In this article, we're going to look for the Zen 3's memory sweet spot and examine DDR4 memory performance with the new Ryzen 5000 series of CPUs. Before we get into this, it is important that we create a reference based on previous tests and findings.
Earlier this year, we compared Intel's Core i9-9900K and Ryzen 9 3900X with two 8 GB modules with four 4 GB modules with exactly the same storage times. We found that using a rather slow DDR4-3000 CL16 memory was aiming for a 3-5% performance increase for the average frame rate at 1080p using an RTX 2080 Ti, but looking at 1, up to 10%% low Results. This has been observed with both AMD and Intel processors. Certainly not massive gains, but this is relatively slow storage and we expected margins to grow a bit with faster storage.
Before we move on to our updated tests, here's a quick explanation of why 4 sticks are faster than 2 …
It boils down to how the memory is configured, or rather the "rank" of the memory. For those of you who don't know, the term "rank" means the number of 64-bit memory banks in a module. Most consumer memories are single-ranked, although higher capacity modules are typically double-ranked, while server memories are often four-ranked.
It can be difficult to determine if your memory has a double or single rank because the software does not always read the modules correctly and not all memory manufacturers record the rank in the module ID. Typically, single-rank modules all have memory chips on one side of the circuit board, while dual-rank memories place chips on both sides of the circuit board. However, this is not always the case. A module with chips on either side of the board is actually only double-sided and can still be a single-tier module, so it's a bit confusing.
Rank configurations of the DIMM module
It can get even more confusing as you insert more memory sticks or modules. A system with more than two single-tier modules actually behaves as if two-tier modules were installed. In fact, there is little difference between a two-tier module and two single-tier modules when connected to the same memory controller, even though the memory chips are on different circuit boards.
So if you use two one-tier modules for dual-channel operation, the memory is configured as one one-tier module. However, if you are using four single-rank modules for dual-channel operation, the memory is now configured as dual-rank.
This can give the configuration with four DIMM modules an advantage, since multiple open DRAM pages are possible in each tier. Although the ranks cannot be accessed concurrently, they can be accessed independently. This means that the controller can send write data to a rank while waiting for read data previously selected from another rank, and as you will see in a moment, it is very important to increase the memory bandwidth.
How much this affects performance depends on the application and the ability of the storage controller to take advantage of open pages. However, all of this means that four modules can improve performance over two modules in a two-channel system.
Our test system was equipped with the GeForce RTX 3090. Here is our comparison of the Ryzen 9 5900X, Ryzen 9 3900X and Core i9-10900K when using two DDR4-3200 modules with four DDR4-3200 modules.
We only tested two games as this is all the data we need to clear up any misunderstandings regarding 4 sticks with Zen 3 compared to other CPUs.
When testing Shadow of the Tomb Raider, a very CPU demanding game, we found that the 3900X's average frame rate increased 14% when switching from 2 to 4 modules. This is a very significant improvement, even though we're using an extreme low-resolution GPU.
In the transition to the Core i9-10900K, we again see a large increase in performance when 4 sticks are used, this time a performance increase of 15%. With the Ryzen 9 5900X, we're now seeing a similar increase in performance, this time by 12%, and it's possible that we're running into a GPU limitation as Amps scales poorly at 1080p, but the point is that all three CPUs see a similar double-digit Performance increase with 4 sticks, so this is not a special feature of Zen 3.
Hitman 2 is another CPU / memory sensitive game and we'd like to connect a GPU to the 5900X. Let's focus our comparisons on the 1% low data. With the 3900X we expect an 8% increase in performance with 4 sticks. However, using the 10900K shows a far more significant increase in performance of 29%, although that's only a 13% increase in average frame rate.
Hitman 2 can be a little strange, and we suspect the 3900X is a performance bottleneck, more related to core-to-core latency than DRAM performance. However, with the 5900X we are also seeing a massive increase in the 1% low performance with 4 sticks, this time a 21% increase in performance. However, we again suspect that we will encounter a GPU limitation in the average frame rate. The point is, the 10900K and 5900X with four memory modules see a similar increase in performance, and not just for the Zen 3.
Now let's look at how the Zen 3 architecture behaves with different memory modules, frequencies and timings. We'll be using the RTX 3090 again, but will later have data with a more general GPU too.
At this point, none of our Ryzen 5000 processors are working with a 2000 MHz FCLK, which limits us to DDR4-3800 memory. Apparently a new BIOS will make a 2000 MHz FCLK more likely, although it didn't help in our case. The good news is that everything worked perfectly with a 1900 MHz FCLK, and it didn't with Zen 2.
For much of these tests, we used G.Skill's TridentZ 3600 CL14 memory, which we manually set to increase the CL to 16 but aggressively tighten the secondary and tertiary timings, increasing the performance of all Ryzen processors massively improved. This configuration is tested on DDR4-4000, 3800, 3600, and an underclocked 3000 configuration.
