斷金計畫 - RSWiki

Report
斷金計畫
Information and Communications
Laboratories (ICL)
Cloud Computing Research Center for
Mobile Applications (CCMA)
Shortage of ICT R&D Talents in ITRI
• Problems:
– Financial compensation not competitive with top-paying HSIP companies
– Low visibility in terms of commercial impacts and research
accomplishments in recent years
• Direct outcome:
– ITRI is not on the radar of many top ICT students when they look for jobs
– Many engineers, but not enough researchers
• Indirect outcomes: less ambitious projects, and fewer success
stories such as hot start-ups and inspiring technology
breakthroughs
• Weaker ICL  Lower momentum to turn around Taiwan’s ICT
industry at a time when it is needed most
– IC design
– 5G
We want to change this by collaborating
with universities in creative ways
眾志成城,其利斷金
Three-Pronged Approaches
1. ITRI-supervised Masters’ thesis project
– Finnish- and Swedish-style industrial apprentice
– Students work with pay in ITRI full time for six months or more to
complete their thesis
– Thesis quality must be approved by university
2. Ph.D. students working in joint projects between ITRI and
universities
– ITRI employees joining Ph.D. program
– University Ph.D. students working in ITRI
– Balance between publication requirements and industrial impacts
– Structured in such a way to increase the matriculation rate of the
Ph.D. program
Three-Pronged Approaches
3. University team (professors and students) works in ITRI for
a period of time
– Over the summer months, or over a year
– Pre-negotiated new project topic or existing project topic
– Fully funded as an internal ITRI project, which is typically lead by
a university faculty member, and may be joined by other ITRI
members
– Designed to enable professors to
•
•
•
Get a taste of what it is like to work on projects that aim to meet industry
needs, which entail additional robustness and usability requirements
Apply and test out technologies developed in universities to real-world
problems
Facilitate or broker transfers of university technologies
Shape of Collaboration
• Worst case
– Small funding
– Topic or deliverable not particularly exciting or on critical path
– Courteous but lukewarm interaction
• Best case
– Larger funding
– Compelling and intellectually challenging problem in a module of
a system ITRI is building
• ITRI covers the baseline implementation
• University tries an alternative and more adventurous approach
– Tighter working relationship with the ITRI team  on site
ITRI Projects Looking for Collaborators
•
•
•
•
•
•
•
5G communication
Hybrid software-defined network architecture
Large-scale data center storage management
Next-generation Smartphone OS
Reconfigurable array of inexpensive batteries architecture
Advanced video/voice/text analysis technologies
Electronic system-level exploration of power/performance
impacts of software/hardware architectural decisions
• PCI Express as a network technology
• In-memory database management for IOT PaaS
• Low-overhead virtual desktop infrastructure for
video/graphics applications
5G Communication
•
•
•
•
•
•
•
Antenna array design
Massive MIMO
Co-channel full-duplex transmission
Millimeter Wave RF circuit and packaging design
Coordinated radio resource management
Distributed/networked MIMO
High-fidelity low-overhead system-level simulation
environment for protocol and HW development and testing
• Network function virtualization
Hybrid Software-Defined Network
•
•
Motivation: Ethernet and OpenFlow switches will co-exist in the next
decade if OpenFlow switches indeed take off
Hybrid Ethernet/OpenFlow network models:
1.
2.
3.
•
•
•
Unmodified Ethernet switches and OpenFlow switches
Ethernet switches and OpenFlow switches under the same controller
Ethernet in the core and OpenFlow at the edge or ECOE
ITRI has an Ethernet-based SDN technology called Peregrine that can
do 2 and 3
ITRI owns the SNMP4SDN module of OpenDaylight, which provides
the southbound API for controlling Ethernet switches: unicast and
multicast forwarding table, and VLAN
ECOE may obsolete OpenFlow switches altogether if its functionalities
are shown to be comparable to OpenFlow-everywhere network
Flash Storage Management
• SATA3-based: 1 million random 4KB writes per second over the
network  faster disk
• PCIe-based: NVMe
• Memory bus-based: slower and lower-power DRAM
– NVDIMM
– Hybrid DIMM
N-GB DRAM +
(128-N)-GB Flash =
128 GB DRAM?
• Other non-volatile
technologies: RRAM
and MRAM
Next-Generation Smart Phone OS
• Target: Local ASEAN brands between MTK and HTC
– Micromax/Karbonn (India), Ninetology/Smartfren (Indonesia), Cherry Mobile
(Philippine)
• Carrier-transparent vertical handoff between Wifi and cellular
• Security vs. convenience problem due to “bring your own device” (BYOD)
– Run a hypervisor on a mobile phone, and use different VMs for different
purposes: Work, Personal, Playground
– Only Work VM can be connected to office network
• Programmatic control of applications
– App as a library with automatically learned API
– User-specific short-cut
– Meta-application that interacts with multiple applications
• Introspective user interest profiling:
– Smartphone understands its owner as much as Google does
– Typed text sequence: “Address” to map app, “Time” to calendar app
Reconfigurable Array of Inexpensive
Batteries Architecture
Battery Resource
Controller (BRC)
Battery
Packs
Programmable
Power
Interconnect
Fabric
(PPIF)
Electric
Loads
Electric
Sources
Advanced Analysis for Law Enforcement
• Cloud-based surveillance video analysis
– Human-in-the-loop intelligent video browsing
– Vehicle color and make identification
– Cross-camera people trajectory tracking
• Cloud-based intercepted voice analysis
– Speech-to-text transcription
– Threading of related calls
– Suspicious conversation detection
• Chinese social opinion text analysis platform
– Crawl public forums and micro-blogging sites
– Perform basic and domain-independent NLP analysis
– Provide API for domain-dependent NLP analysis
• B2B Einvoice analysis for white-collar crime detection
• Wearable device for first responders and care-givers (nurses)
Thank You!
Questions and Comments?
[email protected]
M
• Quantitative

similar documents