1. QuT: A Low-Power Optical Network-on-chip
Parisa Khadem Hamedani
Natalie Enright Jerger
Shaahin Hessabi
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

2. Introduction: Electrical NoC
Scalability limitation
Power
Network channel and buffering power
Latency
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

3. Introduction: Optical NoC
Power is independent of transmission distance
Small transmission latency
Simple modulation, large data bandwidths (Gbps)
Transmitter
Receiver
Off-chip Laser
Waveguide Optical Switches
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

4. Introduction: Optical NoC Challenges
Insertion Loss
The loss of signal power resulting from the insertion in an optical path
Main factor in the power consumption
Number of Microrings
Major source of faults
Number of Wavelengths
Wavelength-division multiplexing (WDM)
Total power is proportional to the number of wavelengths
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

5. Introduction: Quarten Topology (QuT)2 1 3 4 6 5 7 8 10 9 11 12 14 13 15
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

6. Outline Introduction Quartern Architecture Methodology Evaluation
Data Network
Router Microarchitecture
Wavelength assignment
All optical switches
QuT WDM Routing
Control Network
Methodology
Evaluation
Conclusion
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

7. Quartern Architecture
A new all-optical architecture
Based on passive microring resonators
Addressing the optical challenges
Ring-based topology
Strategically placed extra links
To reduce the diameter
To reduce number of wavelengths
A new deterministic wavelength routing
Contention-free network
Optimizing optical switches
With an optical control network
qut uses passive microring resonators which route optical streams based on their wavelength.
QuT is a ring-like topology with strategically placed extra links
to reduce the diameter and number of wavelengths required.
Also we proposed a new deterministic routing algorithm which avoids collisions
between streams that need to be modulated on the same wavelength.
this deterministic wavelength routing provides
contention-free network traversal
And reduces the number of required wavelengths by optimizing the optical switches
In QuT, a destination can only accept data from
a single source at a time;
as a result, QuT requires a control network to serialize access by multiple sources to the
destination.
Both the control and data networks are optical.
(QuT is a contention-free topology; data
streams intended for different destinations do not block each
other. )
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

8. Data Network Ring links Cross links Bypass links Unidirectional
Bidirectional 2 1 3 4 6 5 7 8 10 9 11 12 14 13 15
Cross
Cross links
Bidirectional
Even
Bypass links
Unidirectional
Emanate from odd nodes
This slide shows a 16-node QuT network.
QuT has three different connections between nodes
Thr first one isThe Ring links which are bidirectional
The second one is the Cross links which are bidirctional and connected to the even nodes
The last one is the bypass links.
Although bypass links appear to have the same connectivity
as ring links, there are two important differences.
First, they are unidirectional: they only emanate from odd nodes.
Second, their connection within the switch is different from ring links.
We use bypass links, instead of additional cross links in odd
switches, to prevent the data from being absorbed by the wrong
node.
The link for sending data is chosen based on the
distance between the current node and the destination.
(Ring links are used when the distance between the current
node and the destination is less than N/4. Otherwise, Cross
or Bypass links are used for sending data streams.)
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

9. Router Microarchitecture : Wavelength assignment
Each node has:
Dedicated but not unique wavelength
Source uses this wavelength
In an N-node QuT
N/4 distinct wavelength sets
Node i dedicated wavelength set
(i mod N/4) λ0 λ1 λ2 λ3
Each node in the data network has a dedicated but not unique wavelength set.
other nodes will modulate their data onto this wavelength set to communicate
with this node.
We have designed QuT, its deterministic wavelength routing and its optical switches to use a small
number of wavelengths.
In this figure, nodes with the same color have the same dedicated wavelength set.
In an N-node QuT, we use N/4
distinct wavelength sets.
Each wavelength set can include one or more wavelengths
based on the required network bandwidth.
(A wavelength set of (i mod N/4) is assigned to node i.)
(In QuT optical
Streams are routed based on their wavelengths.)
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

10. QuT WDM Routing : Source is even
Distance (Source, Destination):
< N/4
= N/2
Source
Destination
Ring links
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

