Do you think another 4-core ARM instance on the Oracle cloud would perform the same? Each person gets one free instance but unfortunately Fido didn't. Doesn't Larry know dogs are people too?After studying the output of the Ryzen 5800X3D Fido's tail stopped wagging and the dog developer admitted there was a downward trend over the entire runtime. As an alternative to shoddily constructed on-premise computing resources, I decided to run the same tests on the free 4 core ARM instance in the Oracle cloud.The greatest average Flops rate for the Ryzen 5800X3D occurred with an array size of 767. As averages can be deceiving, I decided to plot the time evolution of Flops over the first 100 seconds of execution to see how the two runs compare.It looks pretty close, but not quite as consistent as the Pi 4B, Celeron or Phenom graphs. I think additional tests need to be made once Summer arrives. I wonder how consistent a laptop would look.
Here is a set of runs comparing two identically configured dual-socket Epyc 7551 servers.
The results for both systems are satisfyingly similar and visually overlay one another. The greatest throughput occurs with an array size of 768. Looking at the time-series for that size yields
There is a slight decaying boost at the beginning of the run, but not as dramatic as the decrease with the 5800X3D. I think this means the heat sinks are working well.
Showing the graphs to the dog developer was met with barking about turning off unnecessary service tasks before going to the Pi races. Since these are not virtual machines, that 10-second periodic drop must be caused by the operating system. I blame SLURM, but this has not been confirmed.
I wonder if one can see the effects of the fan turning on and off on the active cooler in the sustained Mflops per core plots for the Pi 5.
Statistics: Posted by ejolson — Tue Apr 16, 2024 6:44 pm