Photonic Networks-on-Chip for Future Generations of Chip Multiprocessors
read more
Citations
Device Requirements for Optical Interconnects to Silicon Chips
Single-chip microprocessor that communicates directly using light
A role for graphene in silicon-based semiconductor devices
All-photonic quantum repeaters
Silicon Photonics Design: From Devices to Systems
References
Networks on chips: a new SoC paradigm
Principles and Practices of Interconnection Networks
Route packets, not wires: on-chip interconnection networks
The omnet++ discrete event simulation system
Optical Networks: A Practical Perspective
Related Papers (5)
Frequently Asked Questions (21)
Q2. What are the future works in "Photonic networks-on-chip for future generations of chip multiprocessors" ?
Much of this is ongoing work as the authors plan to address many of these challenges in follow-up publications. Detailed study of other design issues such as process integration, design complexity, and area overhead is also important to evaluate the feasibility of this project and is an interesting area for future research. This paper aims at laying the groundwork for future research progress by providing a complete discussion of the fundamental issues that need to be addressed to design a photonic NoC for CMPs. The authors have carefully reviewed the recent major progress made in both academia and industry and they expect that, within a small number of years, the enabling technologies will gradually become available to the designers of silicon-integrated circuits.
Q3. What is the role of photonics in the on-chip global interconnect structures?
The insertion of photonics in the on-chip global interconnect structures for CMP can potentially leverage the unique advantages of optical communication and capitalize on the capacity, transparency, and fundamentally low energy consumption that have made photonics ubiquitous in long-haul transmission systems.
Q4. Why are optical buses limited in the number of terminals?
optical buses are limited in the number of terminals due to the finite launching power and coupling losses incurred by each terminal.
Q5. What are the two functions needed for packet switching?
While photonic technology offers unique advantages in terms of energy and bandwidth, two necessary functions for packet switching, namely, buffering and header processing, are very difficult to implement with optical devices.
Q6. What is the importance of a balanced latency penalty?
The latency penalty that results from the increased hop-count should be balanced against the latency reduction achieved by mitigating contention such that an optimal latency point is found.
Q7. What is the topology of choice in the design?
The topology of choice in their design reflects the characteristics of the entire system—a CMP, where a number of homogeneous processing cores are integrated as tiles on a single die.
Q8. What should be used to validate the network design?
Scientific benchmarks or real applications should be used to validate the network design and assist in exploring routing algorithms, topology, flow control, and other design decisions.
Q9. What is the problem of blocking of dimension-order routing?
Even though injection-ejection blocking situations are completely avoided and so are the associated performance penalty and possible deadlocks, the problem of intradimensional blocking of dimension-order routing still remains.
Q10. What is the role of the interconnect in the future of CMPs?
With the future path clearly toward further multiplication of the on-chip processing cores, CMPs have begun to essentially resemble highly parallel computing systems integrated on a single chip.
Q11. How much power is reduced over a baseline electrical bus?
The work shows a latency reduction as high as 50 percent for some applications and a power reduction of about 30 percent over a baseline electrical bus.
Q12. What is the performance metric used to evaluate the improvement gained by adding the paths?
The performance metric used to evaluate the improvement gained by adding the paths is again the path-setup overhead ratio, which is derived directly from the path-setup latency.
Q13. How many switches can be integrated in a photonic layer?
If the authors assume that the area of the 4 4 switch is about 5,000 m2, then, theoretically, more than 80,000 such switches can be integrated in the photonic layer of a 400 mm2 die.
Q14. What is the effect of the large block size on the latency?
As predicted, for small block sizes ð 1 KbyteÞ, the overall latency is dominated by the path-setup overhead, which is greater than the serialization latency, because of the extremely large transmission bandwidth.
Q15. How many ns can be used to transmit a 6 kbyte message?
The authors note that, with a peak transmission rate of 960 Gbps (using WDM, see Section 6.2), 50 ns can be used to transmit a 6 Kbyte message, e.g., a DMA block.
Q16. What is the average bandwidth of the optical network?
The average bandwidth is calculated as the product of the peak bandwidth and the fractional time, in steady state, that can be allocated for actual transmission of the optical messages, after messages have been set up.
Q17. How many dB is the optically pumped PSE capable of switching?
Recent results reported in [19] (see Fig. 1a) demonstrate an optically pumped PSE with a measured insertion loss of 2.5 dB in the pass-band, capable of simultaneously switching nine 40 Gbps wavelengths.
Q18. How much latency can be reduced by dropping every blocked packet?
The simulation results are reported in Fig. 11: By setting the buffer depth to 0, i.e., dropping every blocked packet and immediately notifying the sender, the path-setup latency can be reduced by as much as 30 percent when compared to the case where path-setup packets are not dropped on contention (buffer depth of 2).
Q19. How can the network handle packet-switched traffic?
when packet sizes are fairly large, the setup time in their NoC can be considered fast and the network can handle packet-switched traffic with reasonable latency.
Q20. What is the energy required to transmit a flit across a link and a?
EFLIT HOP , the energy expended to transmit one flit across a link and a subsequent router, is computed based on the energy estimates in Table 3 as well as the link length and flit-width, which vary for different technology nodes.
Q21. What is the main difference between a photonic and an electronic noC?
In particular, photonic NoCs can deliver a dramatic reduction in power expended on intrachip global communications while satisfying the high bandwidths requirements of CMPs.