How to Choose the Right Memory: A 2019 Guide To DRAM

[Admin]

We break down the easy shortcuts enthusiasts use to choose RAM that delivers the ultimate combination of performance, capacity and price.

How to Choose the Right Memory: A 2019 Guide To DRAM : Read more

 

[Soda-88]

I've bought a Micron E-die 3000 15-16-16 kit for my new Ryzen PC and got it from 42/24/40 GB/s with XMP settings up to 54/30/52 GB/s with fine tuned timings. It did take several hours of trial and error going through all the timings.
Detailed look at my settings is in my forum signature.

 

[tiggers97]

Great article!. But at the end of the article is a link to "Best Memory ". But some of the options are no longer available or headed that way. Is it possible to update the recommended "best memory" list?

 

[Crashman]

Great article!. But at the end of the article is a link to "Best Memory ". But some of the options are no longer available or headed that way. Is it possible to update the recommended "best memory" list?

We're continuously working to add newly awarded products to "Best Memory", but sometimes it takes a while to get old products removed.

 

[Anton Hunter]

Great article (actually, a great series of articles pertaining to memory). Thanks!
I purchased a G.Skills kit (F4-3200C16-8GVGB) dual channel/single rank and XMP 1600MHz 16-18-18-38-56. I couldn't get any stability at all at 3200 MHz, so I spent a few days reading up on the Ryzen/memory correlation and then another couple days going through the long process of discovering the most stable settings I could find...and after benchmarks and stress testing using HWINFO and 3DMark I ended up at 2933MHz and 14-15-15-15-34-51 via my ASUS Bios, and achieved a 97.5 stress score using Firestrike.
The process is long and can be tedious unless one has a mind-set filled with patience and the joy of discovery. I learned a lot, and am grateful for your and AMD's timely articles on the matter.
Thanks again!

 

[ilincarubogdan]

Hello. My mainboard asrock z97 Killer says it can use up to 3200OC ddr3 memory, then i saw that the intel 4690K processor works on maximum 1600mhz. Is that all i can get from ddram ? i'm a little confused

 

[Crashman]

Hello. My mainboard asrock z97 Killer says it can use up to 3200OC ddr3 memory, then i saw that the intel 4690K processor works on maximum 1600mhz. Is that all i can get from ddram ? i'm a little confused

I killed a reference my editor made to CPU specs specifically for this reason.
Intel's processors support high memory speeds
Intel does not support you doing this
Intel validates high-speed memory in its XMP program, then leaves that memory off its CPU compatibility list, therefore, you can't rely on Intel for accurate information regarding what your CPU supports.

In an effort to promote itself as a superior overclocker, ASRock put up memory speeds that you won't likely achieve unless you have perfect hardware.

I'd expect DDR3-2400 to work very easily. DDR3-2666 should work but is hard to find in 8GB-per-DIMM capacity. And I'd give you better than 50/50 odds of running DDR3-2933 successfully, but with the same problem of available capacity at that frequency you'd might as well stick to DDR3-2400.

 

[lubomirz]

Anton, let me react to your post from Saturday (note #5 here). I can't believe that kit rated at 3200 won't run at... 3200 you guessed it. Something else must went wrong in cooperation with your Ryzen5 1600, maybe a combination of motherboard and Ryzen itself (newest BIOS installed ?). I'm absolutely sure that Ryzens 2xxx would be running your memory kit at 3200 and most likely at 34xx/36xx.

Also, to all readers : notice how Anton spent FEW DAYS to make his 3200 kit run at 2933. Similar thing, spending couple of days, happens quite often to my friends who are trying to find the holy grail, the maximum speed their memory can run at. They spend dozens of hours testing, changing and re-testing their configurations... only to find that real performance difference between 3400 and 3833 memories is couple of seconds while running 500+ second benchmarks. They lurk for the last 4 fps difference in games, which is honestly meaningless - in real life, it doesn't matter if you have 118 fps or 122 fps. It doesn't matter if your render finishes in 29 minutes 43 seconds or 29 minutes 11 seconds.

While we learn a lot by trying things like this, never forget - there's the real life out there.