11. QuT WDM Routing : Source is even
Distance (Source, Destination):
>= N/4
Source
Destination
Cross links
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

12. QuT WDM Routing : Source is Odd
Distance (Source, Destination):
<= N/4
Source
Destination 1
Ring links
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

13. QuT WDM Routing : Source is Odd
Distance (Source, Destination):
> N/4
Source
Destination 1
Bypass links
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

14. QuT WDM Routing: example8
Source: N0
Destination: N8
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

15. Example: Switch at N0 1 8 I1 I2 I3 I4 Ring (Left) Bypass (Left)1 8
Ring (Left)
Bypass (Left)
Cross (Left)
Cross (Right)
Ring (Right)
Bypass (Right) I1 I2 I3 I4
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

16. Example: Switch at N1 2 8 I1 I2 I3 I4 1 Ring (Left) Ring (Right)2 1 8
Ring (Left)
Ring (Right)
Bypass (Left)
Bypass (Right) I1 I2 I3 I4
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

17. Example: Switch at N2 2 1 6 8 I1 I2 I3 I4 Ring (Left) Bypass (Left)2 1 6 8
Ring (Left)
Bypass (Left)
Cross (Left)
Cross (Right)
Ring (Right)
Bypass (Right) I1 I2 I3 I4
Node 2 chooses injection channel I1. Therefore, data are transferred over a cross link connected to the node 6.
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

18. Example: Switch at N6 7 8 I1 I2 I3 I4 6 Ring (Left) Bypass (Left)6 7 8
Ring (Left)
Bypass (Left)
Cross (Left)
Cross (Right)
Ring (Right)
Bypass (Right) I1 I2 I3 I4
Then, the data are turned onto the ring link by AmR between the cross link and ring link
in node 6.
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

19. Example: Switch at N7 2 1 6 7 8 I1 I2 I3 I4 Ring (Left) Ring (Right)2 1 6 7 8
Ring (Left)
Ring (Right)
Bypass (Left)
Bypass (Right) I1 I2 I3 I4
Data are transferred through ring link without changing direction in node 7.
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

20. Example: Switch at N8 I1 I2 I3 I4 E 2 1 8 Ring (Left) Bypass (Left)
Cross (Left)
Cross (Right)
Ring (Right)
Bypass (Right) I1 I2 I3 I4 E 2 1 6 7 8
Finally, in Node 8, data are turned into the ejection channel by DmR.
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

21. Router Microarchitecture: All optical switches (Even)
Ring (Left)
Bypass (Left)
Cross (Left)
Cross (Right)
Ring (Right)
Bypass (Right) I1 I2 I3 I4 E
QuT requires two different switches, one for even and one for odd nodes.
This slide shows the optical switch in the even node.
Microring resonators in the even switches are classified in three groups:
Add Microring Resonators (AmR) they switch optical streams from injection channels or cross links to ring links
(b) Bypass Microring Resonators (BmR) they switch optical streams from bypass links to cross links
(d) Drop Microring Resonators (DmR) they send the data, from ring links into the switch’s detector, if data are modulated on the wavelength dedicated to this node.
AmR, BmR, and DmR are arrays of microrings. (Each microring is sensitive to a specific wavelength. )
DmR is only sensitive to the node dedicated wavelength.
AmR and BmR are sensitive to all wavelengths in QuT.
AddμR
BypassμR
DropμR
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

22. Router Microarchitecture: All optical switches (Even)
Ring (Left)
Bypass (Left)
Cross (Left)
Cross (Right)
Ring (Right)
Bypass (Right) I1 I2 I3 I4 E
QuT requires two different switches, one for even and one for odd nodes.
This slide shows the optical switch in the even node.
Microring resonators in the even switches are classified in three groups:
Add Microring Resonators (AmR) they switch optical streams from injection channels or cross links to ring links
(b) Bypass Microring Resonators (BmR) they switch optical streams from bypass links to cross links
(d) Drop Microring Resonators (DmR) they send the data, from ring links into the switch’s detector, if data are modulated on the wavelength dedicated to this node.
AmR, BmR, and DmR are arrays of microrings. (Each microring is sensitive to a specific wavelength. )
DmR is only sensitive to the node dedicated wavelength.
AmR and BmR are sensitive to all wavelengths in QuT.
AddμR
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

