1. Leveraging Optical Technology in Future Bus-based Chip Multiprocessors
Kevin Kane
EEL 6905
1/25/07

2. Content Background Introduction to paper Optical Technology Overview
Transmitter
Waveguide
Receiver
Opto-electric Bus Architecture
Bus design
Estimations
Simulated Design Evaluation
Results and Conclusions
Future Research

3. Background
Optical technology (like 3D integration) may solve current limitations in chip-to-chip communications.
Ten years ago electrical-optical translation costs and CMOS incompatibility seemed insurmountable.
Benefits include:
high speed, high bandwidth, low-chip power, good electrical isolation, low electrometric inference

4. Computer Systems Laboratory
Introduction to Paper
Computer Systems Laboratory
Cornell University
“Investigate the potential of optical technology as a low-latency, high-bandwidth shared bus supporting snoopy cache coherence in future CMP.”

5. C.M.P.
Chip MultiProcessor (CMP) is symmetric multiprocessing (SMP) implemented on a single VLSI integrated circuit. Multiple processor cores (multicore) typically share a common second- or third-level cache and interconnect.
- Wikipedia -
The goal of a CMP system is to allow greater utilization of thread-level parallelism (TLP), especially for applications that lack sufficient instruction-level parallelism (ILP) to make good use of superscalar processors.

6. Snooping?
Bus snooping is a technique used in distributed shared memory systems and multiprocessors aimed at achieving cache coherence. Every cache controller monitors the bus, awaiting for broadcasts which may cause it to invalidate its cache line.
- Wikipedia -
Cache lines are most commonly is in states "dirty", "invalid” , "valid" or "shared". On read misses, a request for read is broadcast on the bus. All cache controllers are monitoring this bus. The one having the copy in the state "dirty" changes it state to "valid" and sends the copy to the requesting node. On write misses, an invalidation of all cache copies is performed. When writing a block in state "valid", its state is changed to "dirty" and a broadcast is sent out to all cache controllers to invalidate their copies.

7. W.D.M.
Wavelength-Division Multiplexing (WDM) technology which multiplexes multiple optical carrier signals on a single optical fiber by using different wavelengths of laser light to carry different signals. This allows for a multiplication in capacity, in addition to making it possible to perform bidirectional communications over one strand of fiber.
- Wikipedia -

8. Optical Technology Overview
“on-chip modulator-based optical transmission”
Transmitter
Waveguide
Receiver
You can have on

9. Transmitter Optical transmission requires:
Laser source (on-chip and off chip)
Modulator
Modulator driver (electrical) circuit
Laser source: on-chip and off chip laser source; using off-chip means signal loss included in the paper
Modulator: “electrical signal changes the refractive index or the absorption coefficient of an optical path”
this is the optical equivalent to an electrical switch
Modulator driver: controls the capacitive load; a smaller capacitance is best

10. Waveguide Silicon (Si) vs. polymer (?) Silicon in this paper
“polymer’ not a real description
Polymer-based modulators are bulky and require high voltage

11. Receiver Optical-to-electrical conversion
Photodetector and trans-impedance amplifier
WDM applications require a wave-selective filter

12. Opto-Electrical Bus Architecture
32nm process technology
Die Area: 400mm sq.
4 “Cores”
4GHz frequency
Three levels of cache
16 L2 caches split among cores
Optical bus
“…look-like structure,…”
Bus for a total of b address, data, and snoop response bits
w wavelengths
Multiplex-by-node organization
Multiplex-by-node has a possibility to save power over the multiplex by address (do to fewer modulators)

13. Simple 4-core CMP Scheme

14. Optical Bus Topology Characteristics (results)
# of nodes
Topology determines:
Frequency
Area
Power
address wavelengths per node
data wavelengths per node
H-nxkAkD
Best are H4x1A2D and H4x1A3D as they:
1) low laser power
2) more flexible (dynamically allocate wavelengths)

15. Two Cache Configurations
16X256KB L2 vs. 16X256KB L2

16. Electrical Networks (baseline)
Used “the later” as it shows similar area and power characteristics to the opto-electrical

17. Electrical Topology Characteristics (results)
Two possible topologies determine:
Area
Power
Different # of snoop requests per bus cycle

18. Applications (benchmark)
Using MIPS binaries compiled with –O3 optimization level

19. Results and Conclusions

20. Average Number of Bus Requests
Assuming infinite bus bandwidth

21. Performance improvement
Interestingly, in spite of the higher snoop request bandwidth H-4x1A1D experiences a significant performance degradation. This is mainly due to its lower per-node data bandwidth (one outgoing port to the optical bus vs. two outgoing ring-ports in the electrical baseline).

22. Latency

23. The Basic Point!
“Indeed, for the configurations under study, the main overall benefits come from reduced contention (and thus effective latency) for data transfers. Our simulations show that the data network struggles to supply the bandwidth needed to satisfy these requests. It is in the data network that the availability of extra wavelengths through WDM yields the largest performance improvements.”

24. Parallel Efficiencies
Scalability improves with the addition of optical technology
Those applications that suffer from more contention in the data network tend to exibit lower parallel efficiencies in all configurations.

25. Results
Optical technology in bus-based CMPs can have a positive impact on performance.
WDM can come at very small additional area and power consumption.
Performance gains from the opto-electric buses could be given back in exchange for power/area savings

26. Future Research
Best use of optical waveguides and WDM for a particular CMP design
Polymer waveguides vs. Silicon waveguides
Polymer (fast but low bandwidth)
Silicon (high bandwidth but slow)
Hybrid system
Further penetrations into the bus protocol
Optical module sensitivity to temperatures variations

27. Thank you