Linux Memory Performance With Intel Kabylake From DDR4-1600 To DDR4-3333MHz

[phoronix]

Phoronix: Linux Memory Performance With Intel Kabylake From DDR4-1600 To DDR4-3333MHz

For those that may be thinking about picking up an Intel Kabylake processor and trying to justify if DDR4-2400 memory is worthwhile for your budget, or even faster DDR4 memory via XMP profiles / overclocking, here are some tests using a Kabylake CPU and testing DDR4 memory at frequencies from 1600MHz up to 3333MHz.

http://www.phoronix.com/vr.php?view=24066

[riklaunim]

Did some tests some time ago - http://www.rkblog.rk.edu.pl/w/p/test...ormance-linux/ - dual channel helps a lot, while some benchmarks (Redis, Postgres) gave bit odd results.

[devius]

The compilation speed had a small but noticeable impact on the DDR4 memory frequency

Cool! I have to improve my compilation speed then

[Michael]

Cool! I have to improve my compilation speed then

Fixed that sentence

[cj.wijtmans]

Did some tests some time ago - http://www.rkblog.rk.edu.pl/w/p/test...ormance-linux/ - dual channel helps a lot, while some benchmarks (Redis, Postgres) gave bit odd results.

quad channel or "gtfo"

[Mark Rose]

In addition to quad channel, some have found that single versus dual rank sticks also make a difference in memory performance.

[cj.wijtmans]

In addition to quad channel, some have found that single versus dual rank sticks also make a difference in memory performance.

Interestingly though that quad channel get more bandwidth its latency may be worse. So its really a trade off on what you want. Personally i used "slower" memory because it has better latency. For tasks such as compiling bandwidth might be better though.

[dh04000]

I'm very surprised that the performance of the intel igpu wasn't improved more by the increasing ram memory speed. I also wonder how a nvidia dedicated gpu would be effected by the ram speed change? I imagine very little.

Same your money and buy more ram and more channels instead of faster ram in gaming I think is the lesson.

[dungeon]

GPUs likes faster RAM as VRAM as always

Looking at iGPUs only, suddenly TF2 is not CPU bound but RAM speed bound

[caligula]

Michael, did the compilation test use flash/hdd storage for intermediate results or a tmpfs? I didn't find this info from the test suite's description file. People compile kernels in tmpfs these days. Don't know about defconfig, but smaller configurations easily fit in RAM even on 8GB machines. For example my kernel and all the temp files is < 800-900 MB. The memory has a lot bigger impact with compilation fully done in tmpfs.