Table of contents
It has been a little over a month since the Ryzen 5000 series CPUs have launched, and I think it is fair to say that they have been well received by consumers since every model is sold out absolutely everywhere. Thankfully, unlike the AMD GPUs at least the processors are being restocked semi-regularly. Anyways, with every new processor release there is always something new to learn and discover and in this article we are going to focus on what you should look for in a memory kit that is going to be paired with a Zen 3 CPU. Yes, I know there are a lot of these around, but our intent here is to approach this as a guide for newcomers. This isn’t meant to point out a single kit and say “hey that one’s the best” but rather to guide you into making a more informed decision. We are going to do this by explaining the effects of three main things: memory ranks, latency, and frequency, and what each of those things means.
Of course, there are also a few other technical things like Infinity Fabric frequency limits and overclocking, along with some of the speed bumps that he can expect along the way. There are also a few shout outs that I need to make too. First of all, I want to thank Crucial for helping us out with this video, even though the advice that we are going to be giving you can apply to pretty much any memory kit out there, they did supply us with a ton of different combinations for more apples to apples testing. This was also a whole team effort as well, Mike, myself and Patrick sunk in over 200 hours of time into this, and we learned a lot of things that we are really excited to share with all of you.
Crucial Ballistix Lineup
The first thing I really want to talk about is the memory kits that Crucial sent our way for this article, because I will be referring to them a lot. The standard Ballistix lineup comes with either a plain heatspreader or RGB that is controllable with your motherboard software. Personally, I prefer the stealth looks of the standard non-RGB modules that come in three different heatspreader colours like white, black, or red. These are available in speeds up to DDR4-3600 right now, and represent the more affordable options in Crucial lineup. Then there is the Ballistix Max series, and I’m sure you have seen me using them in a bunch of builds lately. Their matte black non-illuminated look are just perfect for blending in with pretty much any build, and they are fast too with 32GB kits hitting the DDR4-4400 mark. That is important for this video since it will give us a chance to test at lower frequencies, but much tighter timings. Now since Crucial and Micron, their parent company, are able to control pretty much all aspects of DRAM production these kits can be tuned from the ground up for some pretty unique creations. They have also been able to bin memory chips, which allows them to create kits like the Ballistix Max DDR4-5100, which the fastest memory kit on the planet right now. Later in this article we are going to push the AMD platform to the max with it, but the bigger question is will it make much of a difference?
Now speaking of the platform, we are going to talk about a well-known and settled topic that for some crazy reason people seem to be rediscovering with Ryzen 5000. That topic is single-rank versus dual-rank memory modules, but let’s start with explaining what rank actually is. To put it super simply a rank is a group of memory chips that are all physically linked together, and that can all be accessed at the same time together. You see this memory stick the memory stick above? The eight memory chips are all grouped together in a single-rank, and if I flip the memory over there are no memory chips on the other side. Therefore this is a single-sided single-rank memory stick. On the other hand, a standard dual-rank memory module would have chips on both sides, with each side forming its own grouped rank. Right now the majority of DDR4 kits follow a pretty basic rule. 8GB sticks are single-sided and single-rank, 16GB and higher sticks are dual-sided and dual-rank. However, there are a few very minor exceptions. For example, Crucial has transitioned all of their high-speed Ballistix Max 32GB kits to single-rank in order to push their frequencies. Now I wouldn’t be surprised if other manufacturers started doing the same thing, but we will just have to wait and see. If you want to know the rank count of your current memory kit, just to open CPU-Z, click on the SPD tab, and ranks are listed right over there.
Now why is this all important? Well ever since the launch of the first-gen Ryzen processors memory ranks has been a topic of discussion in a way that never happened on the Intel side. The reason being first-gen and second-gen Ryzen CPUs had pretty weak memory controllers. These older Ryzen CPUs handled single-rank memory pretty well, but not high speed dual-rank modules or other setups. And that led to highly clocked single-rank DDR4 performing better than lower clocked dual-rank DDR4 on early Ryzen systems. This is where the legend that dual-rank memory was bad with all Ryzen CPUs got started. Thankfully nowadays the Ryzen 5000 series processors have very strong memory controllers so they can handle dual-rank memory kits that are clocked as high as your Infinity Fabric or even your motherboard can handle. I hope that all made sense.
