ConclusionFor this article, we completed nearly 500 individual benchmark runs to paint a more complete picture on the DDR memory situation with Coffee Lake. We had no problems reaching our targets thanks to Gigabyte's fine Z370 Ultra Gaming and G.SKILL's outstanding Trident-Z memory. First of all, the scores are sometimes very close together, within the margin of error. That explains why some results appear to be out of place in terms of their position in a very narrow performance band. This is as expected and can't be avoided when doing testing on a system like Windows 10, with background processes active and using a timer resolution in the order of 30 ms.
Also, in case you aren't aware of it yet: Memory timings are expressed as cycles relative to the clock frequency (not absolute time). For example, "CL12" at DDR4-2666 means "9 nanoseconds". Whereas on DDR4-3200 the same CL12 is just 8 nanoseconds - the actual time gets shorter, which might be too quick for the memory chips. That's why the vendors' rated latencies increase as memory speed increases. To achieve the same 9 nanoseconds absolute time you'd have to use CL15 on DDR4-3200 (9.38 ns).
The minimum memory frequency we would recommend for a high-end Coffee Lake system is 2666 MHz. Everything below that seems to have relatively large performance effects. The cheapest DDR4 memory runs at roughly $8 per GB (2133 MHz, CL15). While cheap, that is not the way to go. Rather, get a 2666 MHz kit with CL15, or better CL14. These should be in the range of $9 per GB, which makes for an extra $15 (for 16 GB) that is well spent.
If you have a bit more money to invest, the next good option is 3200 MHz memory, with CL14 or CL15. 3200 MHz CL16 is roughly equal to 3000 MHz CL14 in speed, so consider that option too. On AMD Ryzen processors the InfinityFabric (which links the cores) is clocked at 0.5X DRAM frequency, so you had to prioritize memory frequency over latency. Coffee Lake, on the other hand, also sees good gains from improved timings, which often makes it more sensible to buy lower latency memory than to go for the highest clock speed you can find.
Last but not least, the question of 1T vs. 2T seems to be settled with measurable gains in the ~0.5% range when averaged, which are really small and in my opinion not worth the trouble of paying a lot of extra for, or fiddling with BIOS options for hours just to go from 2T to 1T.
On the topic of super-high-speed memory kits, the gains here are limited, too, often below 1%. So spending all that money on a faster graphics card instead should be more useful, at least if you are a gamer. Modders should also sign up for higher-speed memory which often come with better visuals and lighting options, ie. RGB, etc. Should you be a competitive overclocker, then every little bit of performance matters and money should be no object.