 

[jamescobalt]

I'm confused; can someone clarify? Based on my understanding of the relationship between ranks and channels to performance and stability, the ideal would be one dual rank DIMM per channel or, if necessary, two single rank DIMMs. But is the article suggesting 4 and 8 rank DIMMs for quad channels? Or are the subheadings just weird and what they mean are TOTAL ranks across all channels regardless of DIMMs? I'm guessing the latter but wanted to confirm.

 

[Crashman]

I'm confused; can someone clarify? Based on my understanding of the relationship between ranks and channels to performance and stability, the ideal would be one dual rank DIMM per channel or, if necessary, two single rank DIMMs. But is the article suggesting 4 and 8 rank DIMMs for quad channels? Or are the subheadings just weird and what they mean are TOTAL ranks across all channels regardless of DIMMs? I'm guessing the latter but wanted to confirm.

Correct: Apart from the weird article that had the Zadak four-rank modules that work only with specific two-slot Asus boards, Tom's Hardware's performance tests were only done with single and dual-rank modules. So, eight total ranks using four dual-rank DIMMs is ideal for quad-channel.

 

[jamescobalt]

Thanks a ton for the speedy clarification! It's taken me a lot of cross-eyed reading for my focus-challenged brain to get this far in understanding RAM. I'm working on a new build and I may have dived a bit too deep in trying to understand stuff on a more technical level.

 

[Crashman]

Thanks a ton for the speedy clarification! It's taken me a lot of cross-eyed reading for my focus-challenged brain to get this far in understanding RAM. I'm working on a new build and I may have dived a bit too deep in trying to understand stuff on a more technical level.

Well, here's a good article on the "technical" stuff
https://www.tomshardware.com/reviews/pc-memory-ram-frequency-timings,6328.html

 

[magyarc]

With an Intel desktop CPU can the same performance be achieved with 4 pcs of 8Gb single rank and 2pcs of 16Gb dual rank modules?
I read an excellent review about the Kingston HX436C17PB3K2/32 32Gb dual rank kit. Would two HX436C17PB4K2/16 single rank kits have the same XMP perfomance?

 

[Crashman]

With an Intel desktop CPU can the same performance be achieved with 4 pcs of 8Gb single rank and 2pcs of 16Gb dual rank modules?
I read an excellent review about the Kingston HX436C17PB3K2/32 32Gb dual rank kit. Would two HX436C17PB4K2/16 single rank kits have the same XMP perfomance?

When the frequency and timings are the same, four single rank DIMMs and two dual-rank DIMMs have the same performance.

 

[magyarc]

When the frequency and timings are the same, four single rank DIMMs and two dual-rank DIMMs have the same performance.

Thank you! I frequently run a program under Linux, which uses only around 50Mb of RAM, but it runs for about 45 minutes, and I run several hundreds of it sequentially. I have a 2x16Gb 2133Mhz cl14 dual rank HyperX Fury kit in my computer and was curious if faster RAM would matter, so I tested a couple of faster 2x8Gb (3600c17 and 4000cl19 Patriot) and also a 4x4Gb (3200cl16 HyperX) single rank kits based on your suggestion. While in synthetic benchmarks the performance gain varied from 97,4% (2x4Gb 3200c16) up to 111.2% (3600c17) in the running time of this specific program the performance gain (calculated from averaged tunning times, with very low variance) was in the range of 99.4% to 100.6%. The 4x4Gb 3200MHz CL16 "quasi dual rank" kit was only 0.2% faster in running this job, than the 2x4Gb single rank. In synthetic benchmarks the performance advantage of the 4x4Gb was around 10% against 2x4Gb.
I was quite dissapointed until I tried the G-SKill F4-3400C16D-16GVK Samsung D-die single rank kit. It was consistently 104% faster than the original 2133MHz cl14 dual rank kit. It seems to be small difference, but it means about 3 hours for a 3 day long job.
I tried two other G.Skill F4-3200C14D-16GVK and F4-3600C16D-16GVK Samsung B-die kits. All three kits are single rank and the 3400Mhz kit is the "slowest" on paper with the worst latency. Of course in synthetic benchmarks the faster kit were faster. But both faster B-die kits delivered worse performance with this job (around 100.9% when compared to the 2133Mhz result). I obtained the second best 101.2% result using the 3600C16D-16GVK kit with XMP setting but changing the "base frequency" from 133Mhz to 100Mhz. I rechecked it and the number are reproducible. I even tried to set latencyes manually on the theoretically fastest 3200MHz cl14 kit, I even dig into advanced setting like tWR,tRFC, tRRD_SL,... and tried to copy all setting of the 3400c16, but could not get better than 101%. Do you have any idea what can be the cause of this?

