Couple words
Timings, sub-timings (primary and secondary), frequency, XMP / overclocks / 2 sticks / 4 sticks / 32GB 64GB 128GB – all these questions are going to be answered.
RAM overclocking is relatively difficult and not as straightforward as GPU or CPU overclocking. Many factors come into play, and all the timing settings can be overwhelming, even for pro users. This guide will help you on this journey.
Many of the principles described here can be applied to DDR5 memory, but the exact numeral values and tests provided are based on DDR4.
Memory ranks
Understanding ranks architecture is very important. The problem is that manufacturers don’t include this information on RAM packages, so the only way to find out the rank count of your modules is to look online, check forums, use specialized tools like Thaiphoon Burner, or refer to motherboard manuals, which may contain this info in the QVL RAM lists..
It does not matter whether a stick has chips on one or both sides, if it has one 64-bit channel to transfer data, it is single-rank memory. If it has two 64-bit channels, it is dual-rank memory.
It’s hard to go higher than 3600 MT/s with a total of 8 ranks on one motherboard, your CPU mem controller can’t run it. Fewer ranks always mean higher frequency.
When overclocking this is what you may expect:
2R – the highest frequency 4400 MT/s or even up to 5200 (two sticks, single-rank each).
4R – max speed 4000. (four single-rank sticks or two dual-rank sticks).
8R – too high for any CPU, max stable speed is 3600. Some motherboards can’t handle even that (four dual-rank sticks).
For Ryzen CPUs you can’t run AMD FCLK Infinity Fabric in coupled mode higher than 2000MHz (4000MT/s). Going higher decouples the fabric clock and imposes a severe performance penalty that even high speeds can barely compensate for. It works similarly with Intel’s Gear 1 and Gear 2 options, where Gear 2 decouples the bus.
Note: MHz (megahertz) is commonly misused in place of MT/s (megatransfers per second). If the RAM is rated at 3600, it means megatransfers, not megahertz. This directly corresponds to a motherboard bus speed of 1800 megahertz. For DDR4, the megatransfers rate is twice the bus speed.
Best volume: 32, 64 or 128 Gigs?
Building a PC with 16GB of RAM is outdated, so we did not test this configuration, it is simply not enough in 2024. 16GB is the bare minimum for a usable system. Many applications, such as Chrome or basic video editing software, can use more than 16GB. Some games may also exceed this threshold.
Our tests haven’t shown any significant difference between memory volumes of 32, 64 and 128 GBs.
This is because of the tests we used. Performance and game tests rarely saturate 32GB of RAM, let alone 64 or 128.
• AIDA 64
• Cinebench R20
• 3D Mark Time Spy
• Geekbench
• Total War: Three Kingdoms
• Shadow of the Tomb Raider
• CS:GO
• Red Dead Redemption 2
Speaking of volume, our recommendation is to get at least 32GB, with 2 sticks by 16GB each. This setup will handle most of your daily tasks and gaming needs.
However,
If you edit videos, use 3D modeling tools, or work with After Effects, we highly recommend at least 64GB of RAM.
Lastly, for professional setups, especially when editing 4K videos and adding effects, we do see a benefit in having a 128GB setup. You can get by with 64GB just fine, but 128GB is a nice-to-have.
For applications like Maya / 3ds Max / AutoCAD / Nuke / Flame / Unreal Engine and similar, 128GB is preferable. However, many users manage with less and can still achieve their specific small to medium tasks..
RAM used
1) 4*16 GB G.Skill Royal 4266 @ CL 17-18-18-38
Samsung B-Die 5WB dual rank – F4-4266C17D-32GTRGB
amazon link
height 44 mm
depth 8 mm
weight 66 gram
2) 4*16 GB Ballistix MAX RGB 4400 @ CL 19-19-19-43
Micron rev B single rank – BLM2K16G44C19U4BL
amazon link
height 39 mm
depth 9 mm
weight 61 gram
3) 4*32 GB Ballistix RGB 3600 @ CL 16-18-18-38
Micron rev B dual rank – BL2K32G36C16U4BL
amazon link
height 39 mm
depth 7 mm
weight 42 gram
System specs
CPU: AMD Ryzen 5950X (PBO default ON) – amazon link
CPU Cooling: Noctua NH-D15 (BIOS default speed) – amazon link
GPU: NVidia RTX 3090 Asus ROG Strix OC (perf. fan profile) – amazon link
Motherboard: Aorus X570S Master (BIOS F2) – amazon link
Storage: 3x nvme Samsung 980 Pro 2TB – amazon link
PSU: Seasonic TX1000 Titanium – amazon link
Case: Phanteks P600S – – amazon link
6 fans Noiseblocker NB-eLoop X B14-P – amazon link
Win 10 | ReBAR OFF
This review was not sponsored by any vendors. Some of the links in our video descriptions are affiliate, which means at no extra cost to you, we will earn a small commission if you click them and make a purchase. As Amazon Associates, we earn from qualifying purchases. These help us create more content for you. Thank you for your support!
DDR4 timings: primary and secondary (sub-timings)
‘DRAM Calculator for Ryzen’ is an amazing app you can use to properly calculate your primary and secondary timings.
Secondary timings (aka sub-timings) for all of the tests were always tight and optimized, unless stated otherwise.
All tests were run two times with reboots, in questionable cases – three to five times.
Games were run in two modes:
1st 1440p High Graphics, High RTX (where applicable) – reasonable setting most people actually play games with. More stress on GPU, less on RAM.
2nd 1080p Low Graphics, RTX off – low GPU load test to clearly see how RAM influences benchmark results.
4K and above weren’t tested due to marginal variance, as these resolutions are heavily bottlenecked by the GPU.
Test 1 – two sticks SR (single rank) or two sticks DR (dual rank) – 2 vs 4 ranks in total
2R vs 4R has been a debate for a long time, two camps have their pros and cons. When people buy a setup with two sticks, they often face the question of how many ranks their RAM modules should have.
We had to manually downgrade the G.Skill Samsung B-die modules’ tRFC from 288 to 525 to match them to the Ballistix Micron chips which couldn’t boot with lower tRFC than 525. All kits were set to the same primary and sub-timings.
G.Skill Royal 2*16 DR 3600 16-16-16-34 tRFC 525 (2+2 rank)
Vs.
Ballistix MAX 2*16 SR 3600 16-16-16-34 tRFC 525 (1+1 rank)
Vs.
