DescriptionHW Assignment 4
General policy for all assignments:
• Late or unsuccessful submission of the assignment is 0 credit on the
assignment. This policy is strictly applied. Submit your assignment
in PDF format on the specified due date in course schedule, and
before the deadline that is 11:30 PM. This is an individual
assignment. Sharing your solution, copying from others, or copypasting from lecture slides is considered plagiarism. The penalty in
the course policy for such act is 0 for the assignment in the first
time, and F in the course for the second time. Submit your answer to
BB on or before the due date.
•
Please remember that late submission or failing to submit the
assignment has 0 credit.
•
You may not share your answers with other students.
Part I
1- (10 points) Define the following concepts:
• Ripping
• Frame-rate
• Color-depth
• Bitmap resolution
• Sampling rate
2- (10 points) How much space is required to store 1 hour of fidelity music if the
height of each high-fidelity sound sample is 16 bits, and the sampling rate is
44.1 kHz? Show how you calculated the answer.
3- (10 points) In a two-column table, indicate the differences between bitmap and
vector graphics.
4- (5 points) Pick a color in RGB color system and represent the color in decimal,
binary, and hexadecimal. For instance, a color with decimal representation
(0,255,255) has hex representation (00FFFF) and binary representation (0000
0000 1111 1111 1111 1111).
5- (5 points) Explain how many bits are used to represent a color in a True-color
format? How many different color values are represented in True-Color format.
Include your calculation.
1
6- (10 points) What is the storage needed to store one video frame with 720×486
resolution in True-color format? If frame rate is 32 fps, how much data is
required for one minute of the video?
7- (10 points) Show the RLE for the following image data. Remember that pixels
are scanned in a zigzag fashion.
• (5 points) First show the scanned values in a vector/array.
• (5 points) Then provide the RLE.
9
9
3
0
66
5
5
6
9
9
2
0
66
5
5
6
9
9
3
255
66
5
5
6
9
1
3
255
66
5
6
6
9
1
3
255
255
6
6
6
8
1
2
0
255
6
6
6
8
1
2
0
255
6
6
6
8
1
2
0
255
6
6
6
8- (5 points) Explain the difference between lossy and lossless image
compression with examples of image file format that suppport each category.
9- (5 points) what is the difference between speech synthesis and voice
recognition technology? What type of audio format supports speech synthesis?
Part II
10- (10 points) Download the following paper from Blackboard first (“paper1 for
assignment 4”),
Joo-Young Kim, Junyoung Park, Seungjin Lee, Minsu Kim, Jinwook Oh, Hoi-Jun
Yoo:
A 118.4 GB/s Multi-Casting Network-on-Chip With Hierarchical Star-Ring
Combined Topology for Real-Time Object Recognition. J. Solid-State
Circuits 45(7): 1399-1409 (2010)
Next, rank the references 1-5 in this paper (from high-to-low) according to
Google scholar. Use the following format:
[Ref#], [Title of the paper] , [Citation number in Google Scholar]
11- (20 points) Use DBLP computer science bibliography website to list all of the
publishers (conferences/workshops/journals) the authors of the given paper
have published to (e.g., IEEE Micro).
2
IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 45, NO. 7, JULY 2010
1399
A 118.4 GB/s Multi-Casting Network-on-Chip
With Hierarchical Star-Ring Combined Topology
for Real-Time Object Recognition
Joo-Young Kim, Student Member, IEEE, Junyoung Park, Student Member, IEEE,
Seungjin Lee, Student Member, IEEE, Minsu Kim, Student Member, IEEE, Jinwook Oh, Student Member, IEEE,
and Hoi-Jun Yoo, Fellow, IEEE
Abstract—A 118.4 GB/s multi-casting network-on-chip
(MC-NoC) is proposed as communication platform for a real-time
object recognition processor. For application-specific NoC design, target traffic patterns are elaborately analyzed. Through
topology exploration, we derive a hierarchical star and ring
(HS-R) combined architecture for low latency and inter-processor
communication. Multi-casting protocol and router are developed
to accelerate one-to-many (1-to-N) data transactions. With these
two main features, the proposed MC-NoC reduces data transaction time and energy consumption for the target object recognition
traffic by 20% and 23%, respectively. The 350 k MC-NoC fabricated in a 0.13 m CMOS process consumes 48 mW at 400 MHz,
1.2 V.
Index Terms—Hierarchical star-ring combined topology, multicasting, network-on-chip, object recognition.
I. INTRODUCTION
S THE semiconductor technology scales down and the
number of intellectual properties (IPs) on a single chip
increases, communication among the IP blocks becomes a critical problem in system-on-chip (SoC) design. Conventional bus
based communication architecture cannot continue to scale due
to long signal propagation delay, lack of bandwidth, and large
power consumption [1]. As an alternative, network-on-chip
(NoC) that brings networking theories and methods to on-chip
communication becomes a promising architecture with better
performance and scalability than bus [1]–[4]. Recently, it is
quickly replacing the conventional bus based communication
platforms especially in multi-core based SoCs [5]–[7].
On the other hand, object recognition identifying database
objects out of the input image is an essential technology for
artificial vision applications such as face detection, auto-navigated vehicles, mobile intelligent robots, and surveillance
systems [8], [9]. Since these applications have to seamlessly
A
Manuscript received November 06, 2009; revised March 03, 2010; accepted
March 09, 2010. Current version published June 25, 2010. This paper was approved by Guest Editor Stefan Rusu.
The authors are with the Division of Electrical Engineering, Department of
Electrical Engineering and Computer Science, Korea Advanced Institute of
Science and Technology, Daejeon 305-701, Korea (e-mail: jooyoung.kim@
kaist.ac.kr).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/JSSC.2010.2048085
interact with output environments received from the video
camera, real-time object recognition over 30 frame/sec (fps) is
demanded. However, the object recognition process includes
lots of computationally intensive image processing tasks such
as image filtering, histogram, and database matching, so that
most of recognition processors [10]–[14] adopt a multi-core approach to meet their real-time requirement. As communication
architecture, they employ NoC in order to incorporate multiple
IP blocks as well as satisfy their data bandwidth requirements.
Among these NoCs, the two primitive ones [10], [11] just
provide a simple connectivity to their several IPs without any
consideration of specific data transactions. However, as the
applied recognition algorithms get complicated and the number
of IP blocks increases over ten [12]–[14], the processor’s
communication requirements becomes challengeable to satisfy.
Whether the NoC solves communication requirements of the
processor or not can affect the overall system performance.
The memory centric NoC [15] of 81.6GOPS object recognition
processor [12] supports data synchronization management and
maximum search function among the IP blocks for real-time
object recognition feature extraction based on scale invariant
feature transform (SIFT) algorithm [9]. As a result, the processor achieves 16 fps object feature extraction for QVGA
240) video images. In a real-time object recognition
(320
processor [13], the dynamically reconfigurable NoC [16] is
proposed for processor’s SIMD/MIMD mixed mode operation.
With a low latency crossbar switch containing image-express
channel, the proposed NoC enables 22 fps object recognition
for QVGA video images.
In this paper, we describe an application-specific NoC named
multi-casting NoC (MC-NoC) to enable 60 fps object recog480) screen resolution [14]. Through
nition at VGA (640
elaborate traffic analysis and topology exploration, we derive
its two main features: a hierarchical star and ring combined
topology and broad/multi-casting capability. The hierarchical
star and ring combined topology proposed for low latency data
transfer and fast inter-processor data communication results in
total 118.4 GB/s bandwidth with less than 3 switch hopping latency. For multi-casting implementation, a simple multi-casting
protocol is proposed with the designated source and destination
IPs while cost efficient network switch is developed for it. As a
result, the proposed MC-NoC accelerates data communication
by 20% as well as reduces energy consumption by 23% under
the target object recognition traffic.
0018-9200/$26.00 © 2010 IEEE
1400
IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 45, NO. 7, JULY 2010
Fig. 1. (a) Proposed object recognition processor. (b) Major data transactions.
The rest of the paper is organized as follows. Section II analyzes traffic patterns of the target object recognition processor to
derive its communication requirements. Based on these, Section
III proposes the MC-NoC’s hierarchical star and ring combined
architecture via topology exploration. Section IV describes its
wormhole routing based multi-casting protocol. Then, Section
V describes detailed component design of the MC-NoC. After
Section VI shows the chip implementation results, Section VI
summarizes the paper.
II. OBJECT RECOGNITION PROCESSOR ARCHITECTURE AND
DATA TRAFFIC ANALYSIS
Fig. 1 shows the block diagram of the proposed object
recognition processor [13] with its memory architecture and
major data transactions. The proposed processor is developed
with the target of 60 fps object recognition at VGA (640 480)
video resolution for mobile robot applications. It consists of 21
IP blocks operating at 200 MHz frequency: a neural perception
engine (NPE), a task manager (TM), 16 SIMD processor units
(SPUs), a decision processor (DP), and two external interfaces.
These 21 IPs are fully connected through a NoC whose operating frequency is twice the IP’s, 400 MHz. Different from
previous object recognition processors, the proposed processor
divides the object recognition into three functional stages
[17]: a visual perception (VP) computing regions-of-interest
(ROIs) out of an input image, a main image processing (MIP)
extracting local key-point vectors out of the tiled ROI area, and
a post database matching (PDM) performing matching of the
key-point vectors with a trained database for final decision.
The NPE, 16 SPUs, DP are proposed for the three stages,
respectively, while the TM is proposed to manage the ROI tasks
of the 16 SPUs. In this object recognition processor, distributed
memory architecture is adopted with two 200 MHz SRAM
devices in off-chip domain. The NPE containing instruction
and data cache initializes program codes of the other modules
and then, each module utilizes its own local memory for intermediate data store when it performs the assigned tasks. Because
every memory module is mapped to 32-bit address space in
the top level, each IP can access any other memory modules in
other IPs and two external ones through a fully interconnected
NoC. With stream processing model [18], each IP minimizes
external memory usage.
Fig. 1(b) shows major data transactions of the proposed processor. Three object recognition stages are mapped to a NPE, 16
SPUs, a DP, respectively, and are executed in the pipeline where
the pipelined data are ROI tiles and key-point vectors between
the first/second stage and second/third stage, respectively. The
NPE sends the information of ROI tiles to the message FIFO
in TM to make the TM assign them to 16 SPUs while the 16
SPUs store the resulting key-point vectors to the memory buffer
in the DP. Due to the pipelined operation, lots of data transactions seamlessly occur among the IP blocks throughout the
processing time. Under this condition, the NoC should provide
both sufficient data bandwidth and low latency transfer to enable
the processor’s target 60 fps processing at VGA video resolution. Major data transactions are as follows. At the VP stage,
the NPE distributes 0.8 Mb of instruction kernels to 16 SPUs.