Now the reason why dual-rank is faster is very simple, RAM (Random Access Memory) only does two things: It is accessed and it is refreshed, and one memory rank can only handle one of those operations per cycle. However, when a single stick has two ranks, one rank can be accessed while the other is being refreshed, and this is called Rank Interleaving. Obviously being able to run two operations per cycle instead of just one leads to performance benefits. However, that increase isn’t universal and some apps see significant gains while others see almost none at all. I will be showing you that in just a bit, but the main point is that most CPUs perform best when their memory controller has access to two ranks per memory controller channel or four ranks in total. Now this can either be done with two dual-rank modules or you can achieve the same rank interleaving by using four single-rank modules. There is a problem with this latter approach though, for example if you buy four single-rank sticks you have already maxed out the available motherboard slots. It is also usually more expensive to buy four 8GB sticks rather than two 16GB ones.
So what kind of performance differences can you expect between the various rank configurations? Well we are going to jump into our benchmark results momentarily, but first here is our test system. We are using a Ryzen 9 5950X, but the results can be carried over into other CPUs as well. As for the other items, there is an RTX 3090 to try to eliminate GPU bottlenecks as well. Now this config will be used for all of the benchmarks you see in this article. And here are the various rank configurations that we tested. All of these were done with the timings of 16-16-16-16-36 1T, and all the way from a single single-rank 8GB module to two dual-ranks 16GB modules. Crucial also hooked us up with a few pretty rare 16GB single-rank memory modules, which aren’t common right now but I’m pretty sure we will start to see them as memory speeds increase.
Starting off with Adobe Premiere’s media encoder, and we see the biggest performance improvements going from a single 8GB module to two 8GB modules, which is to be expected, and again when going from single-rank 8GB to single-rank 16GB. There is only a very small jump going from single-rank 16GB to dual-rank 16GB modules, but there is an improvement. In Blender there was zero difference between a single 8GB module installed and two single-rank 16GB modules, which is likely due to this being a super CPU intensive app, but again we saw an improvement going to dual-rank per memory channel with the 32GB dual-rank kit and four 8GB modules. Autodesk Maya rendering tightens up the race even more, with all the results being well within the margin of error. One thing I do want to mention here is that while rendering isn’t impacted by a small 8GB memory footprint, 3D orbiting of high resolution models becomes almost impossible. In Handbrake transcoding there is a clear step down in performance going with single-rank kits, and of course a massive penalty for going with a single-channel populated. Finally, WinRAR is a super memory intensive program, but the difference between single-rank and dual-rank layouts isn’t massive.
The gaming results are also pretty interesting too. It is obvious going with a single channel layout is a big mistake and using dual-rank modules or a 4x8GB layout can have benefits in some titles. On the other hand, single-rank to dual-rank will only net you a few percentage points difference most of the time. So with this first step done, it’s pretty obvious that some programs benefit a lot from going dual-rank modules or populating all four memory slots, but other apps don’t see nearly as much of a performance uplift. Nevertheless, as often as possible, we would still recommend sticking to dual-rank modules since they are consistently better in pretty much a majority of situations if even only by a few percentage points if I’m being perfectly honest.
Memory Timings & Latency
The next step in all of this is timings and latency. The new Zen 3 architecture brought a bunch of physical changes to the CPU die, and the most important of which is the new unified cache that can be accessed by up 8 cores all at once or that one core can have full access to. The result of this new structure is a serious reduction in cache and memory latency compared to Zen 2. The memory latency improvements happen even though the memory controller hasn’t changed at all compared to the previous generation. In the most simple terms, the data has less hoops to jump through and with this comes an automatic reduction in memory latency. The big question is do these Ryzen 5000 CPU’s actually still benefit from using lower latency RAM, and is there now a point of diminishing returns when it comes to tighter timings?Well let’s find out. Since we have already established that two 16GB dual-rank kits achieve the best performance we are going to stick to that, and then we are going to lock it to DDR4-3600 and then test a bunch of timings.