Ballistix MAX 4*16 SR 3600 16-16-16-34 tRFC 525 (1+1+1+1 rank)
AIDA 64
| RAM setup | Read | Write | Copy | Latency |
| G.Skill Royal 2*16 DR (2+2 rank) | 51600 | 50000 | 46700 | 64.3 |
| Ballistix MAX 2*16 SR (1+1 rank) | 54200 +5% | 54140 +8.3% | 48400 +3.6% | 62 +3.7% |
| Ballistix MAX 4*16 SR (1+1+1+1 rank) | 52000 +0.8% | 50100 +0.2% | 47300 +1.3% | 63.4 +1.3% |
Interestingly, for AIDA64, despite some claims that Zen 3 CPU architecture favors 4 ranks over 2, we see a clear win for two single-rank Ballistix modules. Four ranks in total—whether with two G.Skill dual-rank modules or four Ballistix single-rank modules—performed worse. AIDA64 clearly prefers lower rank count systems.
Cinebench R20
| RAM setup | CPU multi-core score |
| G.Skill Royal 2*16 DR (2+2 rank) | 10900 |
| Ballistix MAX 2*16 SR (1+1 rank) | 10940 +0.4% |
| Ballistix MAX 4*16 SR (1+1+1+1 rank) | 10900 |
Cinebench R20 results are very close. Two single ranks are better, but the gain is marginal at just 0.4%, easily wiped out by sample deviation and additional test runs.
Geekbench
| RAM setup | CPU multi-core score |
| G.Skill Royal 2*16 DR (2+2 rank) | 17800 |
| Ballistix MAX 2*16 SR (1+1 rank) | 18020 +1.2% |
| Ballistix MAX 4*16 SR (1+1+1+1 rank) | 18040 +1.3% |
As with Cinebench, Geekbench shows only a marginal difference, though it slightly favors the Ballistix RAM. Rank count makes no difference in performance.
3DMark Time Spy
| RAM setup | CPU multi-core score |
| G.Skill Royal 2*16 DR (2+2 rank) | 15525 +2.6% |
| Ballistix MAX 2*16 SR (1+1 rank) | 15130 |
| Ballistix MAX 4*16 SR (1+1+1+1 rank) | 15600 +3.1% |
3DMark Time Spy is interesting and may actually prove that having 4 ranks in total is indeed beneficial for the Zen 3 architecture.
Unlike AIDA, a purely synthetic test that doesn’t represent real-world performance, 3DMark is known for providing consistent test results that better represent real use cases.
We see that having four ranks is actually beneficial, providing around 3% uplift compared to two ranks, regardless of whether you use 2 dual-rank sticks or 4 single-rank sticks.
Total War: Three Kingdoms
| RAM setup | 1080p – Low graphics | 1440p – High graphics |
| G.Skill Royal 2*16 DR (2+2 rank) | 414 | 156 |
| Ballistix MAX 2*16 SR (1+1 rank) | 424 +2.4% | 157 |
| Ballistix MAX 4*16 SR (1+1+1+1 rank) | 422 +1.9% | 156 |
High graphics, no difference. Low graphics, we see a slight edge of the Ballistix sticks over the G.Skill. This difference does not seem to be related to having 2 or 4 ranks; ranks do not appear to matter.
Shadow of the Tomb Raider
| RAM setup | 1080p – Low graphics – No RTX | 1440p – High graphics – High RTX |
| G.Skill Royal 2*16 DR (2+2 rank) | 273 +3.4% | 123 |
| Ballistix MAX 2*16 SR (1+1 rank) | 277 +4.9% | 123 |
| Ballistix MAX 4*16 SR (1+1+1+1 rank) | 264 | 123 |
SOTR low graphics – two single-rank Ballistix’ showed 1.5% better result than two dual-rank G.Skills, and a 5% better result than four single-rank Ballistix’.
It seems SOTR favors fewer ranks and fewer sticks at the same time. However, this difference is marginal and becomes null if you bump your graphics settings to 1440p High.
CS:GO
| RAM setup | 1080p – Low graphics |
| G.Skill Royal 2*16 DR (2+2 rank) | 707 +0.3% |
| Ballistix MAX 2*16 SR (1+1 rank) | 705 |
| Ballistix MAX 4*16 SR (1+1+1+1 rank) | 712 +1% |
The results are marginal, similar to Cinebench, with little to no relation between different setups.
We see a 1% uplift from using four 1R sticks over two 1R sicks, with two 2R sticks sitting in between.
Red Dead Redemption 2
| RAM setup | 1080p – Low graphics | 1440p – Balanced graphics |
| G.Skill Royal 2*16 DR (2+2 rank) | 174 +0.8% | 117 |
| Ballistix MAX 2*16 SR (1+1 rank) | 172.7 | 117.5 |
| Ballistix MAX 4*16 SR (1+1+1+1 rank) | 174.5 +1% | 117 |
Red Dead Redemption 2: Not much to say here, as all the results are identical. Just like the CS:GO, having two 1R memory sticks showed slightly worse results then the 4R configs.
Conclusions
AIDA in particular and SOTR favor low rank setups.
3DMark favors high rank setups.
Cinebench, GeekBench, Total War, CS:GO don’t care about the rank count.
No definitive answer here: it doesn’t seem that 2 vs 4 ranks matter enough to make a significant performance shift in real-world applications, at least on a Ryzen system.
Based on 3DMark alone, which we consider to be the most balanced test of all, we would recommend getting a 4R setup over a 2R setup. It means that if you’re looking to buy only two sticks of RAM, go with DR sticks like Samsung B-dies.
Two 1R sticks might be suitable if your main goal is to dominate synthetic benchmarks and achieve unparalleled overclocks, which nonetheless will not benefit your daily PC tasks.
The testing involved downgrading the tRFC of Samsung B-die sticks to match them with Micron for fair comparison, so we could test the ranks properly in isolation. Future tests will address the impact of not downgrading the tRFC, which could shift the results in favor of the G.Skill sticks.
Test 2 – four sticks SR or four sticks DR (4 vs 8 ranks in total)
While the previous test showed little to no difference between 2R vs. 4R memory setups, this test can be different. It is known that using 8Rs (4 DR sticks) places significant stress on the CPU controller, along with hindered OC potential it may also result in substantially worse stock performance.
