POLITECNICO DI MILANO

Vincenzo Rana

vincenzo.rana@dresd.org

CITiESCITiES.:: Project Presentation ::..:: Project Presentation ::.

OutlineOutline

• Standard Communication Infrastructures (CIs):• Point-to-point• Bus• Network-on-Chip (NoC)

• The proposed approach:• General overview• Fix part• Reconfigurable slots• A complete example• Adaptation to point-to-point• Adaptation to bus• Adaptation to NoC

• Conclusions and future work

OutlineOutline

• Standard Communication Infrastructures (CIs):• Point-to-point• Bus• Network-on-Chip (NoC)

• The proposed approach:• General overview• Fix part• Reconfigurable slots• A complete example• Adaptation to point-to-point• Adaptation to bus• Adaptation to NoC

• Conclusions and future work

4

Point-to-Point

Point-to-point interconnections• Regular/Uniform

• Well-defined interconnection topology (e.g. full connected graph)

• Flexibility and regularity, but area overhead• Custom

• Ad-hoc interconnection: high-performance and low overhead

• Point-to-point do not scale well, since adding channels requires adding more physical wires

OutlineOutline

• Standard Communication Infrastructures (CIs):• Point-to-point• Bus• Network-on-Chip (NoC)

• The proposed approach:• General overview• Fix part• Reconfigurable slots• A complete example• Adaptation to point-to-point• Adaptation to bus• Adaptation to NoC

• Conclusions and future work

5

Bus (1/3)

• A bus is a set of spatially adjacent links • They define a single, shared communication channel (bus

transparency)• Access to a bus is concurrent, thus contention

resolution is required• An arbiter manages concurrent access requests and assigns

the resource to the user• Area and computational overhead

• Different kind of bus• Hierarchical

• e.g. IBM CoreConnect• Split-bus

• Reduce capacity load

6

Bus (2/3)

Hierarchical bus• The logical communication infrastructure is divided into

subdomains• Each domain is independent to the others• A bridge is used to connect different, independent domains

7

Bus (3/3)

Split bus• Reduce capacity load seen from the user interface• Each segment can be used to address different domains in

the system

OutlineOutline

• Standard Communication Infrastructures (CIs):• Point-to-point• Bus• Network-on-Chip (NoC)

• The proposed approach:• General overview• Fix part• Reconfigurable slots• A complete example• Adaptation to point-to-point• Adaptation to bus• Adaptation to NoC

• Conclusions and future work

8

Network-on-Chip (1/4)

• GENERAL IDEA: borrow theories and applications from the well-known data communication field, e.g. LAN, WAN, MAN...• On-chip network

• RATIONALE: to achieve high-performance communication we need• Reliability• Scalability• Flexibility• Adaptability• Repeatability, ease-to-reuse approach

• Regular structures and concepts

9

Network-on-Chip (2/4)

• XPIPES, first true NoC architecture used for multiprocessing elements based SoC• Highly-parameterizable static NoC with several high-

performance issues• Pipelined inter-router connections• IN/OUT buffering• Reliable communication through communication protocols

• It is defined by a library of network element macros (SystemC defined)

• XPIPES COMPILER, reads the library, reads the user inputs and generate a Verilog instantiable NoC architecture

10

Network-on-Chip (3/4)

• Layered approach to design allows • independent optimization • Separation of concerns• Flexibility

• XPIPES is based on the Smart Stack• Assumptions• The physical layer has non-zero probability of error

• We have to achieve a threshold of reliability• Packet-switched network• End-to-end delivery control based on the use of network

elements

11

Network-on-Chip (4/4)

• Smart Stack layered structure (bottom-up)

DATA LINK LAYER

NETWORK LAYER

TRANSPORT LAYER

Increase reliability of the link (ARQ, FEC)

End-to-end delivery control

Decomposes messages into packets

OutlineOutline

• Standard Communication Infrastructures (CIs):• Point-to-point• Bus• Network-on-Chip (NoC)

• The proposed approach:• General overview• Fix part• Reconfigurable slots• A complete example• Adaptation to point-to-point• Adaptation to bus• Adaptation to NoC

• Conclusions and future work

General overviewGeneral overview

The proposed approach consists of:• a fix part• a set of reconfigurable slots

OutlineOutline

• Standard Communication Infrastructures (CIs):• Point-to-point• Bus• Network-on-Chip (NoC)

• The proposed approach:• General overview• Fix part• Reconfigurable slots• A complete example• Adaptation to point-to-point• Adaptation to bus• Adaptation to NoC

• Conclusions and future work

Fix partFix part

The fix part consists of:• a set of computational components• a set of CI components

• These components cannot be reconfigure at run time, since they have to provide a reliable communication channel between the reconfigurable slots

OutlineOutline

• Standard Communication Infrastructures (CIs):• Point-to-point• Bus• Network-on-Chip (NoC)

• The proposed approach:• General overview• Fix part• Reconfigurable slots• A complete example• Adaptation to point-to-point• Adaptation to bus• Adaptation to NoC

• Conclusions and future work

Reconfigurable slotsReconfigurable slots

Each slot can be filled with:• a computational module• a communication module

• Both these two kind of slot share the same interface, since they have to be interchangeable at run time

Computational modulesComputational modules

Computational modules do not interfere with the communication infrastructure wires• This is possible thanks to the Early Access Partial

Reconfiguration (EAPR) flow

• Computational module logic can use all the resources that are not occupied by the CI logic

CI modulesCI modules

• CI modules can either use the CI wires in order to change their routing or leave them unchanged

• In this way it is possible to dynamically change the CI in order to achieve the desired configuration of communication channels

OutlineOutline

• Standard Communication Infrastructures (CIs):• Point-to-point• Bus• Network-on-Chip (NoC)

• The proposed approach:• General overview• Fix part• Reconfigurable slots• A complete example• Adaptation to point-to-point• Adaptation to bus• Adaptation to NoC

• Conclusions and future work

A complete exampleA complete example

OutlineOutline

• Standard Communication Infrastructures (CIs):• Point-to-point• Bus• Network-on-Chip (NoC)

• The proposed approach:• General overview• Fix part• Reconfigurable slots• A complete example• Adaptation to point-to-point• Adaptation to bus• Adaptation to NoC

• Conclusions and future work

Adaptation to point-to-pointAdaptation to point-to-point

OutlineOutline

• Standard Communication Infrastructures (CIs):• Point-to-point• Bus• Network-on-Chip (NoC)

• The proposed approach:• General overview• Fix part• Reconfigurable slots• A complete example• Adaptation to point-to-point• Adaptation to bus• Adaptation to NoC

• Conclusions and future work

Adaptation to busAdaptation to bus

OutlineOutline

• Standard Communication Infrastructures (CIs):• Point-to-point• Bus• Network-on-Chip (NoC)

• The proposed approach:• General overview• Fix part• Reconfigurable slots• A complete example• Adaptation to point-to-point• Adaptation to bus• Adaptation to NoC

• Conclusions and future work

Adaptation to NoCAdaptation to NoC

OutlineOutline

• Standard Communication Infrastructures (CIs):• Point-to-point• Bus• Network-on-Chip (NoC)

• The proposed approach:• General overview• Fix part• Reconfigurable slots• A complete example• Adaptation to point-to-point• Adaptation to bus• Adaptation to NoC

• Conclusions and future work

Conclusions and future workConclusions and future work

• The proposed approach is just a draft

• In order to use the proposed ideas, it is necessary to explore:• the size of the fix part of the architecture• the size of each reconfigurable slot• the number of slices occupied for the CI for each

reconfigurable slot

The endThe end

•Thank you for your attention

•Do you have any questions?