The TM, in between the NPE and 16 SPUs, assigns the delivered ROI tiles to the 16 SPUs and manage their task threads.
In this process, the TM transfers 2 Mb of image pixel data to
16 SPUs each frame. In the MIP stage, the 16 SPUs perform
the feature description algorithm to generate key-point vectors
out of the selected ROI tiles. While running the algorithm, the
multiple SPUs communicate 200 Kb intermediate data themselves and transfer 100 Kb key-point vectors to the DP each
frame. In the PDM stage, the DP performs vector matching for
the aggregated key-point vectors with a trained object database.
From the three stages, each IP block requests 1 Mb raw image
and intermediate data transactions with external memories each
frame. Table I summarizes detailed traffic characteristics in the
proposed processor.
Based on the obtained traffic patterns, we derived the following five NoC requirements for real-time object recognition.
First of all, regardless of traffic patterns, low latency data
transaction is required for real-time processing. Because every
IP block of the proposed processor performs in-order processing, long data latency for read operations can seriously
affect the overall performance. Second, 1-to-N data transactions are frequently used from the NPE/TM to 16 SPUs, e.
g., when the NPE/TM distribute instruction kernels, ROI tile
tasks, and image data to 16 SPUs. To accelerate these data
KIM et al.: A 118.4 GB/s MULTI-CASTING NOC WITH HIERARCHICAL STAR-RING COMBINED TOPOLOGY FOR REAL-TIME OBJECT RECOGNITION
1401
TABLE I
DETAILED TRAFFIC CHARACTERISTICS
Fig. 2. Application specific NoC requirements.
transactions, broad- and multi-casting capability is demanded
in the NoC. Third, sufficient communication channels are
required among the SPUs for their co-processing. To this end,
a suitable NoC topology should be investigated. Fourth, N-to-1
data transactions are required for the aggregation of vector
packets from 16 SPUs to a DP. An additional output port is
needed to DP to handle this problem. Lastly, the NoC should
provide a multi-clocking environment for the overall SoC to
allow each IP to use a preferable clock frequency. To this end,
every interfacing region between the NoC and the IP blocks
should have data synchronization function to traverse between
heterogeneous clock domains. Fig. 2 summarizes the five
application-specific NoC requirements of the proposed object
recognition processor.
III. HIERARCHICAL STAR AND RING COMBINED TOPOLOGY
A. Topology Exploration
In the design of a NoC, topology selection deciding the structural organization of network switches is a primary concern. By
the topology, not only the performance of NoC such as average
latency and total bandwidth but also the implementation cost is
roughly determined. To find out the most suitable NoC topology
for the target object recognition, we choose four topologies depicted in Fig. 3: mesh [5], [6], hierarchical star [3], [4], [19],
ring [7], and proposed hierarchical star-ring (HS-R) combined
topology. The mesh topology has strength in regularity and scalability, but it is not good in terms of cost efficiency because each
IP requires its own network switch. On the other hand, the hierarchical star topology is a cost efficient solution based on its tree
structure that results in low latency data communication. However, the problem is that significant traffic workloads aggregate
to network switches in higher level. The ring topology also has
good cost efficiency, however, its long hopping latency cannot
continue scaling to many-core era containing tens of IPs. The
proposed HS-R topology combines the hierarchical star and ring
topology for short latency data transmission and inter-processor
data communication, respectively. Based on the hierarchical star
networks, it adds ring networks to achieve shorter hopping latency as well as to mitigate workload aggregation to higher level
network switches. Fig. 4 shows the block diagram of the conventional hierarchical star and proposed HS-R topology where
the number of IPs and network hierarchies scale to 64 and 3,
respectively. Different from the hierarchical star topology, the
HS-R topology additionally interconnects the switches in the
same network hierarchy with the ring topology. In this diagram,
the number in each IP box and grouped IPs by the distance level
means the switch hopping latency from the start IP and the set
of IPs having the same hopping latency from the start IP , respectively. The additional ring networks of the HS-R topology
reduce the hopping latency from the start IP by 1 when it accesses the IPs belongs to the neighbor clusters of each same distance group. As shown in Fig. 4, half of the overall IPs colored
in dark gray are benefited by the HS-R topology.
To evaluate the performance of above four topologies under
target application traffic, we made a cycle accurate NoC simulator broadly consisting of traffic model library and NoC model
library, as shown in Fig. 5. The traffic model library contains
synthetic traffic patterns and traffic generating subjects such
as real IP’s instruction set simulator and traffic generator. The
traffic generator can generate some typical traffic patterns as
well as arbitrary ones by reading a traffic description file (TDF)
that compiles the time and data of each data transaction event.
On the other hand, the NoC model library contains building
components of NoC such as a network switch, communication
socket, and network wrapper. We use transaction level modeling (TLM) [20] in the NoC model library. Because the TLM
operates based on the existence of transactions, it is good to
integrate various IP blocks modeled with different abstraction
levels. In addition, its event driven characteristic is beneficial
1402
IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 45, NO. 7, JULY 2010
Fig. 3. NoC topologies. (a) Mesh; (b) H-star; (c) ring; (d) H-star + ring.
Fig. 4. Comparison between hierarchical star and proposed hierarchical star–ring combined topology.
Fig. 5. NoC simulation framework.
for fast simulation time. A monitoring unit is included in each
network packet and switch model for packet tracing and statistic
extractions. Using these two model libraries, we can constitute
a cycle accurate NoC simulation framework. The traffic generating IP blocks are connected to the NoC using network wrapper
and TLM socket interface, while the NoC topology is built by a
link description file (LDF) that describes the connection status
of whole network switches.
Using the implemented NoC framework, we measured the
average latency, bandwidth utilization, and implementation
cost of the four NoC topologies when the target three-stage
object recognition algorithm runs for several VGA (640 480
pixels) input images. Average 56.49 s is spent by the simulator
to process a single VGA image. In this experiment, cycle
accurate models are used for the TM, 16 SPUs, and DP while
a functional model is used for the NPE. The bandwidth utilization is calculated by dividing the overall amount of packet data
through every network switch by the maximum total bandwidth
of every switch. The cost of the NoC topology is measured by
the amount of employed network switches and interconnecting
wires. Assuming every network switch is square (same input
and output port number, e. g., 5 5, 6 6, and 7 7), the
implementation costs of a network switch itself and the wires
from it are proportional to the number of input ports. Therefore,
the implementation cost of the overall NoC can be represented
by following equation where , , means the cost of a 1
1 sized switch, the cost of a single wire in average length, and
the data bit-width of each port.
Fig. 6 shows the experimental results. The implementation
cost is represented by the total number of input ports from all
switches, a normalized value with the unit cost of a switch
and wires per port. For latency, the hierarchical star and HS-R
topology show better performances than the mesh and ring
topology because their tree-like structure logarithmically reduces the switch hopping number to access other IP blocks. The
HS-R topology further reduces the latency from the hierarchical
star topology by accelerating inter-processor communication
via supplemented ring networks. For bandwidth utilization,
KIM et al.: A 118.4 GB/s MULTI-CASTING NOC WITH HIERARCHICAL STAR-RING COMBINED TOPOLOGY FOR REAL-TIME OBJECT RECOGNITION
1403
Fig. 6. Topology exploration results.
Fig. 7. MC-NoC architecture.
the mesh topology is the lowest among the four topologies
because it has the largest maximum total bandwidth due to its
abundant network switches and links. However they also bring
the most expensive implementation cost to mesh topology.
For implementation cost, the hierarchical star and HS-R show
good efficiency. In conclusion, mesh is not a suitable topology
for the target traffic because it does not show better latency
performance than the others even it requires larger implementation cost. The ring topology seems to be the worst among the
four topologies. This is because its expected switch hopping
number becomes too large in these 21 IP SoC. For smaller sized
SoCs, the ring topology can be a good solution with its low
cost interconnection. We can decide that the proposed HS-R
topology is the most suitable topology for the target traffic with
the lowest data latency and supplemented data bandwidth for
inter-processor communication.
B. MC-NoC Architecture
Fig. 7 shows the overall architecture of the proposed
MC-NoC consisting of a 9
10 system network switch and
four 7 7 local network switches. The 16 SPUs are connected
to the system network through the four local network switches
while the rest of IP blocks such as the NPE, TM, DP, and two
external interfaces are directly connected to the system network switch. As decided in the previous section, the MC-NoC
exploits the HS-R combined topology. It adopts a hierarchical
star topology as a basic starting topology for low latency data
communication, and then, supplements a ring topology to the
local networks for high speed inter-SPU data transactions. The
additional ring networks of the combined topology provide a
GB/s links
input ports
maximum bandwidth of
output ports
byte
MHz, assuming all physical
channels transfer data) between the local networks, and allow
each SPU to access the other SPUs in neighbor local networks
in two switch hops without accessing the system network.
In addition, for N-to-1 data transactions, the output port of
the DP is modified to a dual port. One of them is dedicated
to key-point vector packet aggregation from the 16 SPUs.
In overall, the topology combined MC-NoC provides overall
GB/s
input ports output ports
input ports
output ports
byte
MHz total bandwidth with the
switch hopping latency of less than 3.
IV. MULTI-CAST PROTOCOL
Multi-casting transmitting the same packet to multiple destinations [21] is an effect method when a single source wants
to send the identical packets to multiple destinations. It is also
very useful for the 1-to-N data requirements of the proposed
processor such as program code distribution and image data
download. However, multi-casting operation in NoC can cost a
lot because protocol, routing, and network switch design should
be redesigned. In this chapter, we present a cost efficient multicasting protocol and implementation in the proposed MC-NoC
based on the fact that the multi-casting source and destination
IP are already determined.
Fig. 8 shows the block diagram of the network switch consisting of input ports containing queuing buffer, output arbiters,
crossbar switch, and output ports. It uses a wormhole routing
protocol that controls the packet operation in a smaller flow control unit (flits) to reduce the size of input buffer. To achieve high
throughput, the network switch performs a packet transmission
with four pipeline stages. First, the incoming packet is buffered
at the input port in a flit unit. Then, each input port sends request and obtain grant to/from the output arbiter. After getting
a grant, the packet traverses the crossbar switch. To prevent the
packet loss, the network switch uses a credit based back pressure
scheme [22] that controls the flits in the router by the congested
1404
IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 45, NO. 7, JULY 2010
Fig. 8. Network switch block diagram.