We manually downgraded the tRFC of the G.Skill Samsung B-die modules from 288 to 525 to match the Ballistix Micron chips. All kits were set to the same primary and sub-timings.
Ballistix MAX 4*16 SR 3600 16-16-16-34 tRFC525 (1+1+1+1 rank)
Vs.
G.Skill Royal 4*16 DR 3600 16-16-16-34 tRFC525 (2+2+2+2 rank)
AIDA 64
| RAM setup | Read | Write | Copy | Latency |
| Ballistix MAX 4*16 SR (1+1+1+1 rank) | 52000 +5.5% | 50100 +7.3% | 47300 +9.2% | 63.4 +0.3% |
| G.Skill Royal 4*16 DR (2+2+2+2 rank) | 49300 | 46700 | 43300 | 63.6 |
Although latency is great for both setups, we see a significant drop of ~7% in read/write/copy tests. As expected, having 8 ranks hinders the CPU memory controller. And this is not the case of Samsung vs. Micron architectures, we’ve seen that Samsung B-Die in a 2x2R setup performs on par with Microns with less than 1% deviations. This low score is the result of cramming four 2R sticks together making it a total of 8 ranks.
Cinebench R20
| RAM setup | CPU multi-core score |
| Ballistix MAX 4*16 SR (1+1+1+1 rank) | 10900 +0.6% |
| G.Skill Royal 4*16 DR (2+2+2+2 rank) | 10830 |
CB is rock solid and very hard to influence by any config changes, although marginal at 0.6%, we still see a drop in performance with 8 memory ranks.
Geekbench
| RAM setup | CPU multi-core score |
| Ballistix MAX 4*16 SR (1+1+1+1 rank) | 18040 +3% |
| G.Skill Royal 4*16 DR (2+2+2+2 rank) | 17525 |
Geekbench uses synthetic testing algorithms. A 3% difference, fewer ranks is better.
3DMark Time Spy
| RAM setup | CPU multi-core score |
| Ballistix MAX 4*16 SR (1+1+1+1 rank) | 15600 +2.6% |
| G.Skill Royal 4*16 DR (2+2+2+2 rank) | 15200 |
3DMark, like other tests, shows a 2.6% lower performance with the 8R config. A very insignificant difference to actually matter in real life.
Total War: Three Kingdoms
| RAM setup | 1080p – Low graphics | 1440p – High graphics |
| Ballistix MAX 4*16 SR (1+1+1+1 rank) | 422 +1.7% | 156 +0.6% |
| G.Skill Royal 4*16 DR (2+2+2+2 rank) | 415 | 155 |
TW3K does not favor 8 ranks, a 1.7% difference at low resolutions, and 0.6% at high.
Shadow of the Tomb Raider
| RAM setup | 1080p – Low graphics – No RTX | 1440p – High graphics – High RTX |
| Ballistix MAX 4*16 SR (1+1+1+1 rank) | 264 | 123 |
| G.Skill Royal 4*16 DR (2+2+2+2 rank) | 274 +3.8% | 124 |
SOTR is interesting; for some reason this game doesn’t seem to favor micron sticks in particular. This test is unique and actually shows a 3.8% better result for the 8-rank G.Skill setup.
CS:GO
| RAM setup | 1080p – Low graphics |
| Ballistix MAX 4*16 SR (1+1+1+1 rank) | 712 +1.3% |
| G.Skill Royal 4*16 DR (2+2+2+2 rank) | 703 |
A 1.3% better result for 4 ranks vs. 8 ranks.
Red Dead Redemption 2
| RAM setup | 1080p – Low graphics | 1440p – Balanced graphics |
| Ballistix MAX 4*16 SR (1+1+1+1 rank) | 174.5 +0.9% | 117 |
| G.Skill Royal 4*16 DR (2+2+2+2 rank) | 173 | 116 |
RDR2 at both high and low resolutions shows less than a 1% deviation; fewer ranks still better, though the difference is negligible.
Conclusions
As expected, having 8 ranks in total does pose a penalty, however, that penalty is less than 2% for low resolutions and even less than 1% for high.
If you’re looking for a daily system, having 8Rs will work just fine. However, if you’re aiming for overclocks and supremacy in synth tests, having 8Rs in a no-go. We haven’t started overclocking yet, but we’ll tell you right away: 8 ranks don’t overclock at all.
Sidenote: these results may be misleading and it may seem that having 4*16 SR Ballistix MAX is better in every single test. It is not, remember that we had to manually downgrade the tRFC of our Samsung B-die memory so it could match Micron’s timings to test rank setups. Our next tests will show the full picture.
Test 3 – How tRFC influences the performance
We know one significant difference between the Micron Rev B and Samsung B-die (5WB) chips: Microns can’t handle high tRFC. But how important is tRFC in the first place?
In our previous tests, we had to manually worsen the tRFC for G.Skill Samsung B-Dies to make them perform on par with Microns. Right now, we’ll test the delta and see what would have actually happened if we had kept the Samsung B-dies on their expected low tRFC setting.
G.Skill Royal 2*16 DR 3600 16-16-16-34 tRFC 288
Vs.
G.Skill Royal 2*16 DR 3600 16-16-16-34 tRFC 525
AIDA 64
| RAM setup | Read | Write | Copy | Latency |
| G.Skill Royal 2*16 DR 3600 tRFC 288 | 52850 +2.4% | 50850 +1.7% | 48100 +3% | 61.6 +4.4% |
| G.Skill Royal 2*16 DR 3600 tRFC 525 | 51600 | 50000 | 46700 | 64.3 |
We see around a 2.3% difference in r/w/c tests and around 4.4% change in latency. Lower tRFC is better.
Cinebench R20
| RAM setup | CPU multi-core score |
| G.Skill Royal 2*16 DR 3600 tRFC 288 | 11120 +2% |
| G.Skill Royal 2*16 DR 3600 tRFC 525 | 10900 |
Cinebench has also shown a linear dependency with a 2% performance difference. Lower tRFC is better.
Geekbench
| RAM setup | CPU multi-core score |
| G.Skill Royal 2*16 DR 3600 tRFC 288 | 18335 +3% |
| G.Skill Royal 2*16 DR 3600 tRFC 525 | 17800 |
Somewhat synthetic Geekbench shows a considerable 3% difference, which is the highest result among all tests.
