AOT SLID presentation (engl.)

Download Report

Transcript AOT SLID presentation (engl.)

Overview
• Company overview
• Executive summary
• SLID roadmap
• SLID Key benefits
• SLID target application
• SLID performance comparison
• SLID technology
-
Basic principle
Content addressable Memory
Detection mechanism
Technology information
• Business model
• Evaluation environment
• Support / Contact
• Awards
• Patent status
Company overview
• Company name:
• Location:
Advanced Original Technologies
Matsuba-Cho 4-7-4-101, Kashiwa City
227-0827 Chiba Prefecture, Japan
• CEO:
Katsumi Inoue
• Established:
Sept. 2010
• Capital:
$220.000
• Business Summary: Technology development,
IP sales
Executive summary
1) SLID is a new architectural conceptual processor for recognition and
search purposes.
2) SLID improves the weaknesses of existing DSP and CPU solutions and
improves significantly power consumption for recognition and search uses
cases.
3) SLID has a high affinity towards other devices and is extremely easy to
handle.
4) SLID enables recognition performance beyond super computer capabilities
5) SLID can be a stand alone chip (road map) and can also be integrated into
digital Basebands , Application Co-processors , Sensor and other IC`s.
6) SLID enables total new use cases and generates new business concepts
SLID Roadmap
SLID
(ASIC-Ⅰ)
Idea
Fuzzy SLID
(FPGA)
1st TinySLID -1k
Demo
(FPGA)
2012CES
Introduction
Established
AOT
2010/Sep
2nd Generation
TinySLID -8k
(FPGA)
2011
2011/Jul
2012
Existing HW
2013CES
Introduction
2013
Planned HW
2014
SLID (ASICⅡ)
idea
SLID Key benefits
1)
Recognition of objects in <50μS possible
=> Can recognize >20000 Objects / sec.
=> parallel recognition possible
2)
No difference between exact and fuzzy recognition.
=> Can recognize >20000 fuzzy Objects / sec
3)
Edge Detection in Color possible. Extreme fast Edge detection possible (< 50μS )。
Possible to detect shapes
4)
Fuzzy Search in terms of position and Value possible (see explanation p.x)
5)
SLID can be digitally integrated into any semiconductor, but also build a stand alone
roadmap with different performance characteristics
6)
No need for special HW & SW => Reduces R&D costs
7)
Reduces significantly power consumption for search tasks
8)
Miniturization and weight reduction benefits
SLID Key benefits
1) Speed
pixel
size
PC 1
PC 2
SLID
ASIC
QQVGA
Y
Exact Match
Fuzzy Match
Exact Match
Fuzzy Match
Exact Match
Fuzzy Match
QVGA
VGA
120
240
480
11,000 μs
43,000 μs
143,000 μs
143,000 μs 842,000 μs 11,600,000 μs
18,800 μs 279,000 μs
916,000 μs
511,000 μs 3,200,000 μs 43,500,000 μs
50 μs
50 μs
5 μs
50 μs
50 μs
50 μs
30.000 times faster than Exact Match on PC !
2.000.000 times faster than Fuzzy Match on PC !
SLID Key benefits
2)
Search with exact values
„Blue“
„red“
„Black“
„green“
Normal search pattern has exact values
eg, green, blue, red, black etc.
„yellow“
SLID Key benefits
2)
Search with “fuzzy” values
„Blue“ish
„red“ish
„Black“ish
„green“ish
„yellow“ish
Is it possible to look for close values , eg. Colours who are
close to the original value
SLID Key benefits
2)
Search with fuzzy positions
Black eyes
Face colour
Face size
Pink lips
Is it possible enlarge the search area => fuzzy search
You can combine “fuzzy values” with “fuzzy position” search
SLID Key benefits
3)
Edge detection
SLID Key benefits
3)
Edge detection
- Detects immediately address of red bodies.
- Size and form can be instantly recognized
SLID Key benefits
4) Edge Detection
- Detects immediately address of red
bodies.
- Size and form can be instantly
recognized
=> It’s possible to look for shape.
SLID target applications
Face recognition
Search
Object
Immediate search result !
SLID target applications
Weather pattern recognition
Search
Object
Immediate search result !
SLID target applications
Chart pattern recognition (eg. Stock pattern)
Search
Object
Immediate search result !