Fig. 9. Packet format of MC-NoC.
status of receiver side. For multi-casting operation, it uses a synchronous method [23] that transmits the packet to the multiple
output ports at the same time. By eliminating packet retransmission, it can achieve simpler routing algorithm as well as higher
energy efficiency than asynchronous method.
Fig. 9 shows the detailed packet format of the MC-NoC. It
consists of three flit types—header, address, and data—while
each contains 34-bit information including 2-bit type identifier.
The flit is efficiently controlled by 2-bit flow control signals,
valid and back pressure hold signal.
The header flit contains all information of the packet such
as read/write, 4-bit burst length, 2-bit priority level, and 16-bit
routing information (RI). The address flit indicates the address
of the final destination IP where the data flits are read or written.
When the packet is routed in each network switch, the header filt
forwards first and the address/data flits follow after it. The burst
length information enables burst data transactions up to 8 data
flits. For this, the write burst packet sends header, address, and
multiple data flits of burst length to the target IP while the read
burst packet sends only header and address flit to get multiple
data flits of burst length from the target IP. The 2-bit priority
level indicates the priority of packet in the situation that the
network has contentions. By obtaining the arbiter’s grant earlier in each network switch, the packet with the higher priority
level is more quickly routed than the packet with the normal
priority level. Lastly, the 16-bit source defined routing information (RI) describes where the packet should be routed in each
network switch to the final destination. It allows 4 switch traversals for normal packets and a multi-casting to arbitrary SPUs
for multi-casting packets. For multi-casting case, we use a bit
string encoding [24] to notify the destination SPUs. In this encoding, each bit of 16-bit RI corresponds to each SPU and configuring a bit to 1 means the SPU is one of the multi-casting destinations. Without describing detailed routing path, the 16-bit
RI efficiently implicates the multi-casting destinations out of 16
SPUs. Each network switch decodes the 16-bit multi-casting RI
to generate correct routing paths to multiple destinations.
Fig. 10 shows the decoding process of a multi-casting packet
in the MC-NoC. With a bit string encoding, the example RI
intends to send a packet to SPU 0, 2, 3, 13, and 15. To this
end, the MC-NoC adopts a hierarchical approach using its
hierarchical-star topology: the system network performs the
first multi-casting to four SPU local networks, and then, local
network perform the second multi-casting to final destination
KIM et al.: A 118.4 GB/s MULTI-CASTING NOC WITH HIERARCHICAL STAR-RING COMBINED TOPOLOGY FOR REAL-TIME OBJECT RECOGNITION
1405
Fig. 10. Multi-casting process of MC-NoC.
Fig. 11. Four-stage pipelined multi-casting network switch.
SPUs. In the first multi-casting, the system network sends the
packet to the local networks containing at least one destination
SPU. The destination local networks can be identified by
performing OR operation on each of four 4-bit RI segments. In
the second multi-casting, each destination local network sends
a packet to destination SPUs according to corresponding 4-bit
RI segment. Through this two step multi-casting, the packet
correctly arrives at the destination SPUs. Different from the
mesh based NoC should consider the optimized multi-cast
routes out of lots of possible candidates, the hierarchical star
based MC-NoC can perform the hierarchical and deterministic
multi-cast routing because its upper level network switches
interconnect all of the lower level network switches. This
hierarchical multi-casting method can be expanded to larger
MC-NoCs having hierarchies more than two.
V. DETAILED MC-NOC DESIGN
In this Section, we discuss real implementation of MC-NoC.
The MC-NoC has three design issues to address. First, each
network switch should be able to perform multi-casting for
overall multi-casting capability. Second, the interface between
the MC-NoC and IP blocks can convert the heterogeneous
clock domains for multi-clocking SoC platform. Third, the
MC-NoC should operate with low power consumption for low
overhead interconnection platform.
A. Multi-Casting Network Switch
Fig. 11 depicts the block diagram of multi-casting network
switch composed of input ports, output arbiters, crossbar fabric,
and output ports. It performs packet transmission with four
pipelined stages. First, the incoming flits are buffered at 8-depth
32-bit first-in-first-out (FIFO) queue containing synchronization interface for the two heterogeneous clock domains. Then,
each activated input port sends a request signal to the output
arbiter to get a grant signal to traverse the crossbar fabric.
The arbiters perform round-robin scheduling with two priority
levels. After getting a grant signal, the input port sends the
packet to the destination output port through a fully connected
crossbar fabric. In case of multi-casting, the input port sends
multiple request signals to multiple output arbiters to ensure
1406
IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 45, NO. 7, JULY 2010
Fig. 12. Multi-casting port.
of the processor operation and the TM performs multi-casting
the image data to two SPUs, above conflicts in multi-casting
rarely happen and barely affect the overall performance.
To provide sufficient driving strength for variable loads in
multi-casting, we specially devise a variable strength driver
whose circuit diagram is shown in Fig. 13. It consists of three
parallel tri-state buffers whose driving forces are different to
2 , 4 , and 8 , respectively, while their output wires are
connected into a single final output to enhance the overall
circuits output driving strength. Using 3-bit control signals that
each of them decides the activation of each tri-state buffer, the
variable strength driver can dynamically configure its output
driving strength from 2 to 14 by a 2 step according to the
number of output ports of multi-casting. The table in Fig. 13
shows the operation modes of the variable strength driver by
the control signals and its energy consumptions according to
various multi-casting output loads.
Fig. 13. Circuit diagram of variable strength driver.
all grant signals of desired output ports. The multi-casting
input port waits until all grant signals are returned. For this
operation, the multi-casting input port employs a multi-port
requester and grant checker shown in Fig. 12. The multi-port
requester decodes the 16-bit routing information, generates
request signals to multiple arbiters, and latches the request
signals in a 16-bit register. Meanwhile, the grant checker holds
the multi-casting packet in the FIFO queue until the received
grant signals become equal to the latched request ones. After
gathering all grant signals, the multi-casting is performed
using broad-casted wires of crossbar fabric. In case of arbitrary
destinations are blocked by some reasons, the multi-casting
port should select whether it hold the grants until to gather all
grants or withdraw the request and request later. In this design,
the multi-casting request is withdrawn when the other inputs
send requests to the output arbiters whose grants are already
captured by the multi-casting port which is waiting remaining
grants, not performing multi-casting. In the proposed processor
because the multi-casting is used in cases that the NPE performs
multi/broad-casting program code to 16 SPUs at the beginning
B. FIFO Based Synchronization Interface
To allow the heterogeneous clock frequencies between the IPs
and MC-NoC, the first-in-first-out (FIFO) based synchronization is performed in every input/output port of network switches
interfacing with IP blocks. Fig. 14 shows the block diagram of
FIFO based synchronization interface that exists in each input
port. It consists of an 8-depth FIFO buffer, write control block,
and read control block, which are for data storage, write and
read operation control and pointer management, respectively. A
flag detector, in between the write and read control block, generates combinational flag signals such as empty and full using
relative locations of the two pointers. When the detected empty
or full signal is used in the other clock domain for flow control signal generation like request or hold, they are synchronized
with the clock of the other domain using a pipelined synchronization method [25].
C. Fine-Grained Clock Gating
For low power consumption, the MC-NoC performs a
fine-grained clock gating that activates the only required packet
routing path in each network switch. Fig. 15 shows the block
diagram of fine-grained clock gating and its operating timing
diagram. The monitor unit in each network switch detects the
KIM et al.: A 118.4 GB/s MULTI-CASTING NOC WITH HIERARCHICAL STAR-RING COMBINED TOPOLOGY FOR REAL-TIME OBJECT RECOGNITION
1407
TABLE II
CHIP SUMMARY
Fig. 14. Heterogeneous clock interface.
existence of incoming flits to activate corresponding input
queuing buffers. In addition, it partially decodes the header flit
to generate the list of output ports to activate the output arbiters
and crossbar fabric. Because the proposed MC-NoC supports
multi-casting operation, the destination output ports can be
multiple. With a simple D-flip-flop chain, the activation time of
each switch component such as input queues, output arbiters,
and crossbar fabric is determined in a pipelined manner. The
activated components operate until the end of a single packet
transmission while the monitor unit is always turned on. By
controlling the activation of each switch component in cycle
level including the input buffers consuming almost 90% of
network switch power, the fine-grained clock gating can save
28% power on average.
Fig. 15. Fine grained clock gating.
VI. IMPLEMENTATION RESULTS
To evaluate the performance of the MC-NoC, we measured
the cycle count and energy consumption from gate-level simulation for the all packets transmitted during the target object
recognition algorithm is running for a single VGA (640 480)
video image. We also performed the identical experiments on
a basic hierarchical star NoC without any scheme and a HS-R
combined NoC without multi-casting capability for comparison. Fig. 16 shows the cycle count and energy reduction effect
by the two main features of the MC-NoC. The HS-R combined
topology contributes to inter-SPU data communication. By
using the supplemented ring networks in inter-SPU communication, it shortens the routing path and reduces the number
1408
IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 45, NO. 7, JULY 2010
Fig. 16. MC-NoC evaluation results. (a) Cycle count. (b) Energy consumption.
The overall processor consumes 496 mW at 1.2 V while the
object recognition is running at 60 frame/sec for VGA video
input. Under this condition, the MC-NoC dissipates 48 mW,
only 9.7% of the overall processor. Table II summarizes the
chip and MC-NoC features.
VII. CONCLUSIONS
Fig. 17. Chip micrograph.
of packet buffering. The multi-casting capability contributes
to program distribution and image download. By removing
retransmissions of the same packet to multiple IPs, it not only
significantly reduces the number of packet buffering but also accelerates the data transactions as well. As a result, the MC-NoC
reduces the overall cycle count and energy consumption in data
communication by 20% and 23%, respectively, compared with
a basic hierarchical-star NoC.
The real-time object recognition processor with the proposed
MC-NoC is fabricated in a 0.13 m eight-metal CMOS process.
Fig. 17 shows the chip micrograph. The overall recognition
processor has 49 mm die area with 3.73 M gates while the
MC-NoC accounts for 350 k gates. The operating frequency
is 200 MHz for IP blocks and 400 MHz for MC-NoC. The
MC-NoC provides 118.4 GB/s total bandwidth (92.8 GB/s
from hierarchical star networks, 25.6 GB/s from ring networks)
at 400 MHz frequency and supports multi-casting to 16 SPUs.
