James Demmel

The New BLAS Standard. The BLAS (Basic Linear Algebra Subroutines) Standards Committee has reconvened to consider adding new functions requested by different user communities, including batched BLAS (eg many small or medium sized matrix multiplications done at once), low precision BLAS, and reproducible BLAS (bitwise reproducible results, independent of number of processors, summation order, etc).

Communication-Avoiding Algorithms

The ASPIRE project (Algorithms and Specializers for Provably Optimal Implementations with Resilience and Efficiency) is motivated by the transition from a post-transistor-scaling world to one where whole-stack codesign is the key to improved efficiency. It follows on the ParLab (Parallel Computing Laboratory), an interdisciplinary research center motivated by the multicore revolution.

The Complexity of Accurate Floating Point Computation. See below for recent talks.

Proposal to update the LAPACK and ScaLAPACK linear algebra libraries, in pdf or postscript.

CITRIS is the Center for Information Technology Research in the interest of Society, a new $300M multicampus research center with the mandate of creating information technology to tackle society's most critical needs. Initial CITRIS research will focus on energy efficiency, transportation, disaster response, health care, education and environmental monitoring.

SUGAR is a simulation tool for MEMS (MicroElectroMechanical Systems), tiny sensors and actuators that form a pillar of CITRIS's approach to solving large scale societal problems. SUGAR is used by many researchers in the Berkeley Sensor and Actuator Center (BSAC).

BeBOP is Berkeley Benchmarking and OPtimization Group. We work on automatic performance optimization of numerical kernels, tuning them to run as fast as possible for particular architectures and sometimes particular inputs.

Millennium is a campus-wide project to provide hardward and sofware facilities for high performance computing. My activities are to provide algorithms, software and advice to best facilitate computational science and engineering on campus.

Templates for the Solution of Algebraic Eigenvalue Problems: This book gives a unified overview of the theory, algorithms and practical software for eigenvalue problems. It organizes this large body of material to make it accessible for the first time to experts and non-experts who need to choose the best state-of-the-art algorithms and software for their problems.

LAPACK - Linear Algebra PACKage for high performance workstations and shared memory parallel computers. The LAPACK Manual is available on-line.

(This and much other useful numerical software is available on Netlib.)

ScaLAPACK - Scalable Linear Algebra PACKage for high performance distributed memory parallel computers, The ScaLAPACK Manual is available on-line.

Proposal to update the LAPACK and ScaLAPACK linear algebra libraries.

SuperLU, implementations of sparse Gaussian Elimination for high performance parallel machines

Templates for the Solution of Linear Systems: Building Blocks for Iterative Methods is a hyper-text book on iterative methods for solving systems of linear equations. A similar book project for eigenvalue problems is underway.

Communication-Avoiding Algorithms: Please look above under Research Projects and Books for the most recent talks on this topic.

Der Zahlenlehrling / The Numerical Apprentice, a poem (in ppt) in memory of Gene Golub, presented at the Householder Symposium 2008

A Celebration of William Kahan and Beresford Parlett, a speech given at Bay Area Scientific Computing Day 2008, in PowerPower2003 (8.6MB) or pdf (3.6MB).

The Future of LAPACK and ScaLAPACK (in powerpoint), presented at PARA 06 meeting, Umea, Sweden, June 2006

Towards Accurate Polynomial Evaluation, or When can Numerical Linear Algebra be done accurately? (in pdf), presented at Banff meeting, October, 2005

The future of numerical linear algebra: Automatic Performance Tuning of Sparse Matrix Kernels, and The Next LAPACK and ScaLAPACK. In powerpoint.

The Cost of Accurate Numerical Linear Algebra, or Can we evaluate polynomials accurately?
In pdf. Slides presented at IWASEP 5.

The Complexity of Accurate Floating Point Computation, or Can we do Numerical Linear Algebra in Polynomial Time?
In postscript or pdf. Slides presented at the Householder Symposium XV, 2002 and IWASEP 4. Earlier version presented at ISSAC 2001.

Scaling in Numerical Linear Algebra, Scaling to New Heights, Pittsburgh Supercomputer Center, 20 May 2002 (in PowerPoint)

CITRIS Seminar, 7 Feb 2002 (in PowerPoint)

Guest Lecture on Dense Linear Algebra for CS267 (Fall 2001, in PowerPoint)

High performance Information Retrieval and MEMS CAD on Intel Itanium (in PowerPoint) UCB EECS Department Talk, Oct 12, 2001. (Version from 11 Dec, 2001)

CITRIS - Center for Information Technology Research in the Interest of Society (in PowerPoint) UCB EECS Department Colloquium, Oct 10, 2001.

The Complexity of Accurate Floating Point Computation in postscript or in pdf.

Recent Progress in High Accuracy and High Performance Linear Algebra Algorithms May 4, 1999. Solving systems of linear equations, and finding eigenvalues of symmetric matrices are standard problems that arise in many mathematical modeling problems. Solutions have existed for a long time, but these solutions are no longer good enough, because the demands of applications keep growing to larger problems, and to more ``sensitive'' problems that require higher accuracy. In addition, the larger problems in turn require large parallel computerswhere communication is much more expensive than computation.
We highlight several new algorithms under development, all of which have solved linear systems and eigenvalue problems many times larger, more sensitive or more quickly than before. Though these algorithms are all very different, one common mathematical approach distinguishes them from their predecessors: The new algorithms deliberately throw away just enough information about the problem to run both efficiently yet accurately. In contrast, their predecessors would either have to use much higher precision, send many more messages, or do many more operations to get the same answers.

Recent Progress in High Accuracy and High Performance Linear Algebra Algorithms May 13, 1999. This version of the above talk was an invited presentation at the 1999 SIAM Annual Meeting.

Survey of Parallel Numerical Linear Algebra Libraries Aug 20, 1997. This survey of dense and sparse parallel numerical linear algebra libraries covered a variety of available software for dense and sparse linear algebra problems on parallel computers, including LAPACK, ScaLAPACK, SuperLU and others. It was presented at a short course for NPACI (at the San Diego Supercomputer Center).

In Fall 08 we started a Designated Emphasis in Computational and Data Science and Engineering (CSE), which operates as a "graduate minor". Over 100 faculty from 20 departments and programs are participating.

Math 128a is a one semester upper division course on numerical analysis.

CS 170 is a one semester upper division course on efficient algorithms and intractable problems.

Math 221 is a one semester course in matrix computations. My text for the course, Applied Numerical Linear Algebra, was published by SIAM in August 1997.

CS 267 is a one-semester graduate class in Applications of Parallel Computers.

Computational Science Education Project provides an on-line text book on high-performance computing. I co-authored the chapter on Parallel Numerical Linear Algebra

Math 55 is a one-semester undergraduate course in discrete mathematics, offered Spring semester 1997.