Firefly: Illuminating Future Network-on-Chip with Nanophotonics

Presentation Transcript

Firefly: Illuminating Future Network-on-Chip with Nanophotonics Yan Pan, Prabhat Kumar, John Kim † , Gokhan Memik, Yu Zhang, Alok Choudhary EECS Department Northwestern University Evanston, IL, USA {panyan,prabhat-kumar,g-memik, yu-zhang,a-choudhary} @northwestern.edu † CS Department KAIST Daejeon, Korea jjk12@cs.kaist.ac.kr

Motivation On-Chip Network Topologies Network-on-chip is basic for execution. Work [MIT RAW] [TILE64] [Teraflops] C-Mesh [Balfour'06] [Cianchetti'09] Crossbar [Vantrease'08] [Kirman'06] Others: Torus[Shacham'07], Flattened Butterfly[Kim'07], Dragonfly[Kim'08], Hierarchical(Bus&Mesh)[Das'08], Clos[Joshi'09], Ring[Larrabee], …

Motivation Signaling innovations Electrical flagging Repeater inclusion required Bandwidth thickness (up to 8 Gbps/um) [Chang HPCA'08] Nanophotonics Bandwidth thickness ~100 Gbps/μ m !!! [Batten HOTI'08] Generally separate autonomous power utilization Speed of light  low inertness Propagation Switching [ Cianchetti ISCA'09 ]

Motivation Nanophotonic segments full finders Ge-doped Basic segments coupler waveguide off-chip laser source thunderous modulators

Motivation Resonant Rings Selective Couple optical vitality of a particular wavelength Radius r  Baseline Wavelength Temperature t  Manufacturing mistake amendment Carrier thickness d  Fast tuning by charge infusion

Motivation Putting it together Modulation & location ~100 Gbps/μ m transmission capacity thickness [Batten HOTI'08] 10001011 11010101 64 wavelengths DWDM 3 ~ 5 μ m waveguide pitch 10Gbps for every connection 10001011 11010101 ~100 Gbps/μ m transfer speed thickness

Motivation What's the catch? Control Cost Ring warming Laser Power E/O & O/E changes Distance inhumane For short connections (2.5mm) Nanophotonics Electrical RC lines with repeater inclusion For long connections Nanophotonics Cost remains the same Electrical Cost increments [Batten HOTI'08] [Cheng ISCA'06]

Motivation Here is the thought … Design an engineering that separates activity. Utilize electrical motioning for short connections. Utilize nanophotonics just for long range movement. What do we pick up? Low idleness High transmission capacity thickness High power productivity Localized mediation Scalability

Architecture of Firefly Outline Motivation Architecture of Firefly Evaluation Conclusion

Architecture of Firefly Layout View of 64-center Firefly Concentration 4 centers share a switch 16 switches

Architecture of Firefly Layout View of 64-center Firefly Concentration Clusters Electrically associated Mesh topology 4 switches per group 4 bunches Cluster 0 (C0) Cluster 1 (C1) Cluster 2 (C2) Cluster 3 (C3)

Architecture of Firefly Layout View of 64-center Firefly Concentration Clusters Assemblies Routers from various groups Optically associated Logical crossbars

Architecture of Firefly Layout View of 64-center Firefly Clusters Electrical CMESH Assemblies Nanophotonic crossbars Nanophotonic Crossbars Efficient nanophotonic crossbars required!

Architecture of Firefly Nanophotonic crossbars Single-Write-Multiple-Read (SWMR) [Kirman'06] (CMXbar † ) Dedicated sending direct Multicast in nature Receiver think about & dispose of High fan-out  laser control † [Joshi NOCS'09] SWMR Crossbar

Architecture of Firefly Nanophotonic crossbars Multiple-Write-Single-Read (MWSR) [Vantrease'08] (DMXbar † ) Dedicated accepting channel Demux to channel Global assertion required! † [Joshi NOCS'09] MWSR Crossbar

Architecture of Firefly Reservation-helped SWMR Goal Avoid worldwide mediation Reduce control Proposed plan Reservation channels Narrow Multicast to hold Destination ID Packet length Uni-cast information bundle R-SWMR Crossbar

Architecture of Firefly Router Microarchitecture Virtual-channel switch Added optical connection ports and additional support. Isolate accepting channels from different groups. Devoted sending channel for all activity.

Architecture of Firefly Routing (FIREFLY_dest) Routing Intra-bunch steering Traversing optical connection

Architecture of Firefly – another look Clusters Short electrical connections Concentrated work Assemblies Long nanophotonic joins Partitioned crossbars Benefits Traffic territory Reduced equipment Localized assertion Distributed between group data transfer capacity

Evaluation Outline Motivation Architecture of Firefly Evaluation Conclusion

Evaluation Setup Cycle-precise test system (Booksim) Firefly versus CMESH, Dragonfly † and OP_XBAR Synthetic movement examples and follows Electrical Hybrid Optical Hybrid [ † Kim et al, ISCA'08]

Evaluation Load/Latency Curve Throughput Up to 4.8x over OP_XBAR At slightest +70% over Dragonfly 4.8x 70% Bitcomp, 1-cycle Uniform, 1-cycle

Evaluation Energy Breakdown Reduced equipment by apportioning Reduced warming Throughput affect Locality 34% vitality diminishment over OP_XBAR with region

Evaluation Technology Sensitivity α is warming proportion and β is laser proportion. Firefly favors movement area. bitcomp taper_L0.7D7

Conclusion Technology impacts design New open doors in nanophotonics Low idleness, high transfer speed thickness Tailored structures required Firefly profits by nanophotonics by giving Power Efficiency Hybrid flagging Partitioned R-SWMR crossbars  Reduced equipment/control Scalability Scalable between bunch transmission capacity Low-radix switches/crossbars