3DMark Time Spy
| RAM setup | CPU multi-core score |
| G.Skill Royal 2*16 DR 3600 tRFC 288 | 15870 +2.2% |
| G.Skill Royal 2*16 DR 3600 tRFC 525 | 15525 |
3DMark Time Spy is similar, 2.2% difference.
Total War: Three Kingdoms
| RAM setup | 1080p – Low graphics | 1440p – High graphics |
| G.Skill Royal 2*16 DR 3600 tRFC 288 | 416 +0.5% | 156.5 |
| G.Skill Royal 2*16 DR 3600 tRFC 525 | 414 | 156 |
TW3K is less responsive to tRFC changes. Only 0.5% difference in framerates.
Shadow of the Tomb Raider
| RAM setup | 1080p – Low graphics – No RTX | 1440p – High graphics – High RTX |
| G.Skill Royal 2*16 DR 3600 tRFC 288 | 275 +0.75% | 123 |
| G.Skill Royal 2*16 DR 3600 tRFC 525 | 273 | 123 |
The same with SOTR, 0.75% difference at low and zero difference at high resolutions.
CS:GO
| RAM setup | 1080p – Low graphics |
| G.Skill Royal 2*16 DR 3600 tRFC 288 | 715 +1.1% |
| G.Skill Royal 2*16 DR 3600 tRFC 525 | 707 |
1.1% difference for CS:GO.
Red Dead Redemption 2
| RAM setup | 1080p – Low graphics | 1440p – Balanced graphics |
| G.Skill Royal 2*16 DR 3600 tRFC 288 | 174 | 117 |
| G.Skill Royal 2*16 DR 3600 tRFC 525 | 174 | 117 |
Amazing result for Red Dead; this game does not care about the tRFC whatsoever, with zero difference at both resolutions.
Conclusions
Changing the tRFC showed a linear dependency in performance: 2-3% for synth tests , and less than 1% for gaming.
Judging by this, the difference in tRFC is marginal and even if you choose Micron memory chips over Samsung B-Dies, you’re still getting decent levels of performance with a penalty that, although existing, is close to marginal.
Test 4 – Micron rev B vs. Samsung B-die (2*16 non-OC)
Let’s run a series of tests to compare the Ballistix MAX Micron rev Bs and the G.Skill Royal Samsung B-dies. In which particular configs either of them is better. We’ll run 4 different configs: 2*16 OC, 2*16 non-OC, 4*16 OC, 4*16 non-OC.
2*16 non-OC – If you were to buy 2 sticks of 16GB each and did not plan to overclock, what would you choose?
This time we won’t align the tRFC, allowing each memory module to run at its full potential.
Ballistix MAX 2*16 SR 3600 16-16-16-34 tRFC 525 – Micron Rev B – 2 ranks
Vs.
G.Skill Royal 2*16 DR 3600 16-16-16-34 tRFC 288 – Samsung B-die – 4 ranks
AIDA 64
| RAM setup | Read | Write | Copy | Latency |
| Micron Rev B – 2*16 SR 3600 tRFC 525 | 54200 +2.6% | 54140 +6.5% | 48400 +0.6% | 62 |
| Samsung B-die – 2*16 DR 3600 tRFC 288 | 52850 | 50850 | 48100 | 61.6 +0.7% |
Despite our G.Skill Royal Samsung B-die having a better tRFC, the Ballistix still showed a 2.5% better r/w/c results in AIDA due to its SR architecture. Latency is better for G.Skill though.
Cinebench R20
| RAM setup | CPU multi-core score |
| Micron Rev B – 2*16 SR 3600 tRFC 525 | 10940 |
| Samsung B-die – 2*16 DR 3600 tRFC 288 | 11120 +1.6% |
CB prefers Samsung B-dies with a better tRFC – 1.6% difference.
Geekbench
| RAM setup | CPU multi-core score |
| Micron Rev B – 2*16 SR 3600 tRFC 525 | 18020 |
| Samsung B-die – 2*16 DR 3600 tRFC 288 | 18335 +1.7% |
Geekbench same as CB. G.Skill winning by 1.7%.
3DMark Time Spy
| RAM setup | CPU multi-core score |
| Micron Rev B – 2*16 SR 3600 tRFC 525 | 15130 |
| Samsung B-die – 2*16 DR 3600 tRFC 288 | 15870 +5% |
G.Skill wins by 5% in 3DMark , quite a big difference here.
Total War: Three Kingdoms
| RAM setup | 1080p – Low graphics | 1440p – High graphics |
| Micron Rev B – 2*16 SR 3600 tRFC 525 | 424 +2% | 157 |
| Samsung B-die – 2*16 DR 3600 tRFC 288 | 416 | 156.5 |
2% better result for the Ballistix in Total War on low settings.
Shadow of the Tomb Raider
| RAM setup | 1080p – Low graphics – No RTX | 1440p – High graphics – High RTX |
| Micron Rev B – 2*16 SR 3600 tRFC 525 | 277 +0.73% | 123 |
| Samsung B-die – 2*16 DR 3600 tRFC 288 | 275 | 123 |
SOTR showed a negligible 0.7% better result for the Ballistix.
CS:GO
| RAM setup | 1080p – Low graphics |
| Micron Rev B – 2*16 SR 3600 tRFC 525 | 705 |
| Samsung B-die – 2*16 DR 3600 tRFC 288 | 715 +1.5% |
Vice versa for the CS:GO: 1.5% better result for the G.Skill.
Red Dead Redemption 2
| RAM setup | 1080p – Low graphics | 1440p – Balanced graphics |
| Micron Rev B – 2*16 SR 3600 tRFC 525 | 172.7 | 117 |
| Samsung B-die – 2*16 DR 3600 tRFC 288 | 174 +0.8% | 117 |
Red Dead seems to favor G.Skill, which is 0.8% better.
Conclusions
In gaming tests, we don’t see a clear preference for any setup; some games prefer single-rank Microns, while others prefer dual-rank Samsungs.
AIDA showed a clear preference for SR memory. However, CB, Geekbench, and 3DMark showed consistent domination of the Samsung B-die DR memory setup.
After all, we know that Ryzen chips do slightly prefer having 4 ranks in total over having 2 ranks. So, if we were to choose a 2*16 setup for our daily non-OC tasks while keeping the memory at stock settings, we’d choose dual-rank Samsung B-dies.
Let’s see if OC makes any difference.