SLID target applications
Data search from Server side ( parallel usage of SLID`s)
Server
Adress will be
given back to
server
Data
transfer
to SLID
SLID
SLID
SLID
SLID
SLID
SLID
SLID
SLID
SLID
SLD
SLID
SLID
SLID
SLID
SLID
SLID
SLID
SLID
SLID
SLID
SLID
SLID
SLID
SLID
SLID
SLID
SLID
SLID
SLID
SLID
SLID
SLID
SLID
PC analysis
Software
Immediate search result !
SLID target applications
Traffic control ( eg. number plate recognition)
It is also possible, for instance, to specify the locations of the license plates
on cars in an extremely fast manner.
SLID target applications
DNA Search
Conventional search
・・・GATCATTGA・・・
Search
Object
SLID target applications
DNA Search
Search with SLID
・・・GATCATTGA・・・
Search
Object
Immediate search result !
SLID target applications
Moving object recognition
T1
T2
T3
T4
・・・・No change to the background・・・A car has passed through it・・・・
SLID target applications
Moving object recognition
The area that does not match is the area
that has moved.
SLID target applications
Moving object recognition
T1
T2
T3
Super-simple and fast recognition of a moving body
T4
SLID target applications
Stereo Match
L Image
Measure depth by the difference in
position in the horizontal direction.
R Image
SLID target applications
Further applications ideas !
1)
2)
3)
4)
5)
6)
7)
8)
9)
Compares 2 videos ( piracy identification)
Sound recognition
Character recognition
Finger Print recognition
Smile recognition
3D (Video) recognition
Web Search
Graphic defect search
Moving object tracking
SLID vs CPU
• CPU detection search
mechanism
= serial search mechanism
scanning all memory adresses
CPU search takes extreme
long time !
SLID vs CPU
Immediate search result !
SLID vs CPU
CPU
•
Does set operating only with
values
vs.
•
Inoue
:update
SLID
Does set operating with values
•
Does set Operating with
addresses
•
Does parallel set operating
with addresses and values
•
Give adress out
SLID technology
• CPU is doing information processing only
sequentially and hence extremely slow for set
operating processing. If CPU speed is increased heat
and power consumption will increase
• SLID is processing data en bloc parallel
•SLID is compared to CPU many 1000 times faster
and can reduce power consumption and heat. Address
=> This enables totally new use cases, application
and ideas !!!
0000
h
0001
h
0002
h
0003
h
0004
h
0005
h
Data 0
Data 1
Data 2
Data 3
Data 4
Data 5
000n
h
Data n
Data
Basic principle
Actual data is stored linearly from the first dimension to
the Nth dimension into CAM
SLID is using position and search data as search input
search pattern is matched with stored data
Address shift can detect “fuzzy” data location
Σ(data & data location)= Pattern
Address is used as output
CAM Block Diagram
Input
Output
Value
Designation
Match Address
Data bus
Address bus
memory
c o m par at o r
memory
c o m par at o r
memory
c o m par at o r
memory
c o m par at o r
memory
c o m par at o r
memory
c o m par at o r
memory
c o m par at o r
memory
c o m par at o r
memory
c o m par at o r
memory
c o m par at o r
memory
c o m par at o r
memory
c o m par at o r
memory
c o m par at o r
memory
c o m par at o r
memory
c o m par at o r
memory
c o m par at o r
memory
c o m par at o r
memory
c o m par at o r
memory
c o m par at o r
memory
c o m par at o r
memory
c o m par at o r
memory
c o m par at o r
memory
c o m par at o r
memory
c o m par at o r
memory
c o m par at o r
Priority Address encoder output
c o m par at o r
Address decoder
memory
Priority Address encoder output
c o m par at o r
Address decoder
memory
Priority Address encoder output
Address decoder
Match address bus
SLID Block Diagram
Principle of SLID detection
Example 1
Principle of SLID detection
Because real images are very complex, here we will use an extremely
simple image. This kind of image is stored in SLID’s memory.
Query data
This is the
pattern we
want to
find.
Principle of SLID detection
A mask is placed over the
entire image.
Principle of SLID detection
Counters are attached here
(every pixel).
counter
counter
counter
counter
counter
counter
counter
Principle of SLID detection
Where is
Black?
Principle of SLID detection
Where is
Black?
Principle of SLID detection
Where is
Black?
Windows is made in the Mask
Principle of SLID detection
1
1
Where is
Black?
1
1
There is a possibility that the required image is
somewhere around these 4 pixels.
Principle of SLID detection
The mask,
equipped with a
counter and
punctured with
holes, can be
moved at an
ultra-fast speed
to any arbitrary
coordinate.
