Kingmax Mars DDR2-667 KLCC28F-A8EB5
Since our last system memory review, we've refined some of our Linux memory benchmarking strategies to further focus on the system memory performance due to the lack of adequate alternative OS benchmarking programs at this time. Our revised arsenal for Linux memory benchmarking consists of the following programs - Doom 3, LAME (compilation & encoding), FreeBench, SPECViewPerf, and RAMspeed. Unfortunately, due to some Ubuntu 5.10 conflicts we were unsuccessful with our attempts at getting SPECViewPerf 8.1 to run appropriately for our tests, however, once SPECViewPerf 9.0 is shortly released it should correct these conflicts. With Doom 3 we used the traditional time demo 1 which is bundled with the retail copy and we ran this timed demo benchmark at 1280 x 1024 High Quality and 1280 x 1024 High Quality with 4x, 9-tap Gaussian/8x AF (specified through nvidia-settings). All other video settings in Doom 3 were left stock. Another one of our usual memory benchmarks is the LAME encoder for which we measured the amount of time required to compile LAME 3.96.1 using GCC and then also measured the amount of time to encode a 81.3MB wav file as an MP3 (time lame -h song.wav song.mp3). Onto one of the newly adopted benchmarks by Phoronix, we have FreeBench. FreeBench is a collection of different benchmarks that stress a variety of different points in a computer system. The specific FreeBench programs we've decided to utilize after an extensive amount of collaboration and testing are Analyzer, FourInARow, pCompress2, PiFFT, and Neural. Although Analyzer isn't the most recent of FreeBench programs, it's a program that examines access traces for data dependences and is designed to stress the memory bandwidth, memory latency, and also latency hiding capabilities. The FourInARow benchmark is coded to play a game of "four in a row" against itself. As the name implies, pCompress2 is a file compressor that utilizes a three-stage compression approach and among other system things is quite memory intensive. Next up with FreeBench we have the first of floating point programs and the purpose of PiFFT is to calculate PI to a 4,194,304 decimal places using a FFT (Fast Fourier Transform). Finally, ending off our assemblage of FreeBench memory benchmarks we have Neural. This application is a neural network doing character recognition of characters written out in ASCII graphics. Although we weren't able to run SPECViewPerf v8.1 today, due to software related conflicts, it's a portable OpenGL performance benchmark and happens to be one of our favorite for measuring workstation performance with the 3ds Max, Maya, and Pro/Engineer view-sets. Ending off the official memory benchmarking we have RAMspeed. Although RAMspeed hasn't been updated in quite a while, it serves as a truly terrific memory test for copy, scale, add, and triad. Although these tests are synthetic, they are reflective upon real world performance and are used within STREAM and Windows SiSoft Sandra. For the integer memory testing we used the ramspeed -b 3 -l 5 flags while for the floating point memory testing we used the respective ramspeed -b 6 -l 5. These flags specify the type and LongRun mode with five runs for acquiring the most accurate memory performance results. As always, all of the other variables relating to the performance were kept constant and, as usual, all other benchmarking practices were enforced.