Ryzen Above: Best Memory Settings for AMD's 3000 CPUs, Tested

AMD set the memory spec for Ryzen 3000 at DDR4-3200. But any reviewer or enthusiast worth their salt is going to settle for the base spec, because we know there’s always more to the story. While the base line might be what the company recommends, there's certainly more performance to be attained with the right kit, especially when many motherboards boast supported speeds well above 4,000 MHz.

And of course, performance is about much more than just frequency. Several of our readers for example have stated that DDR4-3200 CAS 14 and DDR4-3600 CAS16 are approximate equals, but that just sounds like another generalized rule. How much truth is behind these claims, and how much farther can an average performance PC builder push their performance by tweaking timings and settings?

Rather than drill down to find the best frequency and timings that one kit can muster (we have reviews for that), we’re going to push a dozen or so of the most common “performance optimized” configurations in order to find out when (and why) you might want to choose a faster kit for your AMD build. G.Skill’s part number F4-3600C16Q-32GTZN Triden Z Neo DDR4-3600 has the proven stability we need to test across a broad range of settings. Other components, including MSI’s MEG X570 Ace motherboard and Gigabyte’s RTX 270 Gaming OC 8G are carried over from that memory review, along with our chosen 4.20 GHz fixed frequency for AMD’s Ryzen 7 3700X.

Frequency: Is DDR4-3200 Really Optimal?

When AMD says that Ryzen 3000 is DDR4-3200 optimized, one might expect that anything less will result in noticeable performance losses and that anything more will produce minuscule gains. Our first response is “prove it.” To focus solely upon data rate means we must minimize the impact of different latency settings in this test, but we’ll get back to that in our next section.

Our attempts to keep latency constant in our first test define the frequencies we’ll use in other tests. Latency is measured in cycles, cycle time is the inverse of frequency, so that both DDR4-3600 CAS 18 and DDR4-2800 CAS 14 have the same ten nanoseconds of real-time latency as DDR4-3200 CAS 16. Though we prefer to set our memory in multiples of 266.667 MHz, none within this range would serve the same matched-latency function.

The greatest performance difference in our combined averages is the 2.5% spread between DDR4-3600 and DDR4-2800 in application startup, so why bother? Digging deeper, we find a few applications with larger differences that were minimized through averaging. We even see that Sandra Memory Latency’s dependence on bandwidth means that all-else-equal…isn’t.

The performance difference between our top and bottom benchmark set jumps to 5% in both F1 2017 and 7-Zip, and 4% in Ashes. And that’s total system performance, so if you use these apps, it might pay to spend a little extra on fast memory.  Five percent of a $1000 build is $50, and that’s typically more than the difference in price between DDR4-2800 and DDR4-3600.

DDR4-3200 performance is dead center between DDR4-2800 and DDR4-3600, making the Ryzen 7 3700X appear “optimized” for a broad range of date rates rather than DDR4-3200 alone.

Latency: What About DDR4-3200 C14?

We started our tests by trying to equalize latency across every testing frequency, but that’s not going to cut it for ultimate performance. Shaving two cycles off DDR4-3600 C18 and DDR4-3200 C16 gives us DDR4-3600 C16 and DDR4-3200 C14, where the later should get a bigger boost since two cycles are a larger portion of sixteen than of eighteen. Is that greater boost enough to validate rumors that DDR4-3600 C16 and DDR4-3200 C14 are roughly equivalent?

CAS 14 did very little for our DDR4-3200 configuration, to the point that it appeared to have lost a step in our overall performance score while falling within the margin of error. But the margin of error is most noticeable in applications where we expect a tie, and we can still separate a few of the tests that are more memory-dependent.

DDR4-3200 C14 did slightly better than DDR4-3600 C18, but worse than DDR4-3600 C16, in Ashes. F1 2017 and our custom 7-Zip file compression timed test show that the extra bandwidth is at least as important as the tighter timings.