A 118.4 GB/s multi-casting network-on-chip (MC-NoC) is
developed as communication platform for a real-time object
recognition processor. Through topology exploration, we derive the HS-R combined topology for low latency and inter-processor data communications. Multi-casting is supported to accelerate 1-to-N data transactions. With these two main features,
the proposed MC-NoC reduces data transaction time and energy
consumption by 20% and 23%, respectively, under the target object recognition traffic. The MC-NoC is fabricated in a 0.13 m
CMOS process in the proposed object recognition processor
while consuming 48 mW at 400 MHz, 1.2 V.
REFERENCES
[1] L. Benini and G. D. Micheli, “Networks on chips: A new SoC paradigm,” IEEE Computer, vol. 35, no. 1, pp. 70–78, Jan. 2002.
[2] W. J. Dally and B. Towles, “Route packets, not wires: On-chip interconnection networks,” in Proc. 38th Design Automation Conf., Jun.
2001, pp. 684–689.
[3] S.-J. Lee et al., “An 800 MHz star-connected on-chip network for application to systems on a chip,” in IEEE Int. Solid-State Circuits Conf.
(ISSCC) Dig. Tech. Papers, Feb. 2003, pp. 468–469.
[4] K. Lee et al., “A 51 mW 1.6 GHz on-chip network for low-power
heterogeneous SoC platform,” in IEEE Int. Solid-State Circuits Conf.
(ISSCC) Dig. Tech. Papers, Feb. 2004, pp. 152–153.
[5] S. Vangal et al., “An 80-tile 1.28TFLOPS network-on-chip in 65 nm
CMOS,” in IEEE Int. Solid-State Circuits Conf. (ISSCC) Dig. Tech.
Papers, Feb. 2007, pp. 98–99.
[6] S. Bell et al., “TILE64TM processor: A 64-core SoC with mesh interconnect,” in IEEE Int. Solid-State Circuits Conf. (ISSCC) Dig. Tech.
Papers, Feb. 2008, pp. 88–89.
[7] T. W. Ainsworth and T. M. Pinkston, “Characterizing the cell EIB
on-chip network,” IEEE Micro, vol. 27, no. 5, pp. 6–14, 2007.
[8] D. A. Forsyth and J. Ponce, Computer Vision: A Modern Approach.
New York: Prentice Hall, 2002.
KIM et al.: A 118.4 GB/s MULTI-CASTING NOC WITH HIERARCHICAL STAR-RING COMBINED TOPOLOGY FOR REAL-TIME OBJECT RECOGNITION
[9] D. G. Lowe, “Distinctive image features from scale-invariant keypoints,” ACM Int. J. Comput. Vis., vol. 60, no. 2, pp. 91–110, Jan.
2004.
[10] R. R. Rojas et al., “Object recognition SoC using the support vector
machines,” J. Appl. Signal Process., pp. 993–1004, 2005.
[11] H.-C. Lai et al., “Communication-aware face detection using NoC architecture,” in Proc. Int. Conf. Computer Vision Systems (ICVS), 2008,
pp. 181–189.
[12] D. Kim et al., “An 81.6 GOPS object recognition processor based on
NoC and visual image processing memory,” in Proc. IEEE Custom Integrated Circuits Conf. (CICC), 2007, pp. 443–446.
[13] K. Kim et al., “A 125GOPS 583 mW network-on-chip based parallel
processor with bio-inspired visual attention engine,” IEEE J. SolidState Circuits, vol. 44, no. 1, pp. 136–147, Jan. 2009.
[14] J.-Y. Kim et al., “A 201.4GOPS 496 mW real-time multi-object recognition processor with bio-inspired neural perception engine,” in IEEE
ISSCC Dig. Tech. Papers, Feb. 2009, pp. 150–151.
[15] D. Kim et al., “Implementation of memory-centric NoC for 81.6 GOPS
object recognition processor,” in Proc. IEEE A-SSCC, Nov. 2007, pp.
47–50.
[16] K. Kim et al., “A 76.8 GB/s 46 mW low-latency network-on-chip for
real-time object recognition processor,” in Proc. IEEE A-SSCC, Nov.
2008, pp. 189–192.
[17] J.-Y. Kim et al., “Real-time object recognition with neuro-fuzzy controlled workload-aware task pipelining,” IEEE Micro, vol. 29, no. 6,
pp. 28–43, Nov./Dec. 2009.
[18] W. J. Dally et al., “Stream processors: Programmability with efficiency,” ACM Queue, vol. 2, no. 1, pp. 52–62, 2004.
[19] K. Lee et al., “Low-power networks-on-chip for high-performance SoC
design,” IEEE Trans. VLSI, vol. 14, no. 2, pp. 148–160, Feb. 2006.
[20] F. Ghenassia, Transaction Level Modeling With SystemC: TLM Concepts and Applications for Embedded Systems. New York: Springer,
2005.
[21] L. L. Peterson and B. S. Davie, Computer Networks: A Systems Approach, ser. The Morgan Kaufmann Series in Networking. : Morgan
Kaufmann, 2007.
[22] H.-J. Yoo, K. Lee, and J. K. Kim, Low Power NoC for High Performance SoC Design. Boca Raton: CRC Press, 2008.
[23] C. M. Chiang and L. M. Ni, “Deadlock-free multi-head wormhole
routing,” in Proc. First High Performance Computing–Asia, Taiwan,
Taiwan, Sep. 1995.
[24] C. Chiang and L. Ni, “Multi-address encoding for multicast,” in Proc.
1st Int. Workshop on Parallel Computer Routing and Communication,
PCRCW’94, Seattle, WA, May 1994, pp. 146–160.
[25] J. N. Seizovic, “Pipeline synchronization,” in Proc. IEEE ASYNC, Nov.
1994, pp. 87–96.
Joo-Young Kim (S’05) received the B.S., M.S., and
Ph.D. degrees in electrical engineering and computer
science from the Korea Advanced Institute of Science
and Technology (KAIST), Daejeon, Korea, in 2005,
2007, and 2010, respectively.
Since 2006, he has been involved with the development of the parallel processors for computer
vision. Currently, he is a postdoctoral researcher at
KAIST. His research interests include multi-core
architecture, network-on-chip, and VLSI implementations for computer vision applications.
Junyoung Park (S’09) received the B.S. degrees in
electrical engineering and computer science from the
Korea Advanced Institute of Science and Technology
(KAIST), Daejeon, Korea, in 2009 and is currently
working toward the M.S. degree in electrical engineering and computer science at KAIST.
Since 2009, he has been involved with the development of the parallel processors for computer vision.
Currently, his research interests are many-core architecture and VLSI implementation for bio-inspired vision processor.
1409
Seungjin Lee (S’06) received the B.S. and M.S. degrees in electrical engineering and computer science
from the Korea Advanced Institute of Science and
Technology (KAIST), Daejeon, Korea, in 2006 and
2008, respectively. He is currently working toward
the Ph.D. degree in electrical engineering and computer science from KAIST.
His previous research interests include low power
digital signal processors for digital hearing aids and
body area communication. Currently, he is investigating parallel architectures for computer vision processing.
Minsu Kim (S’07) received the B.S. and M.S. degrees in electrical engineering and computer science
from the Korea Advanced Institute of Science and
Technology (KAIST), Daejeon, Korea, in 2007 and
2009, respectively. He is currently working toward
the Ph.D. degree in electrical engineering and computer science at KAIST.
His research interests include network-on-chip
based SoC design and bio-inspired VLSI architecture
for intelligent vision processing.
Jinwook Oh (S’08) received the B.S. degree in electrical engineering and computer science from Seoul
National University, Seoul, Korea, in 2008. He is currently working toward the M.S. degree in electrical
engineering and computer science at KAIST, Daejeon, Korea.
His research interests include low-power digital
signal processors for computer vision. Recently, he
has been involved with the VLSI implementation of
neural networks and fuzzy logics.
Hoi-Jun Yoo (M’95–SM’04–F’08) graduated from
the Electronic Department of Seoul National University, Seoul, Korea, in 1983 and received the M.S.
and Ph.D. degrees in electrical engineering from the
Korea Advanced Institute of Science and Technology
(KAIST), Daejeon, in 1985 and 1988, respectively.
His Ph.D. work concerned the fabrication process
for GaAs vertical optoelectronic integrated circuits.
From 1988 to 1990, he was with Bell Communications Research, Red Bank, NJ, where he invented
the two-dimensional phase-locked VCSEL array, the
front-surface-emitting laser, and the high-speed lateral HBT. In 1991, he became Manager of a DRAM design group at Hyundai Electronics and designed
a family of from fast 1 M DRAMs and 256 M synchronous DRAMs. In 1998,
he joined the faculty of the Department of Electrical Engineering at KAIST and
now is a full Professor. From 2001 to 2005, he was the Director of System Integration and IP Authoring Research Center (SIPAC), funded by Korean government to promote worldwide IP authoring and its SOC application. From 2003 to
2005, he was the full time Advisor to Minister of Korea Ministry of Information
and Communication and National Project Manager for SoC and Computer. In
2007, he founded SDIA(System Design Innovation and Application Research
Center) at KAIST to research and develop SoCs for intelligent robots, wearable
computers and bio systems. His current interests are high-speed and low-power
network-on-chips, 3-D graphics, body area networks, biomedical devices and
circuits, and memory circuits and systems. He is the author of the books DRAM
Design (Seoul, Korea: Hongleung, 1996; in Korean), High Performance DRAM
(Seoul, Korea: Sigma, 1999; in Korean), and chapters of Networks on Chips
(New York, Morgan Kaufmann, 2006).
Dr. Yoo received the Electronic Industrial Association of Korea Award for
his contribution to DRAM technology the 1994, Hynix Development Award in
1995, the Korea Semiconductor Industry Association Award in 2002, Best Research of KAIST Award in 2007, Design Award of 2001 ASP-DAC, and Outstanding Design Awards of 2005, 2006, 2007 A-SSCC. He is a member of the
executive committee of ISSCC, Symposium on VLSI Circuits, and A-SSCC.
He is the TPC chair of the A-SSCC 2008.
0
Purchase answer to see full
attachment
![DescriptionHW Assignment 4
General policy for all assignments:
• Late or unsuccessful submission of the assignment is 0 credit on the
assignment. This policy is strictly applied. Submit your assignment
in PDF format on the specified due date in course schedule, and
before the deadline that is 11:30 PM. This is an individual
assignment. Sharing your solution, copying from others, or copypasting from lecture slides is considered plagiarism. The penalty in
the course policy for such act is 0 for the assignment in the first
time, and F in the course for the second time. Submit your answer to
BB on or before the due date.