23. Router Microarchitecture: All optical switches (Even)
Ring (Left)
Bypass (Left)
Cross (Left)
Cross (Right)
Ring (Right)
Bypass (Right) I1 I2 I3 I4 E
QuT requires two different switches, one for even and one for odd nodes.
This slide shows the optical switch in the even node.
Microring resonators in the even switches are classified in three groups:
Add Microring Resonators (AmR) they switch optical streams from injection channels or cross links to ring links
(b) Bypass Microring Resonators (BmR) they switch optical streams from bypass links to cross links
(d) Drop Microring Resonators (DmR) they send the data, from ring links into the switch’s detector, if data are modulated on the wavelength dedicated to this node.
AmR, BmR, and DmR are arrays of microrings. (Each microring is sensitive to a specific wavelength. )
DmR is only sensitive to the node dedicated wavelength.
AmR and BmR are sensitive to all wavelengths in QuT.
BypassμR
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

24. Router Microarchitecture: All optical switches (Even)
Ring (Left)
Bypass (Left)
Cross (Left)
Cross (Right)
Ring (Right)
Bypass (Right) I1 I2 I3 I4 E
QuT requires two different switches, one for even and one for odd nodes.
This slide shows the optical switch in the even node.
Microring resonators in the even switches are classified in three groups:
Add Microring Resonators (AmR) they switch optical streams from injection channels or cross links to ring links
(b) Bypass Microring Resonators (BmR) they switch optical streams from bypass links to cross links
(d) Drop Microring Resonators (DmR) they send the data, from ring links into the switch’s detector, if data are modulated on the wavelength dedicated to this node.
AmR, BmR, and DmR are arrays of microrings. (Each microring is sensitive to a specific wavelength. )
DmR is only sensitive to the node dedicated wavelength.
AmR and BmR are sensitive to all wavelengths in QuT.
DropμR
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

25. Router Microarchitecture: All optical switches (Odd)
Ring (Left)
Ring (Right)
Bypass (Left)
Bypass (Right) I1 I2 I3 I4 E
This slide shows the optical switch in the odd node.
Microring resonators in the odd switches are classified in three groups:
(a) (AmR), DmR, which are similar to the Amr and Dmr in even swithches
(c) Cross Microring Resonators (CmR) they turn optical streams from ring links to bypass links
Also CmR are arrays of microrings.
Even nodes benefit from CmRs in odd switches when
they want to send data to the destination that is a distance
of N/2 from them.
therefore, CmR is sensitive to the wavelength which is assigned to the even node to
the left or right of this optical switch.
AddμR
CrossμR
DropμR
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

26. Router Microarchitecture: All optical switches (Odd)
Ring (Left)
Ring (Right)
Bypass (Left)
Bypass (Right) I1 I2 I3 I4 E
This slide shows the optical switch in the odd node.
Microring resonators in the odd switches are classified in three groups:
(a) (AmR), DmR, which are similar to the Amr and Dmr in even swithches
(c) Cross Microring Resonators (CmR) they turn optical streams from ring links to bypass links
Also CmR are arrays of microrings.
Even nodes benefit from CmRs in odd switches when
they want to send data to the destination that is a distance
of N/2 from them.
therefore, CmR is sensitive to the wavelength which is assigned to the even node to
the left or right of this optical switch.
AddμR
DropμR
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

27. Router Microarchitecture: All optical switches (Odd)
Ring (Left)
Ring (Right)
Bypass (Left)
Bypass (Right) I1 I2 I3 I4 E
This slide shows the optical switch in the odd node.
Microring resonators in the odd switches are classified in three groups:
(a) (AmR), DmR, which are similar to the Amr and Dmr in even swithches
(c) Cross Microring Resonators (CmR) they turn optical streams from ring links to bypass links
Also CmR are arrays of microrings.
Even nodes benefit from CmRs in odd switches when
they want to send data to the destination that is a distance
of N/2 from them.
therefore, CmR is sensitive to the wavelength which is assigned to the even node to
the left or right of this optical switch.
CrossμR
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