 

[Crashman]

Thank you! I frequently run a program under Linux, which uses only around 50Mb of RAM, but it runs for about 45 minutes, and I run several hundreds of it sequentially. I have a 2x16Gb 2133Mhz cl14 dual rank HyperX Fury kit in my computer and was curious if faster RAM would matter, so I tested a couple of faster 2x8Gb (3600c17 and 4000cl19 Patriot) and also a 4x4Gb (3200cl16 HyperX) single rank kits based on your suggestion. While in synthetic benchmarks the performance gain varied from 97,4% (2x4Gb 3200c16) up to 111.2% (3600c17) in the running time of this specific program the performance gain (calculated from averaged tunning times, with very low variance) was in the range of 99.4% to 100.6%. The 4x4Gb 3200MHz CL16 "quasi dual rank" kit was only 0.2% faster in running this job, than the 2x4Gb single rank. In synthetic benchmarks the performance advantage of the 4x4Gb was around 10% against 2x4Gb.
I was quite dissapointed until I tried the G-SKill F4-3400C16D-16GVK Samsung D-die single rank kit. It was consistently 104% faster than the original 2133MHz cl14 dual rank kit. It seems to be small difference, but it means about 3 hours for a 3 day long job.
I tried two other G.Skill F4-3200C14D-16GVK and F4-3600C16D-16GVK Samsung B-die kits. All three kits are single rank and the 3400Mhz kit is the "slowest" on paper with the worst latency. Of course in synthetic benchmarks the faster kit were faster. But both faster B-die kits delivered worse performance with this job (around 100.9% when compared to the 2133Mhz result). I obtained the second best 101.2% result using the 3600C16D-16GVK kit with XMP setting but changing the "base frequency" from 133Mhz to 100Mhz. I rechecked it and the number are reproducible. I even tried to set latencyes manually on the theoretically fastest 3200MHz cl14 kit, I even dig into advanced setting like tWR,tRFC, tRRD_SL,... and tried to copy all setting of the 3400c16, but could not get better than 101%. Do you have any idea what can be the cause of this?

Too many variables for me to mentally compare
When we set up dual rank vs single rank, etc, we always tried to narrow our variables as much as possible. I've noticed that some brands of memory don't work as well with some brands of motherboard and found really strange differences such as tRFC optimization for one brand but not another, to the point that I wrote an article comparing different motherboards with different brands of DRAM (in the same configuration) to determine why Team Group DDR4 had performance issues on MSI boards.

 

[magyarc]

Thank you for your reply! Sorry, my post might be a bit confusing...
Lets ignore comments on dual ranks, and simplify the question. There is a Gigabyte Z390 board with two G.SKill Ripjaws V 2x8Gb single rank RAM kits.
F4-3200C14D-16GVK 14-14-14-34 (XMP setting results tRFC 607)
F4-3400C16D-16GVK 16-18-18-38 (XMP setting results in tRFC 560)
Why do I not get the same performance by manually setting every parameter of the 3400Mhz kit on the more capable 3200Mhz kit? What might I overlook?

 

[Crashman]

Thank you for your reply! Sorry, my post might be a bit confusing...
Lets ignore comments on dual ranks, and simplify the question. There is a Gigabyte Z390 board with two G.SKill Ripjaws V 2x8Gb single rank RAM kits.
F4-3200C14D-16GVK 14-14-14-34 (XMP setting results tRFC 607)
F4-3400C16D-16GVK 16-18-18-38 (XMP setting results in tRFC 560)
Why do I not get the same performance by manually setting every parameter of the 3400Mhz kit on the more capable 3200Mhz kit? What might I overlook?

This is something I constantly asked myself when facing similar situations. There was always something I missed, I just couldn't find it.