![[IMG]](../../../assets/img/network-switch.B.jpg)
Rebuilding ATLAS for the AMD 64 3500+ CPU
Following the disappointing performance of ATLAS on the HP Pavillion a4313w system, I moved the disc and some RAM from that system into the Acer I got from Mildred.
The two systems originally had the following RAM configuration. Both systems were populated with 3 GB RAM, which at the time that was the maximum amount of memory Windows could address:
- Acer: 4 RAM slots populated with 1 x 2GB, 1 empty, 2 x 512 MB; total 3GB
- HP: 2 RAM slots populated with 1 x 2GB and 1 x 1 GB; total 3 GB
I determined 4 GB was the maximum the Acer could handle, so I moved the 1 GB RAM module from the HP into the Acer, ending up with the following:
- Acer: 4 RAM slots populated with 1 x 2GB, 1 x 1GB, 2 x 512 MB; total 4GB
- HP: 2 RAM slots populated with 1 x 2GB and 1 empty; total 2GB
On this system ATLAS built in only 28 minutes. Rather to my surprise (again!) this build performed more poorly than the stock ATLAS. The three trials (n=7296) netted results of 2283, 2288, 2301, for an average of 2,291 MFLOPS.
However, the number I have recorded (2,472) is from an older run, and may even be from a Debian-based distribution. To see what the Fedora 28 ATLAS can do, I installed the RPM, recomplied HPL to link against it, and re-ran the benchmark. The RPM contains two versions of the library, “serial” and “threaded.” Given that there’s only a single core CPU in the system, I decided the threaded version probably would not work, and so used the serial one.
Using the Fedora-provided ATLAS, the three trials (n=7296) had results of
2077, 2079, and 2078, for an average of 2,078 MFLOPS.