Command Rate: What Are The Benefits Of 1T?

Command rate is the number of cycles a command must be presented before a memory operation commences, so that 2T incurs an additional cycle of latency to every memory access. It makes sense that most people would shoot for 1T, except that higher data rates often make this unworkable, particularly when more ranks of memory are used (ie, two “double-sided” or four “single-sided” DIMMs). The reason that we’ve used only half of our four-DIMM kit thus far is that it won’t run all four DIMMs at DDR4-3600 and 1T, hence you’ll want to see the penalty of 2T before we add four-DIMM configurations to the mix.

Based on entire suites, 2T timings appear to impart a maximum penalty of less than 2% and an average penalty of less than 1%. What do our memory-bottlenecked benchmarks show?

We see a 3% penalty in 7-Zip for 2T timings at DDR4-3200, which makes a little more sense than the previous chart because a single cycle of extra time is a little longer at DDR4-3200 than at DDR4-3600. The difference could be little more than 2%, since numbers near a half-percent are either rounded upward or downward.

Capacity: Does More Memory Perform Better?

As we prepare to compare the performance of two DIMMs to four, we thought we should first show how much performance difference there is between 16GB and 32GB. After all, two DIMMs of today’s main test kit have only 16GB, but four have 32GB. Doing so presents a different problem however, as four single-rank DIMMs will perform differently from two, even when the total capacity is the same.

To get 16GB from the same four ranks as our current 32GB set, we needed memory with identical capability but half the capacity. And since those haven’t been produced for several years, we hopped in the wayback machine and grabbed a set from this 2015 test platform.

Amazingly, 16GB outperformed 32GB when frequency, timings, and the organization of ICs on the circuit board are constant. While we could guess that the memory controller is able to address the lower-capacity ICs more quickly, the more important point was to prove that our benchmarks don’t benefit from having more than 16GB of total capacity.

Benchmarks that typically benefit the most from increased data rates or decreased latencies show less favoritism for the smaller kit. None of our benchmarks are designed for 32GB, and our recommendations for that capacity have primarily focused on multitasking rather than benchmark data.

MORE: Best Memory

MORE: DDR DRAM FAQs And Troubleshooting Guide

MORE: All Memory Content

39 comments
    Your comment
  • drivinfast247
    3600mhz is the sweet spot for 3000 series Ryzen.
  • NightHawkRMX
    So another test I want to see with ryzen 3000, faster speed ram with loose timings, or slower speed ram with tight timings.
  • Soaptrail
    So we should get 4x8GB RAM for Ryzen? I see on the reviews for Ryzen 3, the official support for 4 dimms at dual rank is only 2667. Can anyone confirm is it easy or hard to buy 4x8GB of dual rank and use it at 3600? i will do it but if it is a lottery of working or not I want to know.

    I have been building PC's for years but I never really got into overclocking other than XMP profiles. So i am willing to do this if it is easy enough to succeed. Thanks to anyone who helps explain this to me. I enjoyed this article but it was not as noob friendly as I needed.
  • NightHawkRMX
    Ryzen is dual channel only.

    Just get 2x16gb and it will be more compatable and have better performance.
  • salgado18
    Quote:
    Ryzen is dual channel only. Just get 2x16gb and it will be more compatable and have better performance.

    Care to show the tests to support your claim that 2x16 is faster than 4x8? Because the article you're replying to did just that, and proved the opposite.
  • NightHawkRMX
    Traditionally when running 4 sticks in 2 channels, you are usually forcing the command rate to 2T rather than 1T, making the ram latency worse.

    Heres what the article says:
    "Command rate is the number of cycles a command must be presented before a memory operation commences, so that 2T incurs an additional cycle of latency to every memory access. It makes sense that most people would shoot for 1T, except that higher data rates often make this unworkable, particularly when more ranks of memory are used (ie, two “double-sided” or four “single-sided” DIMMs). The reason that we’ve used only half of our four-DIMM kit thus far is that it won’t run all four DIMMs at DDR4-3600 and 1T, hence you’ll want to see the penalty of 2T before we add four-DIMM configurations to the mix."