•
Please remember that late submission or failing to submit the
assignment has 0 credit.
•
You may not share your answers with other students.
Part I
1- (10 points) Define the following concepts:
• Ripping
• Frame-rate
• Color-depth
• Bitmap resolution
• Sampling rate
2- (10 points) How much space is required to store 1 hour of fidelity music if the
height of each high-fidelity sound sample is 16 bits, and the sampling rate is
44.1 kHz? Show how you calculated the answer.
3- (10 points) In a two-column table, indicate the differences between bitmap and
vector graphics.
4- (5 points) Pick a color in RGB color system and represent the color in decimal,
binary, and hexadecimal. For instance, a color with decimal representation
(0,255,255) has hex representation (00FFFF) and binary representation (0000
0000 1111 1111 1111 1111).
5- (5 points) Explain how many bits are used to represent a color in a True-color
format? How many different color values are represented in True-Color format.
Include your calculation.
1
6- (10 points) What is the storage needed to store one video frame with 720×486
resolution in True-color format? If frame rate is 32 fps, how much data is
required for one minute of the video?
7- (10 points) Show the RLE for the following image data. Remember that pixels
are scanned in a zigzag fashion.
• (5 points) First show the scanned values in a vector/array.
• (5 points) Then provide the RLE.
9
9
3
0
66
5
5
6
9
9
2
0
66
5
5
6
9
9
3
255
66
5
5
6
9
1
3
255
66
5
6
6
9
1
3
255
255
6
6
6
8
1
2
0
255
6
6
6
8
1
2
0
255
6
6
6
8
1
2
0
255
6
6
6
8- (5 points) Explain the difference between lossy and lossless image
compression with examples of image file format that suppport each category.
9- (5 points) what is the difference between speech synthesis and voice
recognition technology? What type of audio format supports speech synthesis?
Part II
10- (10 points) Download the following paper from Blackboard first (“paper1 for
assignment 4”),
Joo-Young Kim, Junyoung Park, Seungjin Lee, Minsu Kim, Jinwook Oh, Hoi-Jun
Yoo:
A 118.4 GB/s Multi-Casting Network-on-Chip With Hierarchical Star-Ring
Combined Topology for Real-Time Object Recognition. J. Solid-State
Circuits 45(7): 1399-1409 (2010)
Next, rank the references 1-5 in this paper (from high-to-low) according to
Google scholar. Use the following format:
[Ref#], [Title of the paper] , [Citation number in Google Scholar]
11- (20 points) Use DBLP computer science bibliography website to list all of the
publishers (conferences/workshops/journals) the authors of the given paper
have published to (e.g., IEEE Micro).
2
IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 45, NO. 7, JULY 2010
1399
A 118.4 GB/s Multi-Casting Network-on-Chip
With Hierarchical Star-Ring Combined Topology
for Real-Time Object Recognition
Joo-Young Kim, Student Member, IEEE, Junyoung Park, Student Member, IEEE,
Seungjin Lee, Student Member, IEEE, Minsu Kim, Student Member, IEEE, Jinwook Oh, Student Member, IEEE,
and Hoi-Jun Yoo, Fellow, IEEE
Abstract—A 118.4 GB/s multi-casting network-on-chip
(MC-NoC) is proposed as communication platform for a real-time
object recognition processor. For application-specific NoC design, target traffic patterns are elaborately analyzed. Through
topology exploration, we derive a hierarchical star and ring
(HS-R) combined architecture for low latency and inter-processor
communication. Multi-casting protocol and router are developed
to accelerate one-to-many (1-to-N) data transactions. With these
two main features, the proposed MC-NoC reduces data transaction time and energy consumption for the target object recognition
traffic by 20% and 23%, respectively. The 350 k MC-NoC fabricated in a 0.13 m CMOS process consumes 48 mW at 400 MHz,
1.2 V.
Index Terms—Hierarchical star-ring combined topology, multicasting, network-on-chip, object recognition.
I. INTRODUCTION
S THE semiconductor technology scales down and the
number of intellectual properties (IPs) on a single chip
increases, communication among the IP blocks becomes a critical problem in system-on-chip (SoC) design. Conventional bus
based communication architecture cannot continue to scale due
to long signal propagation delay, lack of bandwidth, and large
power consumption [1]. As an alternative, network-on-chip
(NoC) that brings networking theories and methods to on-chip
communication becomes a promising architecture with better
performance and scalability than bus [1]–[4]. Recently, it is
quickly replacing the conventional bus based communication
platforms especially in multi-core based SoCs [5]–[7].
On the other hand, object recognition identifying database
objects out of the input image is an essential technology for
artificial vision applications such as face detection, auto-navigated vehicles, mobile intelligent robots, and surveillance
systems [8], [9]. Since these applications have to seamlessly
A
Manuscript received November 06, 2009; revised March 03, 2010; accepted
March 09, 2010. Current version published June 25, 2010. This paper was approved by Guest Editor Stefan Rusu.
The authors are with the Division of Electrical Engineering, Department of
Electrical Engineering and Computer Science, Korea Advanced Institute of
Science and Technology, Daejeon 305-701, Korea (e-mail: jooyoung.kim@
kaist.ac.kr).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/JSSC.2010.2048085
interact with output environments received from the video
camera, real-time object recognition over 30 frame/sec (fps) is
demanded. However, the object recognition process includes
lots of computationally intensive image processing tasks such
as image filtering, histogram, and database matching, so that
most of recognition processors [10]–[14] adopt a multi-core approach to meet their real-time requirement. As communication
architecture, they employ NoC in order to incorporate multiple
IP blocks as well as satisfy their data bandwidth requirements.
Among these NoCs, the two primitive ones [10], [11] just
provide a simple connectivity to their several IPs without any
consideration of specific data transactions. However, as the
applied recognition algorithms get complicated and the number
of IP blocks increases over ten [12]–[14], the processor’s
communication requirements becomes challengeable to satisfy.
Whether the NoC solves communication requirements of the
processor or not can affect the overall system performance.
The memory centric NoC [15] of 81.6GOPS object recognition
processor [12] supports data synchronization management and
maximum search function among the IP blocks for real-time
object recognition feature extraction based on scale invariant
feature transform (SIFT) algorithm [9]. As a result, the processor achieves 16 fps object feature extraction for QVGA
240) video images. In a real-time object recognition
(320
processor [13], the dynamically reconfigurable NoC [16] is
proposed for processor’s SIMD/MIMD mixed mode operation.
With a low latency crossbar switch containing image-express
channel, the proposed NoC enables 22 fps object recognition
for QVGA video images.
In this paper, we describe an application-specific NoC named
multi-casting NoC (MC-NoC) to enable 60 fps object recog480) screen resolution [14]. Through
nition at VGA (640
elaborate traffic analysis and topology exploration, we derive
its two main features: a hierarchical star and ring combined
topology and broad/multi-casting capability. The hierarchical
star and ring combined topology proposed for low latency data
transfer and fast inter-processor data communication results in
total 118.4 GB/s bandwidth with less than 3 switch hopping latency. For multi-casting implementation, a simple multi-casting
protocol is proposed with the designated source and destination
IPs while cost efficient network switch is developed for it. As a
result, the proposed MC-NoC accelerates data communication
by 20% as well as reduces energy consumption by 23% under
the target object recognition traffic.
0018-9200/$26.00 © 2010 IEEE
1400
IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 45, NO. 7, JULY 2010
Fig. 1. (a) Proposed object recognition processor. (b) Major data transactions.
The rest of the paper is organized as follows. Section II analyzes traffic patterns of the target object recognition processor to
derive its communication requirements. Based on these, Section
III proposes the MC-NoC’s hierarchical star and ring combined
architecture via topology exploration. Section IV describes its
wormhole routing based multi-casting protocol. Then, Section
V describes detailed component design of the MC-NoC. After
Section VI shows the chip implementation results, Section VI
summarizes the paper.
II. OBJECT RECOGNITION PROCESSOR ARCHITECTURE AND
DATA TRAFFIC ANALYSIS
Fig. 1 shows the block diagram of the proposed object
recognition processor [13] with its memory architecture and
major data transactions. The proposed processor is developed
with the target of 60 fps object recognition at VGA (640 480)
video resolution for mobile robot applications. It consists of 21
IP blocks operating at 200 MHz frequency: a neural perception
engine (NPE), a task manager (TM), 16 SIMD processor units
(SPUs), a decision processor (DP), and two external interfaces.
These 21 IPs are fully connected through a NoC whose operating frequency is twice the IP’s, 400 MHz. Different from
previous object recognition processors, the proposed processor
divides the object recognition into three functional stages
[17]: a visual perception (VP) computing regions-of-interest
(ROIs) out of an input image, a main image processing (MIP)
extracting local key-point vectors out of the tiled ROI area, and
a post database matching (PDM) performing matching of the
key-point vectors with a trained database for final decision.
The NPE, 16 SPUs, DP are proposed for the three stages,
respectively, while the TM is proposed to manage the ROI tasks
of the 16 SPUs. In this object recognition processor, distributed
memory architecture is adopted with two 200 MHz SRAM
devices in off-chip domain. The NPE containing instruction
and data cache initializes program codes of the other modules
and then, each module utilizes its own local memory for intermediate data store when it performs the assigned tasks. Because
every memory module is mapped to 32-bit address space in
the top level, each IP can access any other memory modules in
other IPs and two external ones through a fully interconnected
NoC. With stream processing model [18], each IP minimizes
external memory usage.
Fig. 1(b) shows major data transactions of the proposed processor. Three object recognition stages are mapped to a NPE, 16
SPUs, a DP, respectively, and are executed in the pipeline where
the pipelined data are ROI tiles and key-point vectors between
the first/second stage and second/third stage, respectively. The
NPE sends the information of ROI tiles to the message FIFO
in TM to make the TM assign them to 16 SPUs while the 16
SPUs store the resulting key-point vectors to the memory buffer
in the DP. Due to the pipelined operation, lots of data transactions seamlessly occur among the IP blocks throughout the
processing time. Under this condition, the NoC should provide
both sufficient data bandwidth and low latency transfer to enable
the processor’s target 60 fps processing at VGA video resolution. Major data transactions are as follows. At the VP stage,
the NPE distributes 0.8 Mb of instruction kernels to 16 SPUs.
The TM, in between the NPE and 16 SPUs, assigns the delivered ROI tiles to the 16 SPUs and manage their task threads.
