18
Evaluation experimental environment

• Experimental tools
• RTL design in Modelsim
• Synthesize the design in Synopsys design analyzer
• Power measurement in Nanosim
• NOC designs
• 0.18µm technology
• Bus width 69 bits
• Queuing buffer
• Packet-switched NOC eight entries
• Circuit-switched NOC NO

19
Evaluation Comparisons for NOCs

• Evaluation Workloads
• MPEG-4 decoder
• Partitioned into six modules
• Workloads decoding 15 frames
• The same topology for fair comparisons

20
Evaluation Comparisons for NOCs (contd)

• Experimental results
• Area Implementing the interconnection
• Latency Passing one data unit through a switch
• Power Average power of completing the workload
• Performance Total time of completing the workload

21
Evaluation Overall comparisons

• Area, energy and latency characteristics
• Line-shape topologies
• Random traffic workloads
• Latency means
• For packet-switched NOC
• Transmit one data unit through a router
• For circuit-switched NOC
• Transmit one data unit from source to target

22
Evaluation Localization effects

• Metrics energy saving, delay saving
• saving (P.S. C.S ) / P.S
• The energy and latency savings will increase in
  more localized topologies
• More long-distance communications
• More overhead of retransmissions

23
Overall ObservationsHow to choose a suitable NOC
given area/power/latency constraints?

• Constraints of the proposed circuit-switched NOC
• Latency of transmitting a message will be large
  when the scale of the NOC increases
• Uniformly distributed traffic takes more latency
  and energy consumption

If the hops along the critical path is less than
about ten-something, the proposed
circuit-switched NOC can be a good choice.
24
Conclusion

• Three issues
• Why not the packet-switched NOC for
  application-specific SOC designs
• What alternative Crossroad NOC
• How to choose suitable interconnection
• The proposed circuit-switched Crossroad NOC
• Suitable for SOC designs under few tens of cores
• Could still be a good choice even in a
  large-scale SOC by proper mappings

25
The EndThank you!
Keep going