Satisfying Bandwidth Requirements of GPU Architectures through Adaptive Wavelength Division Multiplexing
Outline
Background: GPU Architecture
Heterogeneous CMP
Networks on Chip
Related Work
Problem Statement
Methodology
Experiments
Speedup Analysis
Heterogeneous NoC
Heterogeneous Photonic Interconnect Model
Detailed Analysis
Bandwidth Selection Policy
BWSP vs. Optimal Performance
Heterogeneous Photonic vs. Wired
Energy-Delay Product
Averages
Conclusions
Questions?
Bandwidth Calculations
Bandwidth Calculations cont.
Photonic NoC
Photonic vs. Wired
Power model
Power Model Cont.
Wired Test Cases
References

Heterogeneous Photonic Interconnect Model

1 Waveguide
Supports up to 64 wavelengths
Two bandwidths
Low: 180 Gbps (18 wavelengths)
Equivalent to 32 B flit size
High: 640 Gbps (64 wavelengths)
Equivalent to 128 B flit size