    I see no way for 4 sticks to be faster as they have no more bandwidth than 2 sticks. If you look at the far right 2, you see the performance is roughly the same, which is very surprising. I have seen 4x8gb would have less compatability with ryzen in the past, but not sure of now.

  • CheckDM
    Hmm. This throws a curveball in my upcoming build, because I already want 32GB.


    I was simply going to go 2x16 with "Corsair Vengeance LPX 32GB (2x16GB) DDR4 DRAM 3200MHz C16" (CMK32GX4M2B3200C16) for $155.
    But now I'm thinking I should go with a 4x8G with "CORSAIR Vengeance LPX 16GB (2 x 8GB) DDR4 3200" (CMK16GX4M2B3200C16) for a total of $140.

    The 4x8 config is a little cheaper, and according to this article using 4 sockets is possibly better.


    (However, the QVL sheet for my chosen motherboard, Gigabyte X570 Auros Elite, supports the 2x16 but the ram selected for 4x8 is only supported in 2x8. So this might be the overriding factor.)
  • Z1NONLY
    4 single rank sticks allow the controller to interleave the ram, making it faster than two single rank sticks. I have experienced this in my own 3800x build. Both time spy physics and shadow of the tombraider did better with 4 sticks of ram.

    As the article points out, with 4 sticks an operation can start on one rank while another has not yet finished on the other...and this is on each of the channels.

    My TS physics score was ~10300 with two sticks and ~10700 with four.
  • NightHawkRMX
    Quote:
    4 single rank sticks allow the controller to interleave the ram, making it faster than two single rank sticks. I have experienced this in my own 3800x build. Both time spy physics and shadow of the tombraider did better with 4 sticks of ram. As the article points out, with 4 sticks an operation can start on one rank while another has not yet finished on the other...and this is on each of the channels. My TS physics score was ~10300 with two sticks and ~10700 with four.

    That makes sense.
  • Crashman
    Quote:
    So we should get 4x8GB RAM for Ryzen? I see on the reviews for Ryzen 3, the official support for 4 dimms at dual rank is only 2667. Can anyone confirm is it easy or hard to buy 4x8GB of dual rank and use it at 3600? i will do it but if it is a lottery of working or not I want to know. I have been building PC's for years but I never really got into overclocking other than XMP profiles. So i am willing to do this if it is easy enough to succeed. Thanks to anyone who helps explain this to me. I enjoyed this article but it was not as noob friendly as I needed.

    You should get four ranks of 8GB. The article says it doesn't matter if it's four single-rank or two dual-rank DIMMs.

    Quote:
    Hmm. This throws a curveball in my upcoming build, because I already want 32GB. I was simply going to go 2x16 with "Corsair Vengeance LPX 32GB (2x16GB) DDR4 DRAM 3200MHz C16" (CMK32GX4M2B3200C16) for $155. But now I'm thinking I should go with a 4x8G with "CORSAIR Vengeance LPX 16GB (2 x 8GB) DDR4 3200" (CMK16GX4M2B3200C16) for a total of $140. The 4x8 config is a little cheaper, and according to this article using 4 sockets is possibly better. (However, the QVL sheet for my chosen motherboard, Gigabyte X570 Auros Elite, supports the 2x16 but the ram selected for 4x8 is only supported in 2x8. So this might be the overriding factor.)
    Maybe I should repeat the above statement from the article body in its conclusion ;)
  • nitrium
    I just built a Ryzen 3700X system with 2x16GB Corsair Vengeance Pro 3200 MHz (Micron DRAM and they are dual rank) and with everything at stock with X.M.P. settings (16 18 18 36 1T) I get 36.6 GB/sec on Sandra which looks about what one would expect from this article. I guess there should be a small advantage in being able to run 4 ranks at 1T as [USER=2458595]remixislandmusic[/USER] suggested(?), but it's not significant by the looks.
    I believe I can OC this to 3600 MHz with roughly the same timings, but I don't think I'll bother based on the mostly small increases in overall performance.
  • alextheblue
    Quote:
    You should get four ranks of 8GB. The article says it doesn't matter if it's four single-rank or two dual-rank DIMMs.

