I am searching for parts for a Ryzen PC-Build for my father and have a question concerning the RAM-modules.
The build will contain either a Ryzen 3300G or 3600G CPU.
(If a had to build the PC today I would buy a 2400G, but the is no need to hurry so I will wait for the 7nm 3rd Gen.)
The PC is supposed to do mainly video editing, recording, file-archiving and office tasks. No Gaming and no overclocking. The OS will be Win10x64 on a Samsung 970 Evo.
The plan is to buy 2x8GB modules.
I understand that Ryzen benefits a lot from fast RAM so here is my question:
Should I rather focus on higher frequency or lower timing when searching for adequate modules ?
I believe Picasso will still be on 12nm as the APU’s have been trailing a generational release schedule.
That said, the reason to get fast RAM on an APU isn’t for CPU performance since it only has one CCX, but for graphics performance. Since he won’t be doing any gaming I’d just stick to whatever 2666 dual stick kit fits his use-case (8GB or 16GB).
For gaming, the second kit (3000MHz CL16) would probably (maybe?) be measurable faster (by low single digit percentages when not GPU limited). Not that it matters in any way.
If you’re not going to OC it, and just want a rule of thumb, take the advertised frequency and divide it by the first number in the timings. Bigger is better.
e.g. DDR4-3000 CL16-18-18-38 … vs … DDR4-2400 CL15-15-15-35:
3000/16 = 188
2400/15 = 160
188 > 160
So, all other things being equal, get the DDR4-3000 CL16/18/18/38 — because it is (very) roughly (188/160 - 1 =) 17.5% better than the DDR4-2400 CL15-15-15-35.
Since Infinity Fabric scales with memory frequency, if the above calculation gives you a result that is a draw, or is within about 5% of each other, go with the higher frequency option. It will help a bit in video editing — but only a bit.
If you want to calculate actual latency, I’m pretty sure you need to halve the advertised frequency because it’s DDR. Thus DDR4-3200 has an actual (I/O bus) frequency of 1600 for the purposes of that calculation.
Single Data Rate (SDR) memory transfers both control signals and data on the rising edge of the clock signal.
Double Data Rate (DDR) memory transfers data on both the rising and falling edges — all the time (not just in burst) — but only transfers control signals on the rising edge. Since CAS latency is based on control signals… that means you halve the advertised (DDR) frequency.
Take DDR4-1866 (0.534ns cycle time). A 14-14-14-34 kit would then exhibit a true latency of arround 15ns.
Then we take some insane memory kit, DDR4-4700 (0.212ns cycle time). However at 19-26-26-46, the 19ns true latency before the data burst happens kills it in responsiveness.
I am just ignoring how the precharge and active command latencys here (they don´t really matter in this context anyway).
Does anyone recommend a Computer Science (actual paper) book that would get into this type stuff, even if its somewhat old (being since its harder to find things on paper)… Some stuff never changes (relatively), like “what are memory timings” “how to measure latency”. Also I would love a good recommendation on a “Intro to Processors”
I know I can google (I do), I know I can read Amazon reviews (I do). I am curious of opinions of the people of this community. A book I really enjoyed going thru recently was “A Practical Guide to Linux” by Mark Sobell, its quite old, 90’s old, but when it came to the core of how things work, the basics, it really was awesome. Just to make my point.
Yeah, a book that covers memory in depth that you enjoyed… Let me know
I still just love myself a paper book on computers, just can’t let go. Its like tobacco, doesn’t really make sense, they are a hassle, they are heavy, they take up to much room, are generally instantly obsolete, or outdated, but I just cant give them up, I have tried so many times.