Experiments with DDR4 Memory Timings

Trying to choose and configure the "best" memory for a PC can be very overwhelming. The manufacturer specifications for your CPU and motherboard will indicate an upper limit for the memory speeds, but the online stores are flooded with memory modules that far exceed these limits. It can be hard to know whether these modules will even work, or how much they will improve system performance.

I figured that a fast way to determine a "maximum expectation" of performance improvement would be to actually reduce my installed memory performance (deliberately underclock) and then plot how much performance degrades. Presumably if I then purchase new memory which can be overclocked then it is reasonable to expect a performance increase on a similar trajectory.

So I ran all of the benchmarks with my memory clocked at all frequency intervals that were available and did not greatly exceed the rated capacity for my CPU, and these were the results:

git status

(3466M/T is 9.3% (1.1x) faster than 1333M/T)

git log

(3466M/T is 11.4% (1.1x) faster than 1333M/T)

ripgrep

(3200M/T is 6.9% (1.1x) faster than 1333M/T, 3466M/T appears to be an invalid result)

pytest - empty test

(3466M/T is 10.3% (1.1x) faster than 1333M/T)

pytest - single test

(3466M/T is 11.0% (1.1x) faster than 1333M/T)

mypy

(3466M/T is 11.0% (1.1x) faster than 1333M/T)

Summary

Other than the broken ripgrep result, massively underclocking the RAM to 1333M/T (the slowest speed supported by my motherboard) couldn't slow the 3200G CPU by more than 12%. Keep in mind that no one would ever run memory that slow - the incredibly cheap 32GB kit that I bought for less than $100 is rated for 3200mhz, and if I left the motherboard speed selection on "auto" it would run at the maximum supported speed of the 3200G which is 2933M/T. Therefore overclocking my memory to 3466mhz only provides a marginal improvement over the default of 2933M/T or 3200M/T.

There are many more factors to consider in terms of RAM performance, but it seems unlikely that a very expensive memory kit would increase this CPU's performance by more than 10-15%. This means that the only way to significantly increase this system's performance I need to install a faster CPU.