    That's the key takeaway. What's annoying is that a lot of kits don't say anything about command rate or number of ranks per stick. That's not great if you tell someone to look for X and it doesn't say anything about it on various vender sites.
    Quote:
    Traditionally when running 4 sticks in 2 channels, you are usually forcing the command rate to 2T rather than 1T, making the ram latency worse.
    The number of ranks is what matters most, not the number of sticks.
    Quote:
    I see no way for 4 sticks to be faster as they have no more bandwidth than 2 sticks.
    Depends on the ranks, well that and the application. The idea is that even if you don't "need" 32GB, if you want the increased performance of 4 ranks (2 per channel) then you're kind of stuck because 8GB sticks are likely all single rank now. So you either need 4 of those, or two 16GB dual rank sticks.



    Look at the 3200 kit that's running at the same clocks and latency. Big difference. Not as big of a difference when you have to jump to 2T, but yeah as discussed above there can be advantages to quad rank.
  • mattkiss
    Do you think using a motherboard with a T-topology memory trace layout would give different results compared to the MSI MEG X570 Ace board you used (it has a daisy-chain topology)?
  • nitrium
    Quote:
    The number of ranks is what matters most, not the number of sticks.

    The point I think is that if you get 2x16GB you can run the optimal four ranks at 1T, as opposed to grabbing 4x8GB which you will need to run at 2T - the article shows there is a small but real difference between 1T and 2T. Therefore, there should be a slight performance difference between these two options (not to mention that 4x 8GB sticks generally cost more than 2x 16GB sticks), right? I'm not sure, but it makes sense.
  • Crashman
    Quote:
    The point I think is that if you get 2x16GB you can run the optimal four ranks at 1T, as opposed to grabbing 4x8GB which you will need to run at 2T (the article shows there is a small but real difference between 1T and 2T). Therefore there should be a slight performance difference between the two options (not to mention that 4x 8GB sticks generally costs more than 2x 16GB sticks), right? I'm not sure.
    No. Two 16GB dual-rank sticks will usually require the same timings as four 8GB sticks. In the test system, 2T was needed to run four ranks at DDR4-3600. It wouldn't have mattered if the four ranks were on two dual-rank or four single-rank DIMMs. Slowing down to DDR4-3200 allowed the system to be stable with four ranks at 1T timings.

    I should note that 2T is standard even at DDR4-3200, and that manually setting 1T is akin to overclocking.
  • nitrium
    Quote:
    I should note that 2T is standard even at DDR4-3200, and that manually setting 1T is akin to overclocking.

    Are you sure? I have pretty average 2x16GB Corsair Vengeance 3200 MHz RAM (not the higher price Dominator stuff), and it's running at 1T directly from its only XMP Profile with no intervention on my part. I'm going to look into this further. Is 1T exclusive to two sticks (regardless of ranks), or not - i.e. can four sticks ever be run at 1T?
  • Crashman
    Quote:
    Are you sure? I have pretty average 2x16GB Corsair Vengeance 3200 MHz RAM (not the higher price Dominator stuff), and it's running at 1T directly from its only XMP Profile with no intervention on my part. I'm going to look into this further. Is 1T exclusive to two sticks (regardless of ranks), or not - i.e. can four sticks ever be run at 1T?

