September 23-28, 2018
University Residential Center
Bertinoro (Forlì-Cesena), Italy

ScalPerf aims at taking an integrated look at the opportunities and constraints on the road to ever higher performance and productivity of computing systems. Distinguished researchers are invited to exchange their perspectives on different areas that can contribute to scalable computing.

The workshop will include invited talks and discussion sessions. The talks can cover established results as well as brain storm on future research directions. The discussion sessions are two, each consisting of a first part to sharpen the formulation of the issues and a second part to explore answers. This year, the topic of discussion is

Linear Algebra for Machine Learning: Algorithms and Accelerators

In the past few years, machine learning has extensively impacted many parts of society, including elections, security, surveillance, engineering operations, scientific research, social media and commerce, etc... Considering that today only about 20% of the digitally available data is subject to some form of mining, it is likely that the impact of machine learning on society will grow over the next several years. Previous editions of ScalPerf have paid attention to some aspects of this trend. ScalPerf'15 included extensive discussions on the state of the art of machine learning of the time, while ScalPerf'11 had speculated on the role of accelerators like GPUs.

The sparse linear algebra runtime, GraphBLAS, is quite useful to efficiently and easily express a variety of AI algorithms, including deep learning. It provides a separation of concerns between the the development of high performance runtime for a particular hardware architecture and the development of algorithms. ScalPerf'18 discussions will focus on the state of the linear algebra approach for machine learning and AI, hardware accelerators (viz. GPUs and beyond), estimate of performance loss due to the use of runtime library that is by and large oblivious to the application at hand, comparison of the sparse linear algebra techniques to those developed for engineering problems. We will also reflect on which characteristics (e.g., bandwidth) constrain the performance of today's system.

ScalPerf'18 attendees are encouraged to address one or more of these issues in their presentations.

Participants will be also be invited to submit suggestions for specific issues to be included in the discussion, a few weeks before the workshop.

In addition to the invited speakers, participation of other interested researchers, particularly graduate students and junior researchers, is welcome. An expression of interest can be sent to, including a brief CV.