[hpc-announce] CFP: IEEE JETCAS Special Issue on Communication-aware Designs and Methodologies for Reliable and Adaptable On-Chip AI SubSystems and Accelerators

[Maurizio Palesi]

(We apologize for possible cross-posting)

IEEE JOURNAL ON EMERGING AND SELECTED TOPICS IN CIRCUITS AND SYSTEMS

CALL FOR PAPERS
Communication-aware Designs and Methodologies for Reliable and Adaptable
On-Chip AI SubSystems and Accelerators
https://ieee-cas.org/sites/default/files/si_cfp_final_2.pdf

GUEST EDITORS
Kun-Chih (Jimmy) Chen, National Sun Yat-sen University, Taiwan (
kcchen at mail.cse.nsysu.edu.tw)
Masoumeh (Azin) Ebrahimi, KTH Royal Institute of Technology, Sweden (
mebr at kth.se)
Maurizio Palesi, University of Catania, Italy (
maurizio.palesi at dieei.unict.it)
Tim Kogel, Synopsys, Germany (tim.kogel at synopsys.com)

SCOPE AND PURPOSE
The Big data and Internet of Things (IoTs) trend helps to drive the
progress of artificial intelligence (AI) technologies in recent years. The
notable benefits of neural network (NN) technologies (e.g., artificial
neural network (ANN), spiking neural network (SNN), etc.) are widely
applied to many real-world applications, such as speech recognition and
image classification, and the resulting accuracies have been well above
human-level. Due to its undoubted significance, research works of “AI
accelerator/subsystem designs” have drawn lots of attention from both
academia and industry.

Due to the massive parallel processing, the performance of the current
large-scale artificial neural network is often limited by the massive
communication overheads and storage requirements. However, the issues of
interconnection, communication, computation synchronized with memory
subsystem, reliability, and flexibility of AI engines receive less
attention. As a result, flexible interconnections and efficient data
movement mechanisms for future on-chip AI accelerator are worthy of study,
such as:

* New efficient data movement in contemporary AI subsystems: Memory access
latency and overhead of NN connections already become the performance
bottleneck. The memory area already dominates the total silicon cost, and
the data access contributes to a large portion of power consumption in
accelerators of deep neural networks (DNN). Many researchers tackled this
issue by improving on conventional Von Neumann-type architectures, such as
pruning and scheduling. In recent years, novel Non-Von Neumann
architectures, such as In-memory Computing and Near-memory Computing
techniques, had been proposed extensively to accommodate the
computing/processing-in-memory issue, which is worthy of investigation in
this special issue.

* Design methodology considering tradeoffs among computing engine (PE
arrays) and data movement/storage unit (memory hierarchy) from
energy/power/timing point of view: In conventional AI subsystem designs,
most works focus on task scheduling to improve the efficiency of the data
movement between the memory and the computing engine. In the emerging
AI-on-Chip designs, in addition to timing efficiency, power and energy
efficiencies become major concerns. Consequently, novel and efficient task
scheduling and data movement methodologies from energy/power/timing
optimization points of view are covered in this special issue as well.

* “Flexible” and “reliable” communication-aware AI subsystems for future
on-chip adaptive learning application: With predefined functional datapath
of dedicated neural network models, the computing flows of the current AI
accelerators are usually “fixed” and “non-adaptive.” On the other hand, the
reliable issue of the SNN design becomes even severe in leading technology
nodes because the spike signals are noise sensitive. Therefore, this
special issue wants to invite research works of novel computing flows, in
consideration of flexible and reliable data movements and communications,
for versatile modern AI applications.

TOPICS OF INTEREST
Topics of interest to this special issue include, but are not limited to:
- Challenges of massive memory data access in deep learning
- Near- and In-memory computing techniques for saving data movement
- ANN design based on emerging non-volatile memory devices
- Data movement optimization through task scheduling of the artificial
neural network
- Energy- and accuracy-aware pruning and quantization mechanism of neural
networks
- Novel interconnection networks for the neural networks (e.g., DNN, RNN,
ANN, SNN, etc.)
- Efficient on-chip communication of multicore-based artificial neural
network
- Communication/traffic-aware artificial neural network architecture and
application
- New topology of on-chip communication for efficient neural network
computing
- NoC design for heterogeneous ANN computing
- Reliable and robust computing method and on-chip interconnection for AI
computing
- Cross-layer optimization for artificial neural network architecture and
application
- Tradeoffs among computing engine and data movement from
energy/power/timing point of view
- Flexible computing flow and reconfigurable neural network architecture
for on-chip learning applications and Artificial General Intelligence (AGI)
- AI-on-Chip interconnections/designs to meet with the requirements and
constraints of emerging application (e.g., 5G, medical applications,
Industry 4.0, etc.)
- “Flexible” and “reliable” communication-aware AI subsystems for future
on-chip adaptive learning application

SUBMISSION PROCEDURE
Prospective authors are invited to submit their papers following the
instructions provided on the JETCAS website:
https://mc.manuscriptcentral.com/jetcas. The submitted manuscripts should
not have been previously published nor should they be currently under
consideration for publication elsewhere.

IMPORTANT DATES
* Manuscript submissions due 2020-05-20
* First round of reviews completed 2020-06-20
* Revised manuscripts due 2020-07-10
* Second round of reviews completed 2020-07-30
* Final manuscripts due  2020-08-15
* Target publication date   2020-09-30

REQUEST FOR INFORMATION
* Kun-Chih (Jimmy) Chen (kcchen at mail.cse.nsysu.edu.tw)