    I don't believe I've ever seen a four-rank DDR4-3200 kit of any flavor that ran at 1T by XMP, but if yours does...heck it might be a credit to the motherboard's firmware. At any rate, four ranks is four ranks no matter how many slots, which is why four ranks WAS run at 1T and DDR4-3200, by manual configuration, in this article.
  • cryoburner
    Quote:
    Would you have believed that you could increase your frame rates by up to 10% in Ashes simply by using four ranks of memory? Or that you could reduce file compression time by up to 13%? Isn’t that performance gain similar to upgrading to the next level of graphics card or CPU? Given that 8GB per-rank capacity fills the enthusiast market, isn’t this a good enough reason for us to recommend 32GB to gamers, even to those who will never need more than half of its total capacity?

    There is something else important to consider though, and that is that these examples are outliers, not the norm. Sure, moving up to 32GB might provide notable performance gains in some rare examples, but in the vast majority of games, the performance differences will tend to range from unnoticeable to nonexistant. That's different from the performance gains from moving up to a higher-end graphics card, where performance can be improved in practically all modern games to a noticeable degree.

    And even these outlier examples are of questionable relevance. Ashes of the Singularity might make for a good multi-threaded CPU performance benchmark, but it was never really among the most popular games, and those kinds of performance gains likely won't translate to many titles that people are actually playing in significant numbers. And how much does a 10% performance difference even matter in an RTS game like this when the game is running at 100+ fps with all memory configurations at those settings? That goes even more so for F1 2017, where we are looking at around a 3% difference for framerates in excess of 200fps, something that would be completely imperceptible.

    And these outliers were only seen with the settings turned down somewhat at 1080p, which is probably not a very common scenario for most of those running an RTX 2070 like the test system appears to be using (I had to search for the specifications in another review, which is not exactly ideal). With a more mid-range graphics card, or this card running at max graphics settings or a higher resolution, these already barely-significant performance differences become much less significant. Most games tend to have their performance limited by the GPU more than anything, so the performance gains from a more expensive memory configuration will tend to be quite minimal. If someone has an extra $75 or $100 to spend on a gaming system, and is deciding between 32GB of RAM or a faster graphics card, the faster graphics card will likely be the better option for the vast majority of games released within the next few years. The additional RAM may provide additional benefit eventually, but for now the benefits of 32GB for gaming and most other tasks will be limited.

    Nice article though! It helped give a better idea of just how much these differences in memory speed, timing and rank can potentially affect performance on 3rd-gen Ryzen. It was interesting to see just how much dual-rank memory can potentially affect performance in certain applications. Getting 2x16GB or 4x8GB of mid-range RAM seems to be more beneficial to performance than getting just 2x8GB of faster, low-latency RAM that could end up costing more.

    Moving from a typical single-rank 16GB kit of 3200 C16 memory (around US $75) to a much more expensive low-latency single-rank 16GB kit of 3600 C16 (close to $150) only improved performance by up to around 4% in those outlier examples (4% in Ashes, 4% in F1 and 2% in 7-Zip). If one were to instead use that money to go with a dual-rank 32GB kit of 3200 C16 (also around $150), that could net them over a 10% increase in performance over the single-rank 3200 kit in those outliers (10% in Ashes, 3% in F1 and 13% in 7-Zip). Plus, you get double the RAM, so if future games or other applications start showing notable performance benefits from having more than 16GB a few years down the line, you will already have that, and won't need to spend more on upgrading. This seems to make a 32GB kit of normal-latency RAM a better value than a 16GB kit of higher-clocked low-latency RAM. And sure, you could get a kit that features both 32GB and higher speeds with lower latency, but it's probably not worth paying too much more for unless you have an unlimited budget for a build. Again, these best-case examples are outliers, and the typical performance gains will likely tend to be much lower.
  • sizzling
    2x8gb vs 4x4gb for gaming with Intel. There is a newer video for a Ryzen comparison but the spoiler is the difference is similar to Intel

    https://forums.tomshardware.com/threads/2x8gb-vs-4x4gb-gaming-comparison-intel.3479216/
  • Crashman
    Quote:
    2x8gb vs 4x4gb for gaming with Intel. There is a newer video for a Ryzen comparison but the spoiler is the difference is similar to Intel https://forums.tomshardware.com/threads/2x8gb-vs-4x4gb-gaming-comparison-intel.3479216/

    We've been doing Intel for years.
    https://www.tomshardware.com/reviews/super-talent-project-x-f3000ux16g-ddr4-memory,5038.html
    Did you want another article like "Does AMD really get a bigger boost from fast memory than Intel?"