28. Control Network Multiple-Writer Single-Reader bus Control Packets
Multiple waveguides
Control Packets
Request, ACK, NACK
Small size: 6 bits
Each source node has a dedicated wavelength
In an N-node QuT
N/16 waveguides
N wavelengths
Our CN is based on a Multiple-Writer Single-Reader optical bus.
It is implemented as several waveguides to reduce the number of required splitter.
16 nodes can accept data from each waveguide by using splitters.
However, all nodes can send control packets on each waveguide.
For example, if node 15 wants to send data to node 1, it will modulate its control
packet onto waveguide 0.
As I explained in previous slides, the CN uses three kinds of control packets: request, acknowledgment
(ACK) and negative acknowledgment (NACK).
Each control packet has 6 bits: 4 bits for addressing 16 nodes and 2 bits for encoding the
packet type.
Each source node has a dedicated wavelength and modulates a control packet on its dedicated wavelength.
Therefore, a destination can accept multiple optical streams from multiple sources at a same time (simultaneously as each will be on a different wavelength.)
Control network in an N-node QuT has N/16 waveguides and N wavelengths.
Since the control packets are small, the CN does
not require large bandwidth.
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

29. Methodology - Phoenixsim 64 and 128-node QuT compared against
An event-driven simulator
Based on OMNet++
64 and 128-node QuT compared against
Topology
Number of Wavelengths
Control Network
λ-router N -
Optical Spidergon: Ring-based
N/2
Optical
Corona: Optical crossbar 8
Slot-token-ring
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

30. Outline Introduction Quartern Architecture Methodology Evaluation
Delay
Power
Energy
Throughput
Area
Conclusion
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

31. Evaluation Constant optical bandwidth for all-optical NoCs
Each node has 8 distinct wavelengths
Data stream is modulated on 8 wavelengths assigned to the destination
Die size: 225 mm
Packet size: 256 bits
10Gb/s modulator and detector
Synthetic traffic patterns:
Random, Bitreverse, Neighbor, Tornado and Hotspot-30%
We use the constant optical bandwidth for all-optical NoCs;
we assign eight distinct wavelengths to each node.
Therefore, each optical data stream is modulated on eight wavelengths assigned to
the corresponding destination.
We assume the size of the die is the same
for 64 and 128 nodes (225 mm2).
We assume a data packet size
of 256 bits and 10Gb/s modulator and detector.
We evaluate the topologies under the synthetic traffic patterns:
random, bitreverse, neighbor, tornado and hotspot in which a
random node receives 30% of all requests and the remaining
70% is uniformly distributed.
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

32. Delay: Packet latency (cycle)
128-node:
Offered Load = 0.5
This slide shows the average packet latency at Offered load = 0.5
for different traffic patterns for 64- and 128-node networks,
Latency is reported in processor cycles. We assume a 5 GHz clock for the processors.
Waiting time in a processor’s output buffer (coupled with) and
the delay (associated with) of modulating the packet, primarily
contribute to the latency in QuT.
Corona uses slot-token-ring arbitration to reserve a destination,
which has better performance than QuT’s control network.
However, it consumes more power and energy.
lambda-router has better latency compared to QuT since we assume
that lambda-router can accept data from all of the sources at a time.
Buffering for one packet per source is added to each destination.
Therefore In 128-node lambda-router, buffer size in each node is 127 times larger than buffer size in other optical nocs
If the buffer size in lambda-router is reduced, the
network latency will increase sharply.
(When the network size grows
from 64 to 128 nodes, the average optical path latency only
increases by two cycles. Altogether, QuT’s latency does not
rapidly increase for 128 nodes compared to 64 nodes, assuming
the packet size, the number of wavelengths used to modulate
a packet and the network offered load are the same.)
(The latency in 64- and 128-node QuT is up to 32% and 24%
higher compared to 64- and 128-node Corona, respectively.)
Waiting time in a processor’s output buffer
The delay of modulating the packet
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