In this process, the TM transfers 2 Mb of image pixel data to
16 SPUs each frame. In the MIP stage, the 16 SPUs perform
the feature description algorithm to generate key-point vectors
out of the selected ROI tiles. While running the algorithm, the
multiple SPUs communicate 200 Kb intermediate data themselves and transfer 100 Kb key-point vectors to the DP each
frame. In the PDM stage, the DP performs vector matching for
the aggregated key-point vectors with a trained object database.
From the three stages, each IP block requests 1 Mb raw image
and intermediate data transactions with external memories each
frame. Table I summarizes detailed traffic characteristics in the
proposed processor.
Based on the obtained traffic patterns, we derived the following five NoC requirements for real-time object recognition.
First of all, regardless of traffic patterns, low latency data
transaction is required for real-time processing. Because every
IP block of the proposed processor performs in-order processing, long data latency for read operations can seriously
affect the overall performance. Second, 1-to-N data transactions are frequently used from the NPE/TM to 16 SPUs, e.
g., when the NPE/TM distribute instruction kernels, ROI tile
tasks, and image data to 16 SPUs. To accelerate these data
KIM et al.: A 118.4 GB/s MULTI-CASTING NOC WITH HIERARCHICAL STAR-RING COMBINED TOPOLOGY FOR REAL-TIME OBJECT RECOGNITION
1401
TABLE I
DETAILED TRAFFIC CHARACTERISTICS
Fig. 2. Application specific NoC requirements.
transactions, broad- and multi-casting capability is demanded
in the NoC. Third, sufficient communication channels are
required among the SPUs for their co-processing. To this end,
a suitable NoC topology should be investigated. Fourth, N-to-1
data transactions are required for the aggregation of vector
packets from 16 SPUs to a DP. An additional output port is
needed to DP to handle this problem. Lastly, the NoC should
provide a multi-clocking environment for the overall SoC to
allow each IP to use a preferable clock frequency. To this end,
every interfacing region between the NoC and the IP blocks
should have data synchronization function to traverse between
heterogeneous clock domains. Fig. 2 summarizes the five
application-specific NoC requirements of the proposed object
recognition processor.
III. HIERARCHICAL STAR AND RING COMBINED TOPOLOGY
A. Topology Exploration
In the design of a NoC, topology selection deciding the structural organization of network switches is a primary concern. By
the topology, not only the performance of NoC such as average
latency and total bandwidth but also the implementation cost is
roughly determined. To find out the most suitable NoC topology
for the target object recognition, we choose four topologies depicted in Fig. 3: mesh [5], [6], hierarchical star [3], [4], [19],
ring [7], and proposed hierarchical star-ring (HS-R) combined
topology. The mesh topology has strength in regularity and scalability, but it is not good in terms of cost efficiency because each
IP requires its own network switch. On the other hand, the hierarchical star topology is a cost efficient solution based on its tree
structure that results in low latency data communication. However, the problem is that significant traffic workloads aggregate
to network switches in higher level. The ring topology also has
good cost efficiency, however, its long hopping latency cannot
continue scaling to many-core era containing tens of IPs. The
proposed HS-R topology combines the hierarchical star and ring
topology for short latency data transmission and inter-processor
data communication, respectively. Based on the hierarchical star
networks, it adds ring networks to achieve shorter hopping latency as well as to mitigate workload aggregation to higher level
network switches. Fig. 4 shows the block diagram of the conventional hierarchical star and proposed HS-R topology where
the number of IPs and network hierarchies scale to 64 and 3,
respectively. Different from the hierarchical star topology, the
HS-R topology additionally interconnects the switches in the
same network hierarchy with the ring topology. In this diagram,
the number in each IP box and grouped IPs by the distance level
means the switch hopping latency from the start IP and the set
of IPs having the same hopping latency from the start IP , respectively. The additional ring networks of the HS-R topology
reduce the hopping latency from the start IP by 1 when it accesses the IPs belongs to the neighbor clusters of each same distance group. As shown in Fig. 4, half of the overall IPs colored
in dark gray are benefited by the HS-R topology.
To evaluate the performance of above four topologies under
target application traffic, we made a cycle accurate NoC simulator broadly consisting of traffic model library and NoC model
library, as shown in Fig. 5. The traffic model library contains
synthetic traffic patterns and traffic generating subjects such
as real IP’s instruction set simulator and traffic generator. The
traffic generator can generate some typical traffic patterns as
well as arbitrary ones by reading a traffic description file (TDF)
that compiles the time and data of each data transaction event.
On the other hand, the NoC model library contains building
components of NoC such as a network switch, communication
socket, and network wrapper. We use transaction level modeling (TLM) [20] in the NoC model library. Because the TLM
operates based on the existence of transactions, it is good to
integrate various IP blocks modeled with different abstraction
levels. In addition, its event driven characteristic is beneficial
1402
IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 45, NO. 7, JULY 2010
Fig. 3. NoC topologies. (a) Mesh; (b) H-star; (c) ring; (d) H-star + ring.
Fig. 4. Comparison between hierarchical star and proposed hierarchical star–ring combined topology.
Fig. 5. NoC simulation framework.
for fast simulation time. A monitoring unit is included in each
network packet and switch model for packet tracing and statistic
extractions. Using these two model libraries, we can constitute
a cycle accurate NoC simulation framework. The traffic generating IP blocks are connected to the NoC using network wrapper
and TLM socket interface, while the NoC topology is built by a
link description file (LDF) that describes the connection status
of whole network switches.
Using the implemented NoC framework, we measured the
average latency, bandwidth utilization, and implementation
cost of the four NoC topologies when the target three-stage
object recognition algorithm runs for several VGA (640 480
pixels) input images. Average 56.49 s is spent by the simulator
to process a single VGA image. In this experiment, cycle
accurate models are used for the TM, 16 SPUs, and DP while
a functional model is used for the NPE. The bandwidth utilization is calculated by dividing the overall amount of packet data
through every network switch by the maximum total bandwidth
of every switch. The cost of the NoC topology is measured by
the amount of employed network switches and interconnecting
wires. Assuming every network switch is square (same input
and output port number, e. g., 5 5, 6 6, and 7 7), the
implementation costs of a network switch itself and the wires
from it are proportional to the number of input ports. Therefore,
the implementation cost of the overall NoC can be represented
by following equation where , , means the cost of a 1
1 sized switch, the cost of a single wire in average length, and
the data bit-width of each port.
Fig. 6 shows the experimental results. The implementation
cost is represented by the total number of input ports from all
switches, a normalized value with the unit cost of a switch
and wires per port. For latency, the hierarchical star and HS-R
topology show better performances than the mesh and ring
topology because their tree-like structure logarithmically reduces the switch hopping number to access other IP blocks. The
HS-R topology further reduces the latency from the hierarchical
star topology by accelerating inter-processor communication
via supplemented ring networks. For bandwidth utilization,
KIM et al.: A 118.4 GB/s MULTI-CASTING NOC WITH HIERARCHICAL STAR-RING COMBINED TOPOLOGY FOR REAL-TIME OBJECT RECOGNITION
1403
Fig. 6. Topology exploration results.
Fig. 7. MC-NoC architecture.
the mesh topology is the lowest among the four topologies
because it has the largest maximum total bandwidth due to its
abundant network switches and links. However they also bring
the most expensive implementation cost to mesh topology.
For implementation cost, the hierarchical star and HS-R show
good efficiency. In conclusion, mesh is not a suitable topology
for the target traffic because it does not show better latency
performance than the others even it requires larger implementation cost. The ring topology seems to be the worst among the
four topologies. This is because its expected switch hopping
number becomes too large in these 21 IP SoC. For smaller sized
SoCs, the ring topology can be a good solution with its low
cost interconnection. We can decide that the proposed HS-R
topology is the most suitable topology for the target traffic with
the lowest data latency and supplemented data bandwidth for
inter-processor communication.
B. MC-NoC Architecture
Fig. 7 shows the overall architecture of the proposed
MC-NoC consisting of a 9
10 system network switch and
four 7 7 local network switches. The 16 SPUs are connected
to the system network through the four local network switches
while the rest of IP blocks such as the NPE, TM, DP, and two
external interfaces are directly connected to the system network switch. As decided in the previous section, the MC-NoC
exploits the HS-R combined topology. It adopts a hierarchical
star topology as a basic starting topology for low latency data
communication, and then, supplements a ring topology to the
local networks for high speed inter-SPU data transactions. The
additional ring networks of the combined topology provide a
GB/s links
input ports
maximum bandwidth of
output ports
byte
MHz, assuming all physical
channels transfer data) between the local networks, and allow
each SPU to access the other SPUs in neighbor local networks
in two switch hops without accessing the system network.
In addition, for N-to-1 data transactions, the output port of
the DP is modified to a dual port. One of them is dedicated
to key-point vector packet aggregation from the 16 SPUs.
In overall, the topology combined MC-NoC provides overall
GB/s
input ports output ports
input ports
output ports
byte
MHz total bandwidth with the
switch hopping latency of less than 3.
IV. MULTI-CAST PROTOCOL
Multi-casting transmitting the same packet to multiple destinations [21] is an effect method when a single source wants
to send the identical packets to multiple destinations. It is also
very useful for the 1-to-N data requirements of the proposed
processor such as program code distribution and image data
download. However, multi-casting operation in NoC can cost a
lot because protocol, routing, and network switch design should
be redesigned. In this chapter, we present a cost efficient multicasting protocol and implementation in the proposed MC-NoC
based on the fact that the multi-casting source and destination
IP are already determined.
Fig. 8 shows the block diagram of the network switch consisting of input ports containing queuing buffer, output arbiters,
crossbar switch, and output ports. It uses a wormhole routing
protocol that controls the packet operation in a smaller flow control unit (flits) to reduce the size of input buffer. To achieve high
throughput, the network switch performs a packet transmission
with four pipeline stages. First, the incoming packet is buffered
at the input port in a flit unit. Then, each input port sends request and obtain grant to/from the output arbiter. After getting
a grant, the packet traverses the crossbar switch. To prevent the
packet loss, the network switch uses a credit based back pressure
scheme [22] that controls the flits in the router by the congested
1404
IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 45, NO. 7, JULY 2010
Fig. 8. Network switch block diagram.
Fig. 9. Packet format of MC-NoC.
status of receiver side. For multi-casting operation, it uses a synchronous method [23] that transmits the packet to the multiple
output ports at the same time. By eliminating packet retransmission, it can achieve simpler routing algorithm as well as higher
energy efficiency than asynchronous method.