    Spoiler, it doesn't. All the myths and rumors about AMD needing fast memory more than Intel needs it are excuses and rationalizations dreamed up by AMD fanboys. Both platforms benefit in a similar manner, and I'm just happy that AMD now supports fast memory without jumping through a bunch of hoops to run it.
  • sizzling
    Quote:
    We've been doing Intel for years. https://www.tomshardware.com/reviews/super-talent-project-x-f3000ux16g-ddr4-memory,5038.html Did you want another article like "Does AMD really get a bigger boost from fast memory than Intel?" Spoiler, it doesn't. All the myths and rumors about AMD needing fast memory more than Intel needs it are excuses and rationalizations dreamed up by AMD fanboys. Both platforms benefit in a similar manner, and I'm just happy that AMD now supports fast memory without jumping through a bunch of hoops to run it.

    The main point isn’t about RAM speed but dual channel 2 bank vs 4 bank. The videos are an interesting comparison to compliment any other information.
  • Crashman
    Quote:
    The main point isn’t about RAM speed but dual channel 2 bank vs 4 bank. The videos are an interesting comparison to compliment any other information.
    Did you click the link I showed you? It's about 2 rank vs 4 rank on Intel. It's from 2017. It's a follow-up to another review that showed the same pattern.

    Edit: Come to think of it, could you imagine the wailing and gnashing of teeth had we titled this article "NO, AMD does NOT meaningfully benefit from faster RAM: Get 4 Ranks Instead!" There would be blood...
  • nitrium
    Quote:
    I don't believe I've ever seen a four-rank DDR4-3200 kit of any flavor that ran at 1T by XMP, but if yours does...heck it might be a credit to the motherboard's firmware. At any rate, four ranks is four ranks no matter how many slots, which is why four ranks WAS run at 1T and DDR4-3200, by manual configuration, in this article.

    I just double checked, and the timings are all definitely still set to "Auto" in the BIOS. The motherboard I'm using is actually one that has been reviewed here, a Gigabyte Aorus Eilte - one of the cheapest X570 mobo's around.
    https://www.tomshardware.com/reviews/gigabyte-x570-aorus-elite-atx-motherboard,6252.html
  • mitch074
    Quote:
    Did you click the link I showed you? It's about 2 rank vs 4 rank on Intel. It's from 2017. It's a follow-up to another review that showed the same pattern. Edit: Come to think of it, could you imagine the wailing and gnashing of teeth had we titled this article "NO, AMD does NOT meaningfully benefit from faster RAM: Get 4 Ranks Instead!" There would be blood...

    Sure there would be! But that would come from ignorant fanbois (I'd like to consider myself being a somewhat informed one) inasmuch as Ryzen 1/2xxx was really dependent on RAM quality to work well, while Ryzen 3xxx just works well with any RAM, same as Intel.
    On a purely editorial matter, negative titles are either flame bait or opinion pieces (and often both).
    As for THIS article, it's been quite a while since I've seen someone take such a close look at RAM timings - these past few years have made us so used to having the BIOS, XMS profiles and the memory controllers deal with everything in our place without big mishaps that I'd pretty much forgotten what it was like to test settings, run benchmarks and tally results out of a RAM kit - it all came down to CAS, frequency and voltage. Seeing us get back to the correlation between RAM banks and command cycles takes me back to the original Duron days (where matching a DDR kit's frequency to the FSB and switching to 1T could mean a 15 to 25% performance boost overall).