Intel E5-2600 Series "Sandy Bridge-EP" Xeons
Posted on: 03/06/2012 05:41 PM

Matrix Multiply v3.5

This Matrix Multiplication benchmark computes the product of 2 square matrices of the same dimension. It can achieve near theoretical peak performance, even on modern processors. It is a key operation for realizing high performance in many of the standard tasks done in scientific computing. This includes solving large dense systems of linear equations (as in the Linpack Benchmark), and QR factorizations (to solve least square problems). Solving large dense linear systems is a key component for electro-magnetic simulations used in stealth design applications. The computation is done by the dgemm subroutine in the Intel Math Kernel Library (MKL).

The default workloads are square matrices of size 5000 to 15000 by steps of 5000. By default each multiplication is run 2 times and the average value is reported.

Nothing surprising here. More of the same from Sandy Bridge-EP. Rumor has it that there is a new build of Matrix Multiply that is optimized for Sandy Bridge-EP's new AVX instructions. For this review I stuck with the version I've used for quite some time now, but I'll try to get a hold of the AVX build next time around.

Matrix Multiply v3.5 Power

MyriMatch

MyriMatch is a tool designed to take experimental data from shotgun proteomics experiments and compare those spectra against sequences in a known database of proteins. Whether the program is being run in a single-computer environment or across an entire cluster of processing nodes, it is able to optimally divide work in a much more efficient way than many other database search programs. This is because it only generates candidate sequences from the known database once for the entire set of spectra instead of once for every spectrum. Thus, for each candidate sequence generated, it is compared against every spectrum. The spectra keep a certain (user-defined) number of candidate sequences that had the highest scores.

MyriMatch is designed to take advantage of (symmetric) multiprocessor systems by multithreading the database search. A search process on an SMP system will spawn one worker thread for each processing unit (where a processing unit can be either a core on a multi-core CPU or a separate CPU entirely). The main thread then generates a list of worker numbers which is equal to the number of worker threads multiplied by this parameter. The worker threads then take a worker number from the list and use that number to iterate through the protein list.

I've proven in previous reviews that MyriMatch loves clock-speed and hates hyper-threading. Luckily, here we are comparing two generations of processors, both with hyper-threading enabled. The performance difference lies solely on the number of cores and the process improvments of Sandy Bridge-EP.

MyriMatch Power




Printed from 2CPU.com (http://www.2cpu.com/contentteller.php?ct=articles&action=pages&page=121,5.html)