Test 5 – Micron rev B vs. Samsung B-die (2*16 OC)
If you were to buy 2 sticks by 16 and planned to OC, what would you choose?
Now it gets interesting. If we buy DR sticks, motherboards can’t support more than 4000 MT/s for such setups. If you wanna go higher than 4000, you need two SR sticks. You can also use one dual-rank stick, but that switches you to single-channel memory mode, which is pointless to use.
At the same time, AMD’s Infinity Fabric FCLK as well as Intel with their gear 1 and gear 2 settings – do not go any higher than 4000 MT/s (2000 MHz) in paired mode. In our case, the CPU can’t go past 3866, and if you go higher with your mem clock, you have to decouple it from the FCLK clock, which drastically lowers the overall performance.
Our overclocks vary significantly for different chips, which is understandable since different rank configs and the chips themselves can achieve very different results. Two DR Samsung B-die G.Skills can’t achieve an OC higher than 4000 due to having 4 ranks in total (motherboard limit, and we decided not to test one-stick setups). However, their timings are much tighter. Two SR Microns, which have two ranks in total, can achieve much higher clock speeds but architecturally have looser timings, even when using the same speed as Samsungs.
Ballistix MAX 2*16 SR 3866 16-17-16-16-34 tRFC 576, FCLK 1933 1:1 – 2 ranks
highest OC achieved in coupled mode
Vs.
Ballistix MAX 2*16 SR 4666 18-19-19 tRFC 680, FCLK 1933 1:2 – 2 ranks
highest OC achieved in decoupled mode
Vs.
G.Skill Royal 2*16 DR 3866 14-14-14-30 tRFC 308, FCLK 1933 1:1 – 4 ranks
highest OC achieved in coupled mode. We didn’t try decoupling
for G.Skill since motherboards don’t support DR sticks going
higher than 4000 anyway.
AIDA 64
| RAM setup | Read | Write | Copy | Latency |
| Ballistix MAX 2*16 SR 3866 FCLK 1933 1:1 | 57900 +1.4% | 57500 +5.1% | 51100 | 61.2 +5% |
| Ballistix MAX 2*16 SR 4666 FCLK 1933 1:2 | 70000 +22.6% | 59500 +8.8% | 62800 +23% | 64.4 |
| G.Skill Royal 2*16 DR 3866 FCLK 1933 1:1 | 57100 | 54700 | 52100 +2% | 58.7 +8.8% |
At first glance, it may seem that the Ballistix 4666 OC shows tremendous results with 23% higher r/w/c speeds over the competing G.Skill. However, if we look at the latency scores, we see the real picture. Latency is the worst for the 4666 in decoupled mode, while G.Skill is the best with a 9% lead. Tight timings make a big difference here.
Cinebench R20
| RAM setup | CPU multi-core score |
| Ballistix MAX 2*16 SR 3866 FCLK 1933 1:1 | 10900 +9.6% |
| Ballistix MAX 2*16 SR 4666 FCLK 1933 1:2 | 9940 |
| G.Skill Royal 2*16 DR 3866 FCLK 1933 1:1 | 11020 +10.9% |
Cinebench is a perfect example of how bad it can get when you decouple your RAM from the FCLK clock. Even though our Ballistix could achieve 4666 MT/s, it resulted in a 10% worse performance, which is a very significant downgrade and one of the lowest CB scores we got throughout all of our tests. Comparing Ballistix and G.Skill in 3866 coupled modes, G.Skill is still better due to its lower timings.
Geekbench
| RAM setup | CPU multi-core score |
| Ballistix MAX 2*16 SR 3866 FCLK 1933 1:1 | 18190 +6.7% |
| Ballistix MAX 2*16 SR 4666 FCLK 1933 1:2 | 17055 |
| G.Skill Royal 2*16 DR 3866 FCLK 1933 1:1 | 18805 +10.3% |
A very similar result in Geekbench. G.Skill leads the way with 10%, followed by the Ballistix 3866 with a 6.7% better performance. The top OC 4666 Ballistix is again the worst performer.
3DMark Time Spy
| RAM setup | CPU multi-core score |
| Ballistix MAX 2*16 SR 3866 FCLK 1933 1:1 | 15500 +4.7% |
| Ballistix MAX 2*16 SR 4666 FCLK 1933 1:2 | 14800 |
| G.Skill Royal 2*16 DR 3866 FCLK 1933 1:1 | 16320 +10.3% |
Identical results in Time Spy. 4666 OC is still the worst.
Total War: Three Kingdoms
| RAM setup | 1080p – Low graphics | 1440p – High graphics |
| Ballistix MAX 2*16 SR 3866 FCLK 1933 1:1 | 421 | 156 |
| Ballistix MAX 2*16 SR 4666 FCLK 1933 1:2 | 427 +1.4% | 156 |
| G.Skill Royal 2*16 DR 3866 FCLK 1933 1:1 | 427 +1.4% | 156 |
No difference in high res in Total War and a marginal difference of 6 frames in low res. Even the decoupled mode of 4666 overclocked Ballistix showed decent scores.
Shadow of the Tomb Raider
| RAM setup | 1080p – Low graphics – No RTX | 1440p – High graphics – High RTX |
| Ballistix MAX 2*16 SR 3866 FCLK 1933 1:1 | 272 | 123 |
| Ballistix MAX 2*16 SR 4666 FCLK 1933 1:2 | 275 +1.1% | 124 |
| G.Skill Royal 2*16 DR 3866 FCLK 1933 1:1 | 294 +7.9% | 124 |
SOTR has also surprisingly shown decent performance for the decoupled OC. However, the real leader here is G.Skill with a 8% better score in the low res benchmark.
CS:GO
| RAM setup | 1080p – Low graphics |
| Ballistix MAX 2*16 SR 3866 FCLK 1933 1:1 | 703 |
| Ballistix MAX 2*16 SR 4666 FCLK 1933 1:2 | 717 +2% |
| G.Skill Royal 2*16 DR 3866 FCLK 1933 1:1 | 719 +2.3% |
Similar results for CS:GO, G.Skill leading.
Red Dead Redemption 2
| RAM setup | 1080p – Low graphics | 1440p – Balanced graphics |
| Ballistix MAX 2*16 SR 3866 FCLK 1933 1:1 | 173 | 117 |
| Ballistix MAX 2*16 SR 4666 FCLK 1933 1:2 | 174.5 +1% | 117 |
| G.Skill Royal 2*16 DR 3866 FCLK 1933 1:1 | 177 +2.3% | 118 |
G.Skill is still leading, though as always with this test, the margin is slim.