33. Power (W) Small Insertion loss, Small number of required wavelengths,
Total power consumption in an optical NoC includes:
off-chip laser power, on-chip microring heating power and electrical to
optical (E/O) and optical to electrical (O/E) circuit power
Total power In QuT, Spidergon and Corona is the summation of
total power consumption in the data and control networks.
Power consumption increases for the 128-node optical NoCs
due to increases in IL, the number of microrings and the total
number of wavelengths.
Reducing the number of required wavelengths can reduce the power consumption.
There is a trade-off between the bandwidth and the power consumption in an optical NoC.
Among these optical NoCs, Corona’s total number of wavelengths in data network remains constant when the network
size grows.
However, Corona and its CN have the largest IL for 64 nodes among the other optical NoCs
Therefore, increasing its IL, its CN IL and the
number of wavelengths in CN for 128 nodes, dramatically increases
Corona’s power consumption.
Small Insertion loss,
Small number of required wavelengths,
Small number of microrings
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

34. Energy-per-bit (pJ) Lower power dissipation
128-node:
Energy consumption in an optical NoC includes the energy
dissipation in the laser, micro-ring heating and back-end
circuitry, E/O and O/E conversion.
In QuT and Spidergon, the average energy-per-bit includes
the energy consumption in the data network and in the CN for
multiple request packets per data packet if necessary.
Energy consumption in Corona includes the energy consumption in
both the data and control network.
For hotspot traffic in QuT and Spidergon, a large number
of requests are sent through the CN for a specific node.
Due to destination contention, energy-per-bit sharply increases ,.
At the saturation point in QuT and Spidergon, a small fraction of energy-per-bit
is related to data network,
Since QuT has lower power dissipation and a smaller average optical path delay, it has
lower energy-per-bit compared to Corona and lambda-router.
(e.g. the data network consumes
9.11% and 7.2% of the energy-per-bit in the 64-node QuT and
Spidergon, respectively. )
Corona’s CN always consumes energy even when the data network is idle
Also, Corona’s data network has an order of magnitude larger energy consumption compared with the CN.
therefore, the average energy-per-bit in Corona at the saturation
point in hotspot traffic does not increase significantly.
Lower power dissipation
Smaller average optical path delay
At the saturation point, a small fraction of energy-per-bit is related to data network
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

35. Normalized Throughput-per-watt
64-node:
Better throughput-per-watt, when the network size increases
128-node:
This slide shows the normalized average
throughput-per-watt for optical NoCs.
64-node QuT has lower throughput-per-watt than Corona.
However, since QuT consumes less power, 128-node QuT achieves throughput-per-watt
improvements over the other optical NoCs,
which indicates better scalability for QuT.
(of up to 43%, 28% and 85% over 128-node
Corona, Spidergon and -router, respectively, )
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

36. Normalized Area
154%
44%
Area overhead in an optical NoC includes the area required for microrings, waveguides and detectors.
This slide shows the normalized area for optical NoCs considered in this work.
128-node Spidergon and lambda-router require (44% and 154%) more
optical area than QuT, since they have more microrings.
Also, Lambda-router uses more detectors to allow each node to accept data
from different sources simultaneously.
Although Corona has fewer microrings, it needs more waveguides and detectors
compared to QuT. Therefore, its area is almost equal to QuT.
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

37. Conclusion Considering optical challenges Insertion loss
Number of microrings
Number of wavelengths
Topology
Insertion Loss
Number of Wavelengths
Number of Microrings
Control Network
QuT
Small
N/4
Optical
λ-router
Large N
Largest -
Spidergon
Smallest
N/2
Corona 8
Slot-token-ring
Consuming Less power and Energy:
Scales better than state-of-art proposals
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip

38. Thank you for your attention!
Question?
Khadem Hamedani et al., QuT: A Low Power Optical Network-on-Chip