[Home] [Version 2] [Run it Now] [Download] [Results] [FAQ] [About] [Resources] [Credits]

(12/10/18) Checksum modules

New logic and checksum modules have been added to all four kernels to guard against over-aggressive and speculative optimizaiton. The results of each kernel are combined in a global checksum with varies with the precise number of iterations and repitions given by the specified time granularity. (Thanks to Lukasz Kuszner for feedback.)

(11/27/00) Bug fix

Tiny bug in off-line version reported FFT performance twice, but did not affect our posted results. Version 2.0a of scimark2lib.zip, scimark2lib.jar, and scimark2src.zip have this fixed. (Thanks to Robert Gottlieb for pointing this out.)

(6/26/00) Improved displays for archived results

Because the number of SciMark submissions is large (currently nearly a thousand) the archived resutls can sometimes take too long to download and display the page. Therefore, the Results page now displays summary information of top scores for each operating system and Java Virutual Machine combination. A link for the complete archived results is provided.

(6/00) Verification of LU and FFT kernels

The Java applet and commandline version of SciMark now check the LU and FFT results to make sure they are reasonable. If any of these fail, a bogus score of 0.0 is reported. This is in place to discourage posting scores which may have cheated on numerical computations. (See the FAQ section for more details.)

(9/99) ANSI C version now available

The Java sources for SciMark 2.0 have been translated to ANSI C for performance comparison with other platforms. See the download area for details.

(4/99) SciMark scores are now reported in approximate Mflops

In the original version of SciMark, results were reported in terms of a performance ratio to a specific machine/platform (a 50 MHz Sun SPARC 10 running Netscape 4.04). This was done because some routines (e.g. Monte Carlo) did a fair amount of integer computations while others utilized transcendental methods which were difficult to count the actual number of floating point operations. Nevertheless, many of you expressed an interest in seeing approximate Mflops numbers, so we changed the metric accordingly. Using Mflops also lets us retain some continuity in scores when we run larger problem sizes.

(4/99) New kernels

• Fast Fourier Transform (FFT)
The FFT kernel has been completely rewritten. It is based on the Gnu Scientific Library (GSL) and uses a one-dimensional Radix-2 version in which both the input and output are complex (treated as a floating-point vector of real/imaginary pairs).
• Sparse Matrix Multiply
The Sparse Matrix Multiply kernel has also been rewritten. The previous version used a coordinate storage format and use a completely random structure pattern. While OK for small cache-contained matrix sizes, it is unrealistic for large problems. SciMark 2.0 uses a compressed-row storage scheme and a scalable structure pattern which more resembles application matrices.
• Monte Carlo integration
In the original version, we approximate the integral of exp(-x*x) between 0 and 1. In SciMark 2.0, we instead compute Pi by evaluating the area of a circle within the unit square. The change was to avoid calling Math.exp() and instead use a simple expression (e.g. x*x + y*y) to get a better count of floating point operations used.
• Successive Over-relaxation (SOR)
The Gauss-Siedel iteration kernel in the previous version has been replaced with a Successive Over-relaxation method.

(4/99) Source code now available

We include the actual Java source code for SciMark 2.0 so that one may use native Java compilers on it, as well develop C and Fortran versions of SciMark for comparison.

New automated updates to our database

SciMark 2.0 now includes an automated submission system (rather than the previous email-based version) to receive contributed scores. If you submit a score, it is updated on our database immediately.

Large out-of-cache version

A command-line version of SciMark 2.0 is available for download which can run large (out-of-cache) problem sizes. This version of the benchmark, SciMark 2.0 LARGE, can be useful in measuring the capability of the memory subsystem under a Java environment.

Portable random number generator

Most of the SciMark kernel use a random number generator to initialize vector and array data. In the original SciMark we utilized Java's own java.util.Random class, but now we include the source for a portable random number generator. This makes it easier to create C and Fortran versions of SciMark.

Commercial equipment and software is referred to on these pages in order to precisely document experimental measurements. Mention of these products does not imply recommendation or endorsement by the National Institute of Standards and Technology, nor does it imply that the products so identified are necessarily the best available for that purpose.