Conclusions
Quite interesting results, though fully expected. We overclocked our Ballistix SR memory to a very high clock of 4666 MT/s, and it still lost by a significant margin to the DR G.Skill Samsung B-dies in every single test, except for AIDA r/w/c, which showed a whopping 23% better score for the Ballistix.
Even though the G.Skill Samsung B-die can’t be overclocked higher than 4000 due to the rank number mobo can support, keeping it in coupled mode and also due to its remarkable timings, the Samsung B-die is the clear winner here.
Buying Microns for a 2×16 setup is not justified. While you can drive them close to 5000 MT/s, it makes no practical sense to do so whatsoever. Unless, of course, overclocks for the sake of overclocks is what you’re looking for.
Test 6 – Micron rev B vs. Samsung B-die (4*16 non-OC)
If you were to buy 4 sticks by 16GB and did not plan to OC, what would you choose?
These results may differ, while G.Skills showed remarkable results in 2*16 setups, using a 4*16 setup poses extreme stress on the CPU mem controller, as 4 sticks of dual-rank RAM make it total an 8-rank setup.
Ballistix MAX 4*16 SR 3600 16-16-16-34 tRFC 525 – Micron Rev B – 4 ranks
Vs.
G.Skill Royal 4*16 DR 3600 16-16-16-34 tRFC 288 – Samsung B-die – 8 ranks
AIDA 64
| RAM setup | Read | Write | Copy | Latency |
| Micron Rev B – 4*16 SR 3600 tRFC 525 | 52000 +4.2% | 50100 +6.2% | 47300 +7% | 63.4 |
| Samsung B-die – 4*16 DR 3600 tRFC 288 | 49900 | 47200 | 44200 | 61.6 +3% |
We do see that AIDA prefers Ballistix over G.Skill in read/write/copy tests; now that we’ve added more ranks, the gap is even wider. However, the latency is still better for the G.Skill.
Cinebench R20
| RAM setup | CPU multi-core score |
| Micron Rev B – 4*16 SR 3600 tRFC 525 | 10900 |
| Samsung B-die – 4*16 DR 3600 tRFC 288 | 11054 +1.5% |
Cinebench favors the better tRFC of G.Skill, even though it is 8 ranks in total.
Geekbench
| RAM setup | CPU multi-core score |
| Micron Rev B – 4*16 SR 3600 tRFC 525 | 18040 +1.1% |
| Samsung B-die – 4*16 DR 3600 tRFC 288 | 17840 |
Geekbench, on the contrary, shows a slightly better result for the Ballistix. Mind that in the previous test where we tested 2*16 setups, G.Skill had an edge of 1.7%, now it loses by 1.1%, so after all, changing the rank count to 8 does negatively impact performance, though only marginally.
3DMark Time Spy
| RAM setup | CPU multi-core score |
| Micron Rev B – 4*16 SR 3600 tRFC 525 | 15600 |
| Samsung B-die – 4*16 DR 3600 tRFC 288 | 15850 +1.6% |
G.Skill is 1.6% better in Time Spy, though it used to be 5% better in the previous 2*16 test, so we confirm again that having 8 ranks instead of 4 ranks is worse.
Total War: Three Kingdoms
| RAM setup | 1080p – Low graphics | 1440p – High graphics |
| Micron Rev B – 4*16 SR 3600 tRFC 525 | 422 +1% | 156 |
| Samsung B-die – 4*16 DR 3600 tRFC 288 | 418 | 156 |
TW3K does not show any significant difference. Micron being a slim 1% better in the low-res test.
Shadow of the Tomb Raider
| RAM setup | 1080p – Low graphics – No RTX | 1440p – High graphics – High RTX |
| Micron Rev B – 4*16 SR 3600 tRFC 525 | 264 | 123 |
| Samsung B-die – 4*16 DR 3600 tRFC 288 | 280 +6% | 124 |
We retested this a couple of times because the result seemed odd, but every single time for some reason SOTR showed a preference for G.Skill in this test, with it being 6% better.
We expected it to be the other way around since, in the previous test, Ballistix was already slightly better, and increasing the rank count of G.Skill to 8 should have widened that gap even further, but it went the other way around. SOTR seems not to like 4 sticks of SR memory for some reason, with the result being 264 fps, when it used to be 277 fps for two sticks. The 8-rank G.Skill setup shows a surprisingly good result here.
CS:GO
| RAM setup | 1080p – Low graphics |
| Micron Rev B – 4*16 SR 3600 tRFC 525 | 712 |
| Samsung B-die – 4*16 DR 3600 tRFC 288 | 717 +0.7% |
A negligible 0.7% advantage for G.Skill. Used to be 1.5%, so increasing the ranks lowered the performance, but again, negligible.
Red Dead Redemption 2
| RAM setup | 1080p – Low graphics | 1440p – Balanced graphics |
| Micron Rev B – 4*16 SR 3600 tRFC 525 | 174.5 | 117 |
| Samsung B-die – 4*16 DR 3600 tRFC 288 | 175 | 117 |
Read Dead – rock solid, no difference worth mentioning.
Conclusions
No clear winner here, but it seems that if you don’t plan to overclock and want to buy a 4*16 setup, going with Samsung B-Die, even though it is 8 ranks combined, is totally safe, and actually due to its superior tRFC, it provides better results in a significant number of tests such as CB, 3Dmark, SOTR, CS:GO.
Test 7 – Micron rev B vs. Samsung B-die (4*16 OC)
If you were to buy 4 sticks by 16 and planned to OC, what would you choose?
OC is something very different. Since our G.Skill 4 sticks, with a total of 8 ranks, can’t maintain a stable OC above 3600, we might see some interesting results compared to the Ballistix 4 sticks – 4 ranks in total which we could overclock to 3866, though at the expense of timings.
Ballistix MAX 4*16 SR 3866 16-17-16-34 tRFC 576 – Micron Rev B – 4 ranks
Vs.
G.Skill Royal 4*16 DR 3600 16-16-16-34 tRFC 288 – Samsung B-die – 8 ranks
We could go 4000 with the Ballistix, but since it required decoupling the memory
from the Infinity Fabric, which consecutively led to bad results, we decided to
stay on 3866 as the optimal highest coupled OC possible with our CPU.