We also added some XMP-loaded memory configurations, one with Corsair Dominator Platinum DDR4-3600 CL18-19-19 memory, one with ADATA XPG Spectrix D50 DDR4-3600 CL18-20-20, but with two different configurations: one with two Dual-rank modules and the other two single-rank modules.
We have also included our G.Skill TridentZ F4-3200 C14 4x8GB test configuration that will be used in our reviews and we will also be adding a 2x8GB configuration for single-rank testing. We have done our best to make this data as simple as possible, but we understand that this will be a bit confusing for some of you. Finally, we'll stick with the Ryzen 9 5900X, but note that these results apply to all Zen 3 processors, including the Ryzen 5 5600X.
Death Stranding is a game where Zen 3 processors went like a bat out of hell, and yet this isn't a particularly memory-sensitive title. For example, if you compare our test configuration with four DDR4-3200 CL14 memory modules, rating the memory to 3800 with precisely matched timings only increases performance by 3 FPS, which is a mere 1.3% increase.
Additionally, we see a 7-10% difference between the fastest and slowest memory configuration tested, which is not that important. It also looks like the two-tier DDR4-3600 CL18 memory is comparable to our two-tier DDR4-3600 CL14 memory, at least in this title.
Moving on to F1 2020, we see little increase in our test system configuration using the manually tuned DDR4-3800 memory. This time we see a performance delta of ~ 5% between the fastest and the slowest memory configuration tested. Hence, F1 2020 is another game that is not particularly sensitive to memory performance.
We know Far Cry New Dawn is a memory sensitive game, especially latency sensitive, and here we see a more noticeable 12% performance gain with optimized DDR4-3800 memory over our 3200 test configuration.
If we use the same 3800 spec and just increase the frequency to DDR4-4000, which currently decouples from a 1: 1 ratio with the Infinity Fabric without 2000 MHz FCLK support, we actually break the performance a bit and get noticed the tuned 3600 spec.
Horizon Zero Dawn like F1 2020 and Death Stranding are not particularly memory sensitive, and again we only see a 5% difference between the fastest and the slowest memory. There is also little difference between one-tier and two-tier memory configurations.
Another game that isn't greatly affected by memory performance is Rainbow Six Siege. Here we see a change of less than 2% between the top and bottom configurations tested.
Storage performance makes a reasonable difference in Watch Dogs Legion. We expect a 6% increase in performance by adding just two more DDR4-3200 modules. Overclocking and tuning increased the performance by a further 3%. Certainly not something that occurs at higher resolutions where there are GPU limitations, but there is a measurable difference.
As mentioned earlier, Hitman 2 is very memory and CPU sensitive. Optimizing your memory can make a huge difference in this game, although it will only do so if the CPU's performance is limited.
We see that the performance with standard DDR4-3600 CL18 kits like the Corsair Dominator Platinum RGB or the ADATA XPG Spectrix D50 in a one-tier configuration is pretty terrible compared to the two-tier or manually tuned configurations. Amazingly, we see a 1% reduction in performance by ~ 23%.
If we look at the 2 and 4 stick DDR4-3200 configurations, we see that our 4 stick test configuration would be up to 17% slower if we removed two of the modules.
We also see Spectrix D50 16GB DDR4-3600 modules which offer a nice increase in performance as they were 6% faster than our 3200 CL14 test configuration. The performance of dual rank operation is strong, and unfortunately we couldn't test the manually tuned DDR4-3800 configuration in dual rank mode because we don't have enough of these modules.
In Hitman, with memory at or above the 3200 spec, we see a performance difference of up to 21% and an increase of up to 12% over our test configuration.
However, if we increase the resolution to 1440p, the 21% margin is reduced to 14%, which is still significant, but is greatly reduced by increasing the GPU load. While the manually tuned DDR4-3800 memory was 12% faster at 1080p than our DDR4-3200 test configuration, it is only 4% faster at 1440p. Still, dual-rank memory makes a big difference in this title.
Like Hitman 2, Shadow of the Tomb Raider is a CPU demanding game that is also responsive to memory performance. Looking at our test configuration, we see that using 4 TridentZ DDR4-3200 CL14 modules improved performance by a whopping 12% compared to just two modules. It is also faster than the single tier DDR4-3600 CL18 configurations.
We see a similar performance to the ADATA XPG 32 GB DDR4-3600 kit, and this increase is of course not due to the additional capacity, but to the two-tier configuration. In addition, we don't gain much from the manually tuned DDR4-3600 and 3800 memory.
We also ran some 1440p tests on the Shadow of the Tomb Raider, and we believe you can expect this to happen in most games, even with something as extreme as an RTX 3090.
While there was a 19% difference between the fastest and slowest configurations tested at 1080p, that margin is reduced to just 4% at 1440p, and if the memory is running at or above the AMD base spec we speak of a 0 difference , 7%.