It used to be that latency meant a lot when it came to achieving the best performance, but for the most part with Zen 3 going from tight CL14 timings to loose CL20 actually nets very little in real-world benchmarks. A lot of this is likely due to AMD’s ultra-fast and heavily upgraded cache design. Gaming on the other hand does see small frame rate bumps with these tighter timings, but those are mostly focused on 1% lows rather than overall averages. So what does this all mean? Well there is still a performance benefit when using lower latency RAM, but not all that much relative to what you are going to be paying for with a kit with super tight timings.
Issues and Ratios
On the issues it that is also getting a lot harder to find memory with true low timings these days, and chasing lower latencies might prevent you from hitting the DDR4-3600 to DDR4-3733 memory speeds that are still the sweet spot just like they were on Zen 2. Now you might be saying “hold on I thought there was an improvement this generation and that a 2000Mhz Infinity Fabric and DDR4-4000 were supposed to be easy, why are you still using DDR4-3600 and DDR4-3733 as the sweet spot?” Well that’s just not the case, at least not yet and maybe not ever. But to understand why, I need to talk about ratios and what the 1:1 ratio mode is and why its so important to getting the best performance on a Ryzen platform. On Ryzen 5000 – just like Ryzen 3000 – the memory clock, the Infinity Fabric clock, and the memory controller clock are all automatically locked in a synchronous 1:1:1 ratio from DDR4-2133 all the way up until DDR4-3600. Past that point those three frequencies are no longer synchronized, the Infinity Fabric is set to 1800MHz and the memory controller ends running at half the memory speed. When this happens you get a memory latency penalty of about 15%, which means a BIG hit on performance in many situations.
Many of you watching this right now with Ryzen 3000 systems and overclocked memory are probably encountering this same problem without even knowing it. An easy way to verify this is to download ZenTimings, and I will make sure to link that down below. All you to do is check out the top right corner. The MCLK is the memory frequency, the FCLK is the Infinity Fabric frequency, and the UCLK is the memory controller frequency. If they aren’t all at exactly the same frequency then you are leaving performance on the table. And by the way, we always use the 1:1:1 ratio rule because it’s simple to understand and it gives extremely good performance in a wide range of workloads.
But as with everything memory related, there are tons of exceptions and extreme configurations. Thankfully, when pushing above DDR4-3600 you can manually adjust both the memory controller and the Infinity Fabric frequencies back in line. At a speed of DDR4-3733 and above head over into the AMD CBS menu, load up the XFR Enhancement, adjust the Infinity Fabric clock so it aligns. and finally make sure the UCLK or memory controller frequency lines up too. DDR4-3733 at 1:1 should be relatively easy on most CPUs but above that point it gets a lot more challenging.
And speaking of challenges, we have achieved Infinity Fabric speeds of 2000MHz and above with 1:1 memory speeds of DDR4-4000+ many times. One 5800X hit 2000MHz at DDR4-4000 flat, the second 5800X hit 2100MHz and DDR4-4200, and our 5950X hit 2033MHz at DDR4-4066. All of those were with very low and safe 24/7 voltages. That’s fantastic and something that was impossible on the previous Zen CPUs.
But look my job is to make sure we offset cool numbers with a healthy dose of reality. Our team has access to the very best motherboards, like the ROG Crosshair VIII Dark Hero and the GIGABYTE X570 AORUS Master. Both of those motherboards have received special attention for AMD and have been fine tuned for Ryzen 5000. Your average B550 or X570 motherboard will simply not hit these numbers right now, and most lower-end motherboards that we have tried so far absolutely hate memory speeds above DDR4-3733. They might boot and load Windows at DDR4-3800, but many are experiencing silent Windows Hardware Errors since the Infinity Fabric is unstable. AMD says that a future AGESA update will improve Infinity Fabric overclocking and as a result 1:1 memory overclocking as well, but at this point who knows if it will help across the board.
Now that you all hopefully have a little better understanding of the various memory related frequencies on Ryzen processors, let’s find out the performance difference between them. And to do that we ran our benchmarking suite at four different memory speeds from DDR4-3200 all the way up to DDR4-4000 and kept identical 18-18-18 timings to isolate the gains to nothing more than faster memory and Infinity Fabric. We also threw in a little wild card here with asynchronous DDR4-4000 numbers.