AIDA 64
| RAM setup | Read | Write | Copy | Latency |
| Micron Rev B – 4*16 SR 3866 tRFC 576 | 55500 +11% | 53700 +14% | 50300 +14% | 61.4 +0.3% |
| Samsung B-die – 4*16 DR 3600 tRFC 288 | 49900 | 47200 | 44200 | 61.6 |
Strong win for the Ballistix, not only latency is now on par with the G.Skill, but also we see a strong overall result of up to +14% across the read/write/copy tests. That’s a significant difference, but that’s a synthetic test.
Cinebench R20
| RAM setup | CPU multi-core score |
| Micron Rev B – 4*16 SR 3866 tRFC 576 | 10900 |
| Samsung B-die – 4*16 DR 3600 tRFC 288 | 11054 +1.5% |
CB does not care about the frequency, the scores are identical as prior to the OC, G.Skill slightly better.
Geekbench
| RAM setup | CPU multi-core score |
| Micron Rev B – 4*16 SR 3866 tRFC 576 | 18233 +2% |
| Samsung B-die – 4*16 DR 3600 tRFC 288 | 17840 |
Geekbench favors the high frequency of the Ballistix, which pulls ahead by 2%.
3DMark Time Spy
| RAM setup | CPU multi-core score |
| Micron Rev B – 4*16 SR 3866 tRFC 576 | 15800 |
| Samsung B-die – 4*16 DR 3600 tRFC 288 | 15850 |
Surprising results in Time Spy, almost identical numbers, G.Skill slightly ahead, but that’s the margin of stat error and could easily change with additional test runs.
Total War: Three Kingdoms
| RAM setup | 1080p – Low graphics | 1440p – High graphics |
| Micron Rev B – 4*16 SR 3866 tRFC 576 | 426 +1.9% | 155.5 |
| Samsung B-die – 4*16 DR 3600 tRFC 288 | 418 | 156 |
Ballistix wins by 1.9%, but only in the low-res test.
Shadow of the Tomb Raider
| RAM setup | 1080p – Low graphics – No RTX | 1440p – High graphics – High RTX |
| Micron Rev B – 4*16 SR 3866 tRFC 576 | 277 | 124 |
| Samsung B-die – 4*16 DR 3600 tRFC 288 | 280 +1.1% | 124 |
G.Skill wins, but only by 1.1%.
CS:GO
| RAM setup | 1080p – Low graphics |
| Micron Rev B – 4*16 SR 3866 tRFC 576 | 709 |
| Samsung B-die – 4*16 DR 3600 tRFC 288 | 717 +1.1% |
G.Skill also wins, 1.1%.
Red Dead Redemption 2
| RAM setup | 1080p – Low graphics | 1440p – Balanced graphics |
| Micron Rev B – 4*16 SR 3866 tRFC 576 | 172.5 | 117 |
| Samsung B-die – 4*16 DR 3600 tRFC 288 | 175 +1.5% | 117 |
G.Skill wins again, 1.5%.
Conclusions
The results are quite surprising. We expected the Ballistix to pull ahead significantly. Being SR sticks you have unparalleled opportunity to overclock 4*16 setups to 3866 MT/s and potentially higher if we had a better binned FCLK CPU chip.
The Ballistix did pull ahead in AIDA by 14%, which was a direct result of the increased frequency, and by 2% in Total War and GeekBench. But it marginally lost in Cinebench, 3DMark, CS:GO, Red Dead, and Tomb Raider.
If you increase the frequency, you see improved synthetic test results, this also adds stability risks, raised voltage and temperature concerns in case you want to daily this type of overclock.
All of this makes buying 4 sticks of the Ballistix SR memory and overclocking it, unreasonable as a daily setup. Surprisingly buying 4 sticks of DR Samsung B-die memory with tight timings proves to be a better choice. You can run it daily with stock voltages and in many cases it still provides better results than the overclocked Micron.
This test was particularly important for us because we were deciding whether to sell G.Skills or Ballistix’, and we needed a clear understanding of which to keep, we’ll keep the G.Skills.
Test 8 – 2 by 32 or 4 by 16 ?
Most of you have 4 RAM slots in your motherboards. Imagine you need to have 64GB in total. A question we’ve been asked a lot of times, what is better: two by 32 or four by 16? The initial answer is, of course, 2 by 32, as this leaves you with two empty slots for future upgrades. However, 32GB sticks generally offer worse performance because high-speed 32GB sticks are not available. For example, Samsung B-die memory only comes in 16GB sticks. So, let’s check if there is any performance hit if we abandon the 4 by 16 ideal, and opt for a more reasonable two 32GB setup.
Ballistix MAX 4*16 SR 3600 16-16-16-34 tRFC 525 Micron Rev B – 4 ranks
Vs.
G.Skill Royal 4*16 DR 3600 16-16-16-34 tRFC 288 Samsung B-Die – 8 ranks
Vs.
Ballistix 2*32 DR 3600 16-16-16-34 tRFC 525 Micron Rev B – 4 ranks
AIDA 64
| RAM setup | Read | Write | Copy | Latency |
| Ballistix MAX 4*16 SR 3600 1+1+1+1 rank | 52000 +4.2% | 50100 +6.1% | 47300 +7% | 63.4 +0.8% |
| G.Skill Royal 4*16 DR 3600 2+2+2+2 rank | 49900 | 47200 | 44200 | 61.6 +3.7% |
| Ballistix 2*32 DR 3600 2+2 rank | 51800 +3.8% | 50100 +6.1% | 46900 +6.1% | 63.9 |
Almost identical results for the Ballistix were expected since they share the same hardware architecture. G.Skill is slightly behind due to 8 ranks in total but with a better latency score thanks to a better timings-optimized architecture.
Cinebench R20
| RAM setup | CPU multi-core score |
| Ballistix MAX 4*16 SR 3600 1+1+1+1 rank | 10900 |
| G.Skill Royal 4*16 DR 3600 2+2+2+2 rank | 11054 +1.4% |
| Ballistix 2*32 DR 3600 2+2 rank | 11000 +1% |
Marginal difference with G.Skill being slightly better, Ballistix 2*32 behind, and Ballistix MAX 4*16 the last, but very close.