Fig. 9 shows the detailed packet format of the MC-NoC. It
consists of three flit types—header, address, and data—while
each contains 34-bit information including 2-bit type identifier.
The flit is efficiently controlled by 2-bit flow control signals,
valid and back pressure hold signal.
The header flit contains all information of the packet such
as read/write, 4-bit burst length, 2-bit priority level, and 16-bit
routing information (RI). The address flit indicates the address
of the final destination IP where the data flits are read or written.
When the packet is routed in each network switch, the header filt
forwards first and the address/data flits follow after it. The burst
length information enables burst data transactions up to 8 data
flits. For this, the write burst packet sends header, address, and
multiple data flits of burst length to the target IP while the read
burst packet sends only header and address flit to get multiple
data flits of burst length from the target IP. The 2-bit priority
level indicates the priority of packet in the situation that the
network has contentions. By obtaining the arbiter’s grant earlier in each network switch, the packet with the higher priority
level is more quickly routed than the packet with the normal
priority level. Lastly, the 16-bit source defined routing information (RI) describes where the packet should be routed in each
network switch to the final destination. It allows 4 switch traversals for normal packets and a multi-casting to arbitrary SPUs
for multi-casting packets. For multi-casting case, we use a bit
string encoding [24] to notify the destination SPUs. In this encoding, each bit of 16-bit RI corresponds to each SPU and configuring a bit to 1 means the SPU is one of the multi-casting destinations. Without describing detailed routing path, the 16-bit
RI efficiently implicates the multi-casting destinations out of 16
SPUs. Each network switch decodes the 16-bit multi-casting RI
to generate correct routing paths to multiple destinations.
Fig. 10 shows the decoding process of a multi-casting packet
in the MC-NoC. With a bit string encoding, the example RI
intends to send a packet to SPU 0, 2, 3, 13, and 15. To this
end, the MC-NoC adopts a hierarchical approach using its
hierarchical-star topology: the system network performs the
first multi-casting to four SPU local networks, and then, local
network perform the second multi-casting to final destination
KIM et al.: A 118.4 GB/s MULTI-CASTING NOC WITH HIERARCHICAL STAR-RING COMBINED TOPOLOGY FOR REAL-TIME OBJECT RECOGNITION
1405
Fig. 10. Multi-casting process of MC-NoC.
Fig. 11. Four-stage pipelined multi-casting network switch.
SPUs. In the first multi-casting, the system network sends the
packet to the local networks containing at least one destination
SPU. The destination local networks can be identified by
performing OR operation on each of four 4-bit RI segments. In
the second multi-casting, each destination local network sends
a packet to destination SPUs according to corresponding 4-bit
RI segment. Through this two step multi-casting, the packet
correctly arrives at the destination SPUs. Different from the
mesh based NoC should consider the optimized multi-cast
routes out of lots of possible candidates, the hierarchical star
based MC-NoC can perform the hierarchical and deterministic
multi-cast routing because its upper level network switches
interconnect all of the lower level network switches. This
hierarchical multi-casting method can be expanded to larger
MC-NoCs having hierarchies more than two.
V. DETAILED MC-NOC DESIGN
In this Section, we discuss real implementation of MC-NoC.
The MC-NoC has three design issues to address. First, each
network switch should be able to perform multi-casting for
overall multi-casting capability. Second, the interface between
the MC-NoC and IP blocks can convert the heterogeneous
clock domains for multi-clocking SoC platform. Third, the
MC-NoC should operate with low power consumption for low
overhead interconnection platform.
A. Multi-Casting Network Switch
Fig. 11 depicts the block diagram of multi-casting network
switch composed of input ports, output arbiters, crossbar fabric,
and output ports. It performs packet transmission with four
pipelined stages. First, the incoming flits are buffered at 8-depth
32-bit first-in-first-out (FIFO) queue containing synchronization interface for the two heterogeneous clock domains. Then,
each activated input port sends a request signal to the output
arbiter to get a grant signal to traverse the crossbar fabric.
The arbiters perform round-robin scheduling with two priority
levels. After getting a grant signal, the input port sends the
packet to the destination output port through a fully connected
crossbar fabric. In case of multi-casting, the input port sends
multiple request signals to multiple output arbiters to ensure
1406
IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 45, NO. 7, JULY 2010
Fig. 12. Multi-casting port.
of the processor operation and the TM performs multi-casting
the image data to two SPUs, above conflicts in multi-casting
rarely happen and barely affect the overall performance.
To provide sufficient driving strength for variable loads in
multi-casting, we specially devise a variable strength driver
whose circuit diagram is shown in Fig. 13. It consists of three
parallel tri-state buffers whose driving forces are different to
2 , 4 , and 8 , respectively, while their output wires are
connected into a single final output to enhance the overall
circuits output driving strength. Using 3-bit control signals that
each of them decides the activation of each tri-state buffer, the
variable strength driver can dynamically configure its output
driving strength from 2 to 14 by a 2 step according to the
number of output ports of multi-casting. The table in Fig. 13
shows the operation modes of the variable strength driver by
the control signals and its energy consumptions according to
various multi-casting output loads.
Fig. 13. Circuit diagram of variable strength driver.
all grant signals of desired output ports. The multi-casting
input port waits until all grant signals are returned. For this
operation, the multi-casting input port employs a multi-port
requester and grant checker shown in Fig. 12. The multi-port
requester decodes the 16-bit routing information, generates
request signals to multiple arbiters, and latches the request
signals in a 16-bit register. Meanwhile, the grant checker holds
the multi-casting packet in the FIFO queue until the received
grant signals become equal to the latched request ones. After
gathering all grant signals, the multi-casting is performed
using broad-casted wires of crossbar fabric. In case of arbitrary
destinations are blocked by some reasons, the multi-casting
port should select whether it hold the grants until to gather all
grants or withdraw the request and request later. In this design,
the multi-casting request is withdrawn when the other inputs
send requests to the output arbiters whose grants are already
captured by the multi-casting port which is waiting remaining
grants, not performing multi-casting. In the proposed processor
because the multi-casting is used in cases that the NPE performs
multi/broad-casting program code to 16 SPUs at the beginning
B. FIFO Based Synchronization Interface
To allow the heterogeneous clock frequencies between the IPs
and MC-NoC, the first-in-first-out (FIFO) based synchronization is performed in every input/output port of network switches
interfacing with IP blocks. Fig. 14 shows the block diagram of
FIFO based synchronization interface that exists in each input
port. It consists of an 8-depth FIFO buffer, write control block,
and read control block, which are for data storage, write and
read operation control and pointer management, respectively. A
flag detector, in between the write and read control block, generates combinational flag signals such as empty and full using
relative locations of the two pointers. When the detected empty
or full signal is used in the other clock domain for flow control signal generation like request or hold, they are synchronized
with the clock of the other domain using a pipelined synchronization method [25].
C. Fine-Grained Clock Gating
For low power consumption, the MC-NoC performs a
fine-grained clock gating that activates the only required packet
routing path in each network switch. Fig. 15 shows the block
diagram of fine-grained clock gating and its operating timing
diagram. The monitor unit in each network switch detects the
KIM et al.: A 118.4 GB/s MULTI-CASTING NOC WITH HIERARCHICAL STAR-RING COMBINED TOPOLOGY FOR REAL-TIME OBJECT RECOGNITION
1407
TABLE II
CHIP SUMMARY
Fig. 14. Heterogeneous clock interface.
existence of incoming flits to activate corresponding input
queuing buffers. In addition, it partially decodes the header flit
to generate the list of output ports to activate the output arbiters
and crossbar fabric. Because the proposed MC-NoC supports
multi-casting operation, the destination output ports can be
multiple. With a simple D-flip-flop chain, the activation time of
each switch component such as input queues, output arbiters,
and crossbar fabric is determined in a pipelined manner. The
activated components operate until the end of a single packet
transmission while the monitor unit is always turned on. By
controlling the activation of each switch component in cycle
level including the input buffers consuming almost 90% of
network switch power, the fine-grained clock gating can save
28% power on average.
Fig. 15. Fine grained clock gating.
VI. IMPLEMENTATION RESULTS
To evaluate the performance of the MC-NoC, we measured
the cycle count and energy consumption from gate-level simulation for the all packets transmitted during the target object
recognition algorithm is running for a single VGA (640 480)
video image. We also performed the identical experiments on
a basic hierarchical star NoC without any scheme and a HS-R
combined NoC without multi-casting capability for comparison. Fig. 16 shows the cycle count and energy reduction effect
by the two main features of the MC-NoC. The HS-R combined
topology contributes to inter-SPU data communication. By
using the supplemented ring networks in inter-SPU communication, it shortens the routing path and reduces the number
1408
IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 45, NO. 7, JULY 2010
Fig. 16. MC-NoC evaluation results. (a) Cycle count. (b) Energy consumption.
The overall processor consumes 496 mW at 1.2 V while the
object recognition is running at 60 frame/sec for VGA video
input. Under this condition, the MC-NoC dissipates 48 mW,
only 9.7% of the overall processor. Table II summarizes the
chip and MC-NoC features.
VII. CONCLUSIONS
Fig. 17. Chip micrograph.
of packet buffering. The multi-casting capability contributes
to program distribution and image download. By removing
retransmissions of the same packet to multiple IPs, it not only
significantly reduces the number of packet buffering but also accelerates the data transactions as well. As a result, the MC-NoC
reduces the overall cycle count and energy consumption in data
communication by 20% and 23%, respectively, compared with
a basic hierarchical-star NoC.
The real-time object recognition processor with the proposed
MC-NoC is fabricated in a 0.13 m eight-metal CMOS process.
Fig. 17 shows the chip micrograph. The overall recognition
processor has 49 mm die area with 3.73 M gates while the
MC-NoC accounts for 350 k gates. The operating frequency
is 200 MHz for IP blocks and 400 MHz for MC-NoC. The
MC-NoC provides 118.4 GB/s total bandwidth (92.8 GB/s
from hierarchical star networks, 25.6 GB/s from ring networks)
at 400 MHz frequency and supports multi-casting to 16 SPUs.
A 118.4 GB/s multi-casting network-on-chip (MC-NoC) is
developed as communication platform for a real-time object
recognition processor. Through topology exploration, we derive the HS-R combined topology for low latency and inter-processor data communications. Multi-casting is supported to accelerate 1-to-N data transactions. With these two main features,
the proposed MC-NoC reduces data transaction time and energy
consumption by 20% and 23%, respectively, under the target object recognition traffic. The MC-NoC is fabricated in a 0.13 m
CMOS process in the proposed object recognition processor
while consuming 48 mW at 400 MHz, 1.2 V.