8 game average
If we average the 1080p data over the 8 games tested, it gives a clear picture of the kind of performance differences you can expect from an extreme low-resolution GPU.
In game scenarios that are more CPU-limited, a manually adjusted DDR4-3800 configuration brings about 7% more performance compared to a standard memory kit such as the Corsair Dominator Platinum RGB. Compared to our test setup, we only expect an average increase of 3%.
AIDA64 memory bandwidth
For those of you who are wondering, here's a look at the memory bandwidth performance of these various configurations. Although the DDR4-4000 performance was average as we couldn't work with an FCLK of 2000 MHz, the bandwidth at 55 GB / s is still very impressive.
The manually tuned DDR4-3800 memory managed 53 GB / s, which wasn't much faster than the ADATA Spectrix 32 GB kit, which hit nearly 52 GB / s. Our test configuration was good at almost 47 GB / s, which is about the best you can hope for in DDR4-3200 memory.
Here is the reason why the DDR4-4000 configuration with 55 GB / s did not dominate the gaming benchmarks: The latency of 60 ns is rather inconspicuous and that is only marginally better than our DDR4-3200 test setup. The optimized DDR4-3800 memory reduced DRAM latency by 9%, which is why it did so well in memory-sensitive titles like Hitman 2, Far Cry New Dawn and Shadow of the Tomb Raider.
DDR4 frequency scaling
Speaking of Hitman 2, here's a look at memory frequency scaling. So we're using the same memory and timings with just changing the memory frequency and FCLK which were believed to be 1: 1 for optimal performance. except for the DDR4-4000 configuration.
In Hitman 2, we see pretty consistent scaling as the memory bandwidth and / or latency is improved up to DDR4-3800. If we could get a 2000 MHz FCLK working, we would expect another 3% increase in performance for the DDR4-4000 configuration.
Shadow of the Tomb Raider fits in better with other demanding games, and this is where we find DDR4-3600 to be the sweet spot, as was the case with Zen 2. For those of you who want as much performance as possible without exceeding memory prices, the DDR4-3600 CL16 seems like the way to go.
RTX 2070 super benchmarks
But wait, there's more. Here is the same memory configuration we just saw in the scaling benchmarks, but this time with the RTX 2070 Super …
Even in Hitman at 1080p, with an RTX 2070 Super you are almost completely tied to the GPU, and if you happen to be playing at 1440p, you are completely tied to the GPU no matter how much bang for your buck on memory or how many modules you use the performance will be the same.
We see exactly the same thing in Shadow of the Tomb Raider, even at 1080p we see balanced performance across the board due to the GPU bottleneck, and the 2070 Super is no problem, we're talking about Radeon RX 5700 XT-like performance here, so very solid GPU performance in the middle range. Of course, increasing the resolution will only further neutralize the results. Now it doesn't matter whether you're using DDR4-2800 or 3800. The performance is the same in a very CPU demanding title.
What we learned
Bottom Line: If you are a gamer looking to maximize performance when you run into CPU limited situations and want the most bang for your buck, we recommend getting DDR4-3600 CL16 memory. For most, we suspect 16GB is fine. However, if you can afford more, 32GB is fine and if you buy two 16GB kits you also have the benefit of dual ranking operations.
Right now, something like Crucial's Ballistix 16GB DDR4-3600 CL16 kit looks great and costs just $ 75. They offer a great deal of tunability if you want. G.Skill also offers an affordable DDR4-3600 CL16 kit for around $ 80.
Assuming we were able to get a 2000MHz FCLK working with future BIOS revisions, we don't think it makes sense to spend more than $ 100 on these kits. You will likely never notice the difference. This stuff is more reserved for overclockers looking to get bigger 3DMark scores or some other ranking that gets them excited these days.
As for the debate about 2 sticks versus 4 sticks of memory. Since our own test almost a year ago, there is nothing new to report. The performance increase for Zen 3 is no different from that of Zen 2 or competing Intel processors. The margins also depend on the quality settings used and of course on the hardware. As you lower the quality settings in games you will further exaggerate the margins, which will take you further from reality for most gamers.
We know that most of you will just be installing your memory and playing games, and frankly, if you don't enjoy playing with these things, it's just a waste to spend hours tweaking and tinkering with memory times which will likely mean very little real world gains by time.
This may be more important to us reviewers trying to run scientific tests, and sometimes it's a bit difficult to isolate a particular component to see what the differences in performance might be, but it's really important to remind readers In most cases it is unlikely that these winnings will be achieved under realistic gaming conditions. So keep that in mind.
- Decisive Ballistix DDR4-3600 CL16 on Amazon
- G.Skill DDR4-3600 CL16 at Amazon
- AMD Ryzen 5 5600X on Amazon
- AMD Ryzen 7 5800X on Amazon
- AMD Ryzen 9 5900X on Amazon