ISC HIGH PERFORMANCE 2021

FRIDAY, JULY 2, 2021, DIGITAL

THE EVENT FOR HIGH PERFORMANCE COMPUTING, NETWORKING AND STORAGE

WORKSHOP

Approximate and Transprecision Computing on Emerging Technologies (ATCET)

ORGANIZERS

Dr Cristiano Malossi
Prof. Luca Benini
Prof. Norbert Wehn
Prof. Roger Woods
Dr Andrew Emerson

Approximate and Transprecision Computing on Emerging Technologies (ATCET) 2nd Edition

In the last 10 years, the demand for new computing strategies driven by energy-efficiency has grown exponentially. Flop-per-watt (thus, per-euro) has become de-facto a driving model in hardware design. Results in this direction have been significant, leveraging first multi-core parallelism and then recently moving toward heterogeneous architectures (e.g., multicore CPU coupled with GP-GPUs). However, these evolutions will not be sufficient in the long term. To maintain an exponential increase in computational efficiency, we will need to rely either on an unlikely breakthrough discovery in hardware technology, or on a fundamental change in computing paradigms.

This workshop is dedicated to experts who explore approximation in hardware and software from both a statistical and a deterministic viewpoint, as a computing paradigm shift to break the current performance and energy-efficiency barriers of systems at all scales, from sensors to supercomputers. Approximate computing is a viable method for building more efficient, scalable and sustainable systems. However, it also places formidable challenges across the entire computing software and hardware stack. Addressing these challenges requires balanced expertise in mathematics, algorithms, software, architecture design and emerging computing platforms. The objective of this workshop is to bring together experts across these areas to present the latest findings and discuss future opportunities for approximate computing. In more detail, the workshop will cover the following areas:

  1. Approximate and transprecision computing: from the physical limits to the architecture and circuit design; from the algorithm design to the error analysis; from innovative technology to real applications.
  2. Programming abstractions: from structured and disciplined approximation in computation, communication and data transfers, to quality control and techniques to recover from over-approximation.
  3. Computing platforms: from tiny low-power devices for IoT applications, up to classical HPC systems embedding imprecise massively parallel accelerator.
  4. Applications: examples from data analytics, machine learning, deep learning, and scientific computing, where uncompromised quality with scalable order-of-magnitude time- and energy-to-solution reduction is reachable relying on approximation for a significant amount of calculations.

Key Topics

The workshop will cover the following key topics:

  • Beyond Moore’s law
  • Future challenges for programming models and languages
  • Exascale Systems

The workshop provides an opportunity to have in-depth discussions, presentations, and interactions on these topics. This will promote future collaborations and better coordination around the development on approximate and transprecision computing techniques.

Expected Outcomes

  • Promote research and development in approximate and transprecision computing
  • Align developments in algorithms, software, and hardware design towards unified and successful platforms for approximate and transprecision computing
  • Foster a common discussion across multiple disciplines
  • Raise energy-awareness in the big data community as well as in HPC
  • Promote collaboration between academia, industry and SMEs
  • Strengthen the community in energy efficient computing

Agenda

The workshop will be held in the afternoon of July 2, between 14:00 and 18:00.

The full agenda is detailed in the following table.

START END Duration FRIDAY, JULY 2 2021
14:00 15:00 01:00 Invited Keynote: “Leveraging open source HW for transprecision computing research”, Frank K. Gurkaynak, ETH – Zurich, Switzerland
15:00 15:30 00:30 Invited Talk: “Approximate Computing: Test and Reliability issues and opportunities”, Alberto Bosio, École Centrale de Lyon, France
15:30 16:30 01:00 Invited Keynote: “Deploying Deep Neural Networks in in the embedded space: Challenges and Opportunities”, Christos-Savvas Bouganis, Imperial College, UK
16:30 17:00 00:30 Coffee Break
17:00 17:30 00:30 Invited Talk: “On the Resilience of Heterogeneous Memory Systems: Extended Margins, Errors, and New Opportunities for Approximation”, Dimitrios S. Nikolopoulos, Virginia Tech, USA
17:30 18:00 00:30 Invited Talk: “Opportunities for Approximate vs Transprecision Computing in High Performance Scientific Computing”, Enrique S. Quintana Ortí, Universitat Jaume I De Castellon, Spain

