For the build

  • MSI Z170A GAMING PRO (Z170 3xPCI-E DDR4)
  • Intel i5-6500 3.20GHz 6MB BOX
  • A-DATA 240GB 2,5'' SATA SSD Premier SP550
  • Monitor: Eizo FlexScan S1901 (1280 x 1024)
  • GeForce GTX960 4096MB 128bit OC (Armor 2X)

What memory would you buy, and why?

Specific problems I consider:

  • Currently sold processors support up to 2133 MT/s (or did I miss some model?), but (I guess) the faster dices work as well (with reduced speed), so maybe ex. 2400 if it has better CAS latency and not much of a price difference?
  • If I choose Kingston, should I go with furry or savage? I read, that savage is more overclocking friendly, but do not know what it means specifically.
  • I see the "PC4-17000" or "PC3-17000" often on dices sold, but on wikipedia, the PC3-17000 is under DDR3 section?
  • What should be the upgrade scheme? Incremental (one 8GB dice per year, for most dice 'recycling') or batch (2*4GB, than 4*4GB, than 4*8 GB to gain maximum of dual channel)?


My use scenarios are: programming in couple of IDE opened simultanously (60% time), 100+ Mb data analysis (10%) and games like Fallout 4 (30%).

  • Are you only considering Kingston or did I miss another contender somewhere?
    – Andy
    Commented Nov 22, 2015 at 1:32
  • @Andy Currently I consider Kingstone, but I'd be happy to find better alternative. Commented Nov 22, 2015 at 2:27
  • 2
    PC3-17000 is DDR3, and PC4-17000 is DDR4, hence the different numbers in front of 'PC'.
    – Rubydesic
    Commented Nov 23, 2015 at 0:38

2 Answers 2


There are very few memory bound tasks for modern CPUs. The key factors to look for are capacity, cost, latency, voltages and speed, usually in that order. It doesn't really matter if a CPU doesn't support a RAM speed. You just need motherboard support and go into BIOS to activate the XMP setting, so that it'll run at the rated speed rather than the SPD speed.

Without any further details about your specific workload (which is very important), my general recommendation, besides not to get a Z170 board unless you really need the extra IO, Smart Sound, or want to get a K CPU (overclocking), is to get 16 GB RAM straight away. It is very easy to use up 8 GB of RAM if you do more than one thing at a time, or if you're using anything that uses a lot of RAM. The price of DDR4 has dropped as well, so it is reasonably affordable to get 2 x 8 GB kits, in fact, cheaper than 2 x 4 GB kits of DDR3 last year: The G.Skill Ripjaws 4 series has a 2400 MT/s 15-15-15-35 kit for $74.99 at newegg.com. G.Skill is one of the largest manufacturers of RAM, and has a reputation for good quality. Their Sniper series generally has slightly more headroom, but it's just going to shave a few seconds off a 10 minute compression or video rendering, to name some memory bound tasks.

Basically, "overclocking friendly" just means that the factory clocks are slightly less aggressive, and that you can increase them a bit more while still being stable. The PC* ***** naming is just a different way of naming DDR* ****. The number directly after the PC and DDRs is the version number, and the second number is the bandwidth in MB/s for the PC scheme, and the clock rate in MT/s (twice the real clock rate) for the DDR scheme. Since DDR RAM has a 64 bit wide bus and there are 8 bits in a byte, the PC number is just 8 times the DDR number.

It doesn't really matter how you upgrade, just keep in mind that using memory from different kits may be unstable. The chance is low if they are the same model, but it's still there. Since you only have 4 RAM slots though, I'd just get a 2 x 8 GB kit and another when you decide it's not enough. Good luck.

Addendum: On memory timings

Memory timings tell the CPU how long to wait before expecting an action to be completed on the memory chip, and in SDRAM (clocked + refreshed RAM), is measured in RAM clock cycles (which is half the DDR number). Memory is arranged in rows, which are divided in to columns.

CAS Latency determines the amount of time the CPU to wait after issuing a read command before actually reading the data from the output pin. If this is shorter than the time it takes for the RAM to respond fully, what is read is a voltage that is transitioning between the previous voltage on the output pin and the desired one. This is bad and sometimes causes the computer to crash.

But CAS Latency is the full latency only if the right row is already open! If there is no row open, the CPU has to issue an Activation command first, and wait. This is the next number, tRCD, the RAS to CAS delay.

The third number is tRP. The Row Precharge time is the time you need to wait between a Precharge command and an Activate command; the time it takes to close a row, if you have the wrong one open.

Finally, the last number is basically CAS + tRCD, and then a bit more. It's the minimum amount of time a row can be active. There's a bit more again to let the memory finish passing on the voltage.

  • Thanks for the answer. My requirements are: programming in couple of IDE opened simultanously (60% time), 100+ Mb data analysis (10%) and games like Fallout 4 (30%). I'd like to futureproof the MoBo and would like to overclock the i5-6500 (not K). Commented Dec 6, 2015 at 12:45
  • @PiotrFalkowski I'm not entirely sure it is enabled with non-K CPUs, but with Skylake, just as with Nehalem, the PEG and DMI clocks are fully separated, and you should be able to overclock the system bus with no problems. I can see why you're looking for the locked i5s. The price of the 6600K is vastly inflated unless you happen to be lucky enough to live near a microcenter. I'd still suggest 16 GB RAM, but that might just because Brackets + 50 tabs Chrome uses heaps of RAM comparatively. I wish I had a second (or bigger) monitor though.
    – timuzhti
    Commented Dec 7, 2015 at 0:35

The clock speed of RAM doesn't affect the performance of the computer much (if at all). Running a RAM at a faster speed would increase the CAS latency, which is the amount of clock cycle lag the RAM has (lower is better). But the advantage would be a faster transfer speed. Decreasing the speed means a lower CAS latency but a slower transfer speed. For normal purposes, the different clock speeds would not have much of a difference. But if you plan on running super high performance RAMDisk (or similar applications) a faster RAM would benefit you significantly. When I choose RAM, I buy the cheapest RAM that is from a trusted brand and is aesthetically pleasing (has a nice looking heat spreader).

To choose between the Kingston Fury and the Savage, it would depend on what you want to do with the RAM. If you have a higher budget and want more overclock, go with the Savage. If you have a lower budget and/or will not overclock your RAM, go with the Fury. They really don't have a huge difference.

You really shouldn't have a upgrading "scheme". I would usually leave 2 or more slots empty (unless I'm maxing out my RAM right when I build my computer) so I can upgrade it later. I only upgrade it when I need it (e.g. the stuff I do requires more RAM) and not in a "schedule" and you should do this too.

  • Is it correct than, that a dice with 2133 Mhz CL15 will have a response time 1/2133000000 * 15 = ~7 ns and 2400 MHz CL12 1/2400000000 * 12 = 5 ns? Second question: Is it the CPU cycles we are waiting on, so having CPU clock twice as fast as the RAM clock would cut the CAS latency in half? Commented Nov 30, 2015 at 13:05
  • 2
    Answer to question 1: You are correct. Answer to question two: Taken from Wikipedia: "Synchronous dynamic random access memory (SDRAM) is dynamic random access memory (DRAM) that is synchronized with the system bus." So increasing the clock of the system bus will decrease the CAS latency, but overclocking the system bus is very dangerous.
    – Peter Zhu
    Commented Nov 30, 2015 at 14:01

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.