Let’s just dive right in. As you can see there is definitely a small but steady increase in performance in a number of workloads when you boost the memory speed. Once again though DDR4-4000 is a level that is outside of the reach of most consumers at the moment and we did have to manually synchronize the ratios as we mentioned earlier. In some cases DDR4-3800 and DDR4-4000 do give a bit high performance, but they smash into the law of diminishing returns. Meanwhile, asynchronous DDR4-4000 results really do show the benefits of running 1:1. If we zero in on the DDR4-3600 results those do tend to be a sweet middle spot when it comes to performance, and that is good news since there are a ton of pretty affordable 16GB and 32GB DDR4-3600 memory kits around.
All you need to do to take advantage of the performance of these memory kits is to pop them onto your system, and then go into the BIOS to enable XMP or A-XMP or D.O.C.P depending on the brand motherboard you have. Now the only downside is those affordable kits tend to have really loose timings, and as we showed earlier timings still do matter a bit. Thankfully there are ways to overcome that even for novices.
The first step is to download a simple program called Thaiphoon Burner that will let you know exactly what chips your memory kit is made with. Just press read and a bunch of information is going to pop up. You need to know the number of ranks, the PCB revision, and the memory type, and them when you have this information you can enter it into the DRAM calculator for Ryzen, and that will give you precise timings that you can enter into the BIOS in order to achieve lower latencies and better performance. There are safe and fast timing options, so check which ones work best for you. The DRAM calculator is not really optimized for Zen 3 yet, but since very little has changed and the whole memory side of the platform is basically the same it still seems to work quite well as long as you stick to the timings suggestions and ignore the voltage recommendations. Then take a picture of the timings and apply the numbers in your BIOS. If and when your system boots up with the new tighter timings you can test for stability in DRAM calculator by using its built-in memory stress test called MEMbench. Last but not least, you should fire up the ZenTimings utility we showed you earlier to verify that all your frequencies are running in that golden 1:1 ratio, if they are then success. Now obviously this was just a speedy overview of what you need to do, but if you want us to go into even more detail let us know in the comments.
There is one last thing that I promised you and that is to push this platform to the absolute max with the Crucial Ballistix Max DDR4-5100 kit. Can you believe that it’s so fast that we actually hit the wall on the new ROG Crosshair VIII Dark Hero at DDR4-4866? Switching to the B550 AORUS Master allowed us to push it full speed since the kit and the motherboard are validated to run together. And no this isn’t at a 1:1 ratio since right now that would just be impossible. We could never have imagined achieving these memory speeds when we were testing the first and second generation Ryzen processors.
The results really are impressive, but they also show how not being able to run at the magical 1:1 ratio can limit performance. Now sure there are huge speedups in the right applications, but ultimately this kind of kit is going to be used by overclockers who want to set world records in synthetic bandwidth of benchmarks rather than just normal users. Nevertheless, it is really cool to see that these modules can do this on this mature AM4 platform.
So that pretty much concludes this article, hopefully it comes in handy if and when you manage to get ahold of a Ryzen 5000 series CPU. We tried to share with you every bit of knowledge that we had after hours – actually hundreds and hundreds of hours – of testing without making it all sound too complicated and I hope it worked out. I also have to thank Crucial once again since without them the apples-to-apples comparisons just wouldn’t have been possible. This has shown us that every little bit of extra performance matters, and that when you combine all the elements of dual-rank frequencies and low latency together many small increases can lead to a pretty significant change. Ultimately, if you can afford it our recommendation is to just buy the cheapest dual-rank 32GB DDR4-3600 kit that you can find now. How do we know if it is dual-rank before buying? Well what we would do is advise you do is check out user reviews or just do a quick Google search of the product name with dual-rank added to your search keyword, chances are someone on Reddit or a PC tech forum has probably mentioned it before.
App Download links below:
ZenTimings – http://bit.ly/3mwZb03
Taiphoon Burner – https://bit.ly/3nw3Z7b
DRAM Calculator for Ryzen – http://bit.ly/34q5GeE
Buy items in this video from Amazon at the links below:
Crucial Ballistix Max – https://geni.us/BALLMAX
Crucial Ballistix – https://geni.us/BALLISTIX16
Crucial Ballistix RGB – https://geni.us/BALLISTIX32