 

Talks

Leveraging open source HW for transprecision computing research

Frank K. Gurkaynak, ETH – Zurich, Switzerland

About 5 years ago when we first started setting up the OPRECOMP project on Transprecision Computing, at ETH Zurich we were at the early stages of our involvement in open source hardware through our PULP project. In only five years, open source hardware has evolved at a tremendous pace, and we are proud to have been part of this movement. In this talk, I will briefly introduce what PULP and open source hardware is and explain how this approach has helped shape the results of OPRECOMP with concrete examples.

 Approximate Computing: Test and Reliability issues and opportunities

Alberto Bosio, École Centrale de Lyon, France

Approximate Computing (AxC) is today one of the hottest topics related to system design and optimization. Thanks to this computing paradigm, designers are able to reduce area, power consumption, and even production costs in the case the target application can accept a given degree of inaccuracy in the final computations. This presentation discusses the impact of Approximate Computing on the test and reliability. More in particular, it aims at showing that it is possible to use Approximate Computing to implement low cost but still efficient test mechanisms and fault tolerant architectures to be used in safety-critical applications.

On the Resilience of Heterogeneous Memory Systems: Extended Margins, Errors, and New Opportunities for Approximation

Dimitrios S. Nikolopoulos, Virginia Tech, USA

Memory systems are becoming increasingly heterogeneous, both in terms of the device technologies used to synthesize the main memory of a computing systems, and in terms of their inherent operating parameters. We will explore new results that reveal maximal operating margins for both DDR and NVM technologies and show how operating systems and virtual machine monitors can leverage heterogeneity, non-uniform access and persistence to improve performance and energy-efficiency, as well as scale memory capacity to meed the demand of data-intensive applications. We will also explore how new memory technologies  create new opportunities for approximate computing, particularly in the domain of deep learning

Deploying Deep Neural Networks in in the embedded space: Challenges and Opportunities

Christos-Savvas Bouganis, Imperial College, UK

Deep Neural Networks have attracted the attention of researchers and practitioners from various fields, as they have shown that can result in systems with performance that matches or even exceeds that of humans. However, their excessive computational requirements oppose challenges when such models need to be deployed In the embedded space, where resources are sparse and energy and power consumption of the compute platform are crucial for their  successful deployment. The talk will discuss methodologies and current approaches for addressing the above challenges, highlighting the opportunities in that space.

Opportunities for Approximate vs Transprecision Computing in High Performance Scientific Computing

Enrique S. Quintana Ortí

The convention in scientific computing is to employ IEEE double-precision (64-bit) arithmetic for all computations involving floating-point data. Nonetheless, appealing benefits from the adoption of mixed precision schemes have been reported for the solution of dense and sparse linear systems on graphics processing units (GPUs) via iterative refinement.In this talk, we will illustrate the benefits of a generalization of the mixed precision strategy, known as Transprecision Computing (TC), in terms of execution time and energy efficiency. For this purpose, we will employ several case studies arising in the iterative solution of sparse linear systems on GPUs,with codes currently integrated the Ginkgo library (https://ginkgo-project.github.io).In some detail, this research effort exploits the fact that, for sparse linear algebra operations, the cost is dominated by the memory accesses while the arithmetic is largely irrelevant. To leverage this property,the Ginkgo solvers store certain parts of the data in reduced precision in memory, but operate in “full” 64-bit precision in order to bound the accumulation of rounding errors.Reduced-precision storage can be leveraged to maintain approximation operators, such as a preconditioner, or in a solver thatgradually augments the precision of the operands as the iteration converges to the solution.

Organising Committee and Contacts

Cristiano Malossi
IBM Research GmbH
Zürich, Switzerland
Prof. Luca Benini
ETH Zürich
Switzerland
Prof. Norbert Wehn
Dr. Andrew Emerson
CINECA
Italy

Register Here

Please register to ISC 2021 workshop at this link: https://www.isc-hpc.com/registration-2021.html