NAG's SMP Library now optimised for the IBM POWER4

Oxford UK – September 2002. The world's leading provider of mathematical and statistical algorithms, the Numerical Algorithms Group (NAG), announces that it has optimised its SMP (Symmetric Multi-Processor) Library for the IBM POWER4 architecture.

IBM provided early access to a POWER4 system at the Hursley Solution Partnership Centre in the UK. "The support provided by IBM and the performance of the POWER4 system were excellent," commented Stefano Salvini, Senior Technical Consultant at NAG.

Some of the results from extensive work are listed below:

On the IBM eServer pSeries p690 system with 1.1 GHz POWER4 processors:

  • Common dense linear algebra factorization routines (eg LU, Cholesky and QR), and their associated solvers, sustain a performance of up to 2.8 GFLOPS on a single processor, and scale in performance up to 21 GFLOPS on 8 processors [1].
  • The QR-algorithm symmetric eigenproblem routine (F08JEF/DSTEQR) achieves 2.4 times the performance on 1 processor, and 16.6 times the performance on 8 processors, of the algorithmically equivalent LAPACK routine DSTEQR [2].
  • The QR-algorithm Singular Value Decomposition (SVD) routine (F08MEF/DBDSQR) achieves over 6 times the performance on 1 processor of the algorithmically equivalent LAPACK routine [3].
  • While the standard LAPACK code does not scale in performance on multiple processors, the SMP routine scales by, for example, a factor of 3.7 on 4 processors [4].

Brian Ford, Director of NAG said, "We were seeing demand from users of the SMP Library to produce an optimised version of the library on IBM POWER4. We are delighted to see how IBM and NAG work together to respond to our user requirements. NAG is committed to ensuring that users of our numerical libraries can utilise the products on IBM and other vendor platforms."

[1] Performance may depend upon problem size. Values quoted are typical for problems of order N=2000 or higher.

[2] Example quoted is for a problem of order N=4000.

[3] Problem of order N=1000

[4] Problem of order N=2000

About NAG
With over thirty years of experience in developing mathematical and statistical components, compilers and tools plus 3D data visualization software, NAG and its products have a worldwide reputation for accuracy, robustness and reliability. Over 10,000 organisations spanning the globe rely on NAG’s software and services to enhance their applications. NAG software is used by business, industry and education to solve problems, however complex, in areas such as research, engineering, life and earth sciences, and financial analysis. NAG has its headquarters in the UK and has offices in Germany, Japan and North America, plus distributors worldwide.

About IBM
IBM is the world's largest information technology company, with 80 years of leadership in helping businesses innovate. Drawing on resources from across IBM and key Business Partners, IBM offers a wide range of services, solutions and technologies that enable customers, large and small, to take full advantage of the new era of e-business. For more information about IBM, please visit www.ibm.com. For more information about IBM eServers, please visit www-1.ibm.com/servers/.

' NAG is the registered trademark of the Numerical Algorithms Group Ltd.

IBM, pSeries are trademarks of International Business Machines in the United States, other countries, or both

Please understand that product prices, specifications and other details are current on the day of issue of the press release. These details, however, may change thereafter without notice.

For more information, please see our Press Contacts page.