Principle of SLID detection
Where
is Red?
Principle of SLID detection
Where
is Red?
Principle of SLID detection
What’s the relationship between the Black and the Red?
2
2
Principle of SLID detection
Where
is Green?
Principle of SLID detection
Where
is Green?
Principle of SLID detection
What’s the relationship between the Black and the Green?
3
3
Principle of SLID detection
Where
is Blue?
Principle of SLID detection
Where
is Blue?
Principle of SLID detection
What’s the relationship between the Black and the Blue?
4
Principle of SLID detection
Where
is Yellow?
Principle of SLID detection
Where
is Yellow?
Principle of SLID detection
What’s the relationship between the Black and the Yellow?
5
Here it is!
This is fully parallel
detection.
Principle of SLID detection
Example 2
Principle of SLID detection
This object we would like
to search
Query image
Small size Image
10 columns x 5 rows = 50 pixels
Principle of SLID detection
Query image
Principle of SLID detection
Relative distance is a constant
value in this image space.
This is the basics of SLID.
Query image
Principle of SLID detection
(Primary judgement)
Query image
Primary
Judgment
Principle of SLID detection
(Secondary judgment -1)
Query image
Primary
Judgment
Principle of SLID detection
(Secondary judgment -2)
Query image
Primary
Judgment
Secondary
Judgment
Principle of SLID detection
(Tertiary judgment -2)
Query image
Primary
Judgment
Secondary
Judgment
Principle of SLID detection
(Tertiary judgment -2)
Query image
Primary
Judgment
Secondary
Judgment
Tertiary
Judgment
Principle of SLID detection
(Tertiary judgment -2)
Query image
Here!
Primary
Judgment
Secondary
Judgment
Tertiary
Judgment
Technology information
• FE process:
• package:
• Die size:
• Power consumpt.:
• Deliverables:
TMSC 90 nm
QFN 48
(see next page)
xx mA
RTL code in Verilog source
Software in C-code source
Integration testbench with set of testcases
Synthesis scripts
Documentation: functional specifications,
integration guide
Test reports
FPGA Platform (additional cost)
Technology information
Spec
One Dimension SLID
QVGA
320×240pix
data processing content
Data Bus
Adress Bus
Pattern Match condition input
possible ?
Data matching search function
Data area search function
Adress matching search
function
Adress area search function
Pattern match function
High Speed Pattern match
function
Matching counter function
Matching adress output
function
Chip Size (90 nm)
Memory Size
Data transfer Speed
2 Dimension SLID (Fuzzy Slid for graphics)
QVGA-II
VGA
VGA
320×240pix
640×480pix
640×480pix
SXGA
1280×1024pix
R, G、B each 4bit
Yes
Yes
R, G、B each 4bit
Yes
Yes
R, G、B each 4bit
Yes
Yes
R, G、B each 4bit
Yes
Yes
R, G、B each 4bit
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
no
yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
no
no
Yes
no
Yes
Yes
Yes
no
Yes
Yes
Yes
Yes
Yes
5mm×5mm
125KB
< 200μs
Yes
7mm×7mm
306KB
< 200μs
Yes
9mm×9mm
506KB
< 800μs
Yes
12mm×12mm
900KB
< 800μs
Yes
17mm×17mm
1600KB
< 800μs
yes
yes
yes
yes
no
yes
yes
yes
100MByte/s
Pattern match time
One dimension:
1 Set ( 5 points)
< 1μs
2 Dimension: 1 Set ( 5 points) fuzzy pattern match
< 5μs
Application example
Thesaurus text search
Relational database
datamining , DNA analysis
High speed precise recognition
face recogintion, OCR, graphic positiong
graphic defect search, stereomatch , moving picture tracking
Business model
• Full access to source code (RTL)
• License fee plus royalties
• Flexible terms in regards to single/multiple use
• License Fee includes training and initial support
• Maintenance Program
• Customization and IP Integration Design Services available from
AOT Technologies
Support / Contact
Sales for SLID is handled by Cross Border Technologies
For Japan / US (Axel Bialke)
Email:
[email protected]
Mobile Phone: +81 80 8030 3330
• For Korea / Taiwan / China (Eric Kim)
Email:
[email protected]
Mobile Phone: +82 10 2371 3532
• For Europa (Andreas vom Felde)
Email:
[email protected]
Mobile Phone: +49 176 3235 5412
Award
Demonstrator
Add picture
Patent
add