Assaf Shacham, Keren Bergman, Luca P. Carloni Presented for HPCAN Session by: Millad Ghane NOCS’07
Scaling Transistor Speed and Integrity Tighter Logic Power Dissipation Increasing Number of Cores per Chip Bottleneck: Global Intrachip Communications ▪ Bandwidth ▪ Power Performance-per-watt
Low Power Dissipation Ultra-high Throughput Minimal Latency End-to-end Transmission No Repeating No Regeneration No Buffering A silicon ring resonator
Photonic Interconnection High Bandwidth Messages Electronic Interconnection Low Bandwidth Messages Short Control Messages
Advantages Bit-rate Transparency ▪ Not Switching by Every Bit of Data ▪ Switching Once per Message Low Loss in Optical Waveguide ▪ Independence of Transmission Distance Disadvantages No Storage Element E/O and O/E Conversions ▪ Off-chip Lasers
D D S Sending path-setup packet Electronic Network Optic Network S
D S D S Sending optical packet
D S D S Sending path-teardown packet
D S D S Sending path-blocked packet path-teardown packet
70 µm Photonic Switching Element Electrical Router
No Injection/Ejection Port Not Deadlock-Free Wide Turns
Injection Torus Address Ejection Gateway Switch Injection Switch Ejection Switch Torus Network Nodes Access Points Address Format
Injection-Ejection Blocking [previous slide] Intra-dimentional Blocking (Torus Network) Virtual Channel Flow Control ▪ Circuit Switching terminate-on-timeout packet
POINTS Simulator Based on OMNET++ 36-core CMP 6x6 planar Chip size: 20 x 20 mm Uniform Traffic 3 Cases Deadlock Message Size Optimization Increasing Path Diversity 2
Long path-setup latency Nonoseconds Super fast transmission Overhead Ratio:
Message Duration: 50ns Message Size: 2KB Suitable for DMA Trans
PD=2 Less Hardware Less Overhead Difference