REFERENCES
[1] L. Benini and G. D. Micheli, “Networks on chips: A new SoC paradigm,” IEEE Computer, vol. 35, no. 1, pp. 70–78, Jan. 2002.
[2] W. J. Dally and B. Towles, “Route packets, not wires: On-chip interconnection networks,” in Proc. 38th Design Automation Conf., Jun.
2001, pp. 684–689.
[3] S.-J. Lee et al., “An 800 MHz star-connected on-chip network for application to systems on a chip,” in IEEE Int. Solid-State Circuits Conf.
(ISSCC) Dig. Tech. Papers, Feb. 2003, pp. 468–469.
[4] K. Lee et al., “A 51 mW 1.6 GHz on-chip network for low-power
heterogeneous SoC platform,” in IEEE Int. Solid-State Circuits Conf.
(ISSCC) Dig. Tech. Papers, Feb. 2004, pp. 152–153.
[5] S. Vangal et al., “An 80-tile 1.28TFLOPS network-on-chip in 65 nm
CMOS,” in IEEE Int. Solid-State Circuits Conf. (ISSCC) Dig. Tech.
Papers, Feb. 2007, pp. 98–99.
[6] S. Bell et al., “TILE64TM processor: A 64-core SoC with mesh interconnect,” in IEEE Int. Solid-State Circuits Conf. (ISSCC) Dig. Tech.
Papers, Feb. 2008, pp. 88–89.
[7] T. W. Ainsworth and T. M. Pinkston, “Characterizing the cell EIB
on-chip network,” IEEE Micro, vol. 27, no. 5, pp. 6–14, 2007.
[8] D. A. Forsyth and J. Ponce, Computer Vision: A Modern Approach.
New York: Prentice Hall, 2002.
KIM et al.: A 118.4 GB/s MULTI-CASTING NOC WITH HIERARCHICAL STAR-RING COMBINED TOPOLOGY FOR REAL-TIME OBJECT RECOGNITION
[9] D. G. Lowe, “Distinctive image features from scale-invariant keypoints,” ACM Int. J. Comput. Vis., vol. 60, no. 2, pp. 91–110, Jan.
2004.
[10] R. R. Rojas et al., “Object recognition SoC using the support vector
machines,” J. Appl. Signal Process., pp. 993–1004, 2005.
[11] H.-C. Lai et al., “Communication-aware face detection using NoC architecture,” in Proc. Int. Conf. Computer Vision Systems (ICVS), 2008,
pp. 181–189.
[12] D. Kim et al., “An 81.6 GOPS object recognition processor based on
NoC and visual image processing memory,” in Proc. IEEE Custom Integrated Circuits Conf. (CICC), 2007, pp. 443–446.
[13] K. Kim et al., “A 125GOPS 583 mW network-on-chip based parallel
processor with bio-inspired visual attention engine,” IEEE J. SolidState Circuits, vol. 44, no. 1, pp. 136–147, Jan. 2009.
[14] J.-Y. Kim et al., “A 201.4GOPS 496 mW real-time multi-object recognition processor with bio-inspired neural perception engine,” in IEEE
ISSCC Dig. Tech. Papers, Feb. 2009, pp. 150–151.
[15] D. Kim et al., “Implementation of memory-centric NoC for 81.6 GOPS
object recognition processor,” in Proc. IEEE A-SSCC, Nov. 2007, pp.
47–50.
[16] K. Kim et al., “A 76.8 GB/s 46 mW low-latency network-on-chip for
real-time object recognition processor,” in Proc. IEEE A-SSCC, Nov.
2008, pp. 189–192.
[17] J.-Y. Kim et al., “Real-time object recognition with neuro-fuzzy controlled workload-aware task pipelining,” IEEE Micro, vol. 29, no. 6,
pp. 28–43, Nov./Dec. 2009.
[18] W. J. Dally et al., “Stream processors: Programmability with efficiency,” ACM Queue, vol. 2, no. 1, pp. 52–62, 2004.
[19] K. Lee et al., “Low-power networks-on-chip for high-performance SoC
design,” IEEE Trans. VLSI, vol. 14, no. 2, pp. 148–160, Feb. 2006.
[20] F. Ghenassia, Transaction Level Modeling With SystemC: TLM Concepts and Applications for Embedded Systems. New York: Springer,
2005.
[21] L. L. Peterson and B. S. Davie, Computer Networks: A Systems Approach, ser. The Morgan Kaufmann Series in Networking. : Morgan
Kaufmann, 2007.
[22] H.-J. Yoo, K. Lee, and J. K. Kim, Low Power NoC for High Performance SoC Design. Boca Raton: CRC Press, 2008.
[23] C. M. Chiang and L. M. Ni, “Deadlock-free multi-head wormhole
routing,” in Proc. First High Performance Computing–Asia, Taiwan,
Taiwan, Sep. 1995.
[24] C. Chiang and L. Ni, “Multi-address encoding for multicast,” in Proc.
1st Int. Workshop on Parallel Computer Routing and Communication,
PCRCW’94, Seattle, WA, May 1994, pp. 146–160.
[25] J. N. Seizovic, “Pipeline synchronization,” in Proc. IEEE ASYNC, Nov.
1994, pp. 87–96.
Joo-Young Kim (S’05) received the B.S., M.S., and
Ph.D. degrees in electrical engineering and computer
science from the Korea Advanced Institute of Science
and Technology (KAIST), Daejeon, Korea, in 2005,
2007, and 2010, respectively.
Since 2006, he has been involved with the development of the parallel processors for computer
vision. Currently, he is a postdoctoral researcher at
KAIST. His research interests include multi-core
architecture, network-on-chip, and VLSI implementations for computer vision applications.
Junyoung Park (S’09) received the B.S. degrees in
electrical engineering and computer science from the
Korea Advanced Institute of Science and Technology
(KAIST), Daejeon, Korea, in 2009 and is currently
working toward the M.S. degree in electrical engineering and computer science at KAIST.
Since 2009, he has been involved with the development of the parallel processors for computer vision.
Currently, his research interests are many-core architecture and VLSI implementation for bio-inspired vision processor.
1409
Seungjin Lee (S’06) received the B.S. and M.S. degrees in electrical engineering and computer science
from the Korea Advanced Institute of Science and
Technology (KAIST), Daejeon, Korea, in 2006 and
2008, respectively. He is currently working toward
the Ph.D. degree in electrical engineering and computer science from KAIST.
His previous research interests include low power
digital signal processors for digital hearing aids and
body area communication. Currently, he is investigating parallel architectures for computer vision processing.
Minsu Kim (S’07) received the B.S. and M.S. degrees in electrical engineering and computer science
from the Korea Advanced Institute of Science and
Technology (KAIST), Daejeon, Korea, in 2007 and
2009, respectively. He is currently working toward
the Ph.D. degree in electrical engineering and computer science at KAIST.
His research interests include network-on-chip
based SoC design and bio-inspired VLSI architecture
for intelligent vision processing.
Jinwook Oh (S’08) received the B.S. degree in electrical engineering and computer science from Seoul
National University, Seoul, Korea, in 2008. He is currently working toward the M.S. degree in electrical
engineering and computer science at KAIST, Daejeon, Korea.
His research interests include low-power digital
signal processors for computer vision. Recently, he
has been involved with the VLSI implementation of
neural networks and fuzzy logics.
Hoi-Jun Yoo (M’95–SM’04–F’08) graduated from
the Electronic Department of Seoul National University, Seoul, Korea, in 1983 and received the M.S.
and Ph.D. degrees in electrical engineering from the
Korea Advanced Institute of Science and Technology
(KAIST), Daejeon, in 1985 and 1988, respectively.
His Ph.D. work concerned the fabrication process
for GaAs vertical optoelectronic integrated circuits.
From 1988 to 1990, he was with Bell Communications Research, Red Bank, NJ, where he invented
the two-dimensional phase-locked VCSEL array, the
front-surface-emitting laser, and the high-speed lateral HBT. In 1991, he became Manager of a DRAM design group at Hyundai Electronics and designed
a family of from fast 1 M DRAMs and 256 M synchronous DRAMs. In 1998,
he joined the faculty of the Department of Electrical Engineering at KAIST and
now is a full Professor. From 2001 to 2005, he was the Director of System Integration and IP Authoring Research Center (SIPAC), funded by Korean government to promote worldwide IP authoring and its SOC application. From 2003 to
2005, he was the full time Advisor to Minister of Korea Ministry of Information
and Communication and National Project Manager for SoC and Computer. In
2007, he founded SDIA(System Design Innovation and Application Research
Center) at KAIST to research and develop SoCs for intelligent robots, wearable
computers and bio systems. His current interests are high-speed and low-power
network-on-chips, 3-D graphics, body area networks, biomedical devices and
circuits, and memory circuits and systems. He is the author of the books DRAM
Design (Seoul, Korea: Hongleung, 1996; in Korean), High Performance DRAM
(Seoul, Korea: Sigma, 1999; in Korean), and chapters of Networks on Chips
(New York, Morgan Kaufmann, 2006).
Dr. Yoo received the Electronic Industrial Association of Korea Award for
his contribution to DRAM technology the 1994, Hynix Development Award in
1995, the Korea Semiconductor Industry Association Award in 2002, Best Research of KAIST Award in 2007, Design Award of 2001 ASP-DAC, and Outstanding Design Awards of 2005, 2006, 2007 A-SSCC. He is a member of the
executive committee of ISSCC, Symposium on VLSI Circuits, and A-SSCC.
He is the TPC chair of the A-SSCC 2008.
0
Purchase answer to see full
attachment](data:image/png;base64,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)
Why Choose Us
- 100% non-plagiarized Papers
- 24/7 /365 Service Available
- Affordable Prices
- Any Paper, Urgency, and Subject
- Will complete your papers in 6 hours
- On-time Delivery
- Money-back and Privacy guarantees
- Unlimited Amendments upon request
- Satisfaction guarantee
How it Works
- Click on the “Place Your Order” tab at the top menu or “Order Now” icon at the bottom and a new page will appear with an order form to be filled.
- Fill in your paper’s requirements in the "PAPER DETAILS" section.
- Fill in your paper’s academic level, deadline, and the required number of pages from the drop-down menus.
- Click “CREATE ACCOUNT & SIGN IN” to enter your registration details and get an account with us for record-keeping and then, click on “PROCEED TO CHECKOUT” at the bottom of the page.
- From there, the payment sections will show, follow the guided payment process and your order will be available for our writing team to work on it.