Geekbench
| RAM setup | CPU multi-core score |
| Ballistix MAX 4*16 SR 3600 1+1+1+1 rank | 18040 +1.1% |
| G.Skill Royal 4*16 DR 3600 2+2+2+2 rank | 17840 |
| Ballistix 2*32 DR 3600 2+2 rank | 17920 +0.45% |
Geekbench is marginal the other way around, the best result for the Ballistix MAX and the worst for G.Skill.
3DMark Time Spy
| RAM setup | CPU multi-core score |
| Ballistix MAX 4*16 SR 3600 1+1+1+1 rank | 15600 +2% |
| G.Skill Royal 4*16 DR 3600 2+2+2+2 rank | 15850 +3.6% |
| Ballistix 2*32 DR 3600 2+2 rank | 15300 |
3DMark is not marginal anymore, G.Skill is 3.6% better than the 2*32 setup of the Ballistix.
Total War: Three Kingdoms
| RAM setup | 1080p – Low graphics | 1440p – High graphics |
| Ballistix MAX 4*16 SR 3600 1+1+1+1 rank | 422 +1.2% | 156 |
| G.Skill Royal 4*16 DR 3600 2+2+2+2 rank | 418 +0.25% | 156 |
| Ballistix 2*32 DR 3600 2+2 rank | 417 | 156.4 |
Total War, marginal, 1.2% better result for the Ballistix MAX.
Shadow of the Tomb Raider
| RAM setup | 1080p – Low graphics – No RTX | 1440p – High graphics – High RTX |
| Ballistix MAX 4*16 SR 3600 1+1+1+1 rank | 264 | 123 |
| G.Skill Royal 4*16 DR 3600 2+2+2+2 rank | 280 +6% | 124 |
| Ballistix 2*32 DR 3600 2+2 rank | 278 +5.3% | 123 |
SOTR Interesting result, while the Ballistix 2*32 and G.Skill showed similar results, the Ballistix MAX was considerably slower.
CS:GO
| RAM setup | 1080p – Low graphics |
| Ballistix MAX 4*16 SR 3600 1+1+1+1 rank | 712 |
| G.Skill Royal 4*16 DR 3600 2+2+2+2 rank | 717 +0.7% |
| Ballistix 2*32 DR 3600 2+2 rank | 713 +0.15% |
CS:GO, marginal difference again, close to the same results for all configs.
Red Dead Redemption 2
| RAM setup | 1080p – Low graphics | 1440p – Balanced graphics |
| Ballistix MAX 4*16 SR 3600 1+1+1+1 rank | 174.5 | 117 |
| G.Skill Royal 4*16 DR 3600 2+2+2+2 rank | 175 | 117 |
| Ballistix 2*32 DR 3600 2+2 rank | 174 | 118 |
Read Dead as always, no difference across the board.
Conclusions
Amazing results, our Ballistix 2*32 setup can easily compete with much more premium 4*16 Samsung and Micron chips. It’s almost a tie across the board, with less than 1% variance here and there, except for Time Spy, which shows a strong 3.6% preference for the 4*16 memory setups.
Considering the fact that running 2 sticks generates less heat and leaves 2 extra slots for a future upgrade to 128GB, the choice of a 2*32 setup seems to be the best here.
Unfortunately, this memory is no longer manufactured, and the only 32GB sticks on the market are Hynix-based with atrocious timings like 3600 18-22-22, that can’t be tightened further. We had this memory in stock and could never push the timings any lower than the stock specification. Let’s test what you can expect from timings of this sort.
Test 9 – Samsung vs Micron vs Hynix; cheap or expensive RAM?
Is it reasonable to look for expensive RAM in the first place? Maybe it’s much easier and better to get whatever is cheaper and spend the money on a CPU or GPU instead?
We are comparing stock frequencies here. But you can also overclock a good bin, whereas a bad bin will not only have worse default specs, but also will hardly ever overclock. So factor that in when making a decision.
Good bin Samsung B-die stock – DR 3600 16-16-16-34 tRFC-288
Vs.
Good bin Samsung B-die OC – DR 3866 14-14-14-30 tRFC-288
Vs.
Bad bin (Samsung, Micron, Hynix) – DR 3200 16-18-18 tRFC-525
Vs.
Hynix medium bin – DR 3600 16-19-19 tRFC-525
Vs.
Hynix medium bin 2021 edition 2*32 setups – DR 3600 18-22-22 tRFC-525 (typical for 32GB sticks)
AIDA 64
| RAM setup | Read | Write | Copy | Latency |
| Sams B-die 3600 16-16-16-34 tRFC-288 | 52850 | 50850 | 48100 | 61.6 +10.5% |
| Sams B-die 3866 14-14-14-30 tRFC-288 | 57100 | 54700 | 52100 | 58.7 +14.7% |
| Bad bin RAM 3200 16-18-18 tRFC-525 | 46500 | 43900 | 41150 | 68.8 – baseline |
| Hynix medium 3600 16-19-19 tRFC-525 | 51200 | 48900 | 45700 | 64.7 +6% |
| Hynix 2*32 – 3600 18-22-22 tRFC-525 | 50700 | 48300 | 44800 | 68.7 +0.1% |
AIDA 64 results are all over the place, but check out the latency, how bad the ‘bad bins’ get. When comparing stock Samsung B-Die RAM to the bad bins, we see up to a 10% difference in performance, and even 15% when overclocked.
Cinebench R20
| RAM setup | CPU multi-core score |
| Sams B-die 3600 16-16-16-34 tRFC-288 | 11120 +1.4% |
| Sams B-die 3866 14-14-14-30 tRFC-288 | 11020 +0.5% |
| Bad bin RAM 3200 16-18-18 tRFC-525 | 10980 +0.1% |
| Hynix medium 3600 16-19-19 tRFC-525 | 10970 – baseline |
| Hynix 2*32 – 3600 18-22-22 tRFC-525 | 11070 +0.9% |
However, CB is totally fine with very little variance. Any RAM will work just fine here.
Geekbench
| RAM setup | CPU multi-core score |
| Sams B-die 3600 16-16-16-34 tRFC-288 | 18335 +7.3% |
| Sams B-die 3866 14-14-14-30 tRFC-288 | 18805 +10% |
| Bad bin RAM 3200 16-18-18 tRFC-525 | 17084 – baseline |
| Hynix medium 3600 16-19-19 tRFC-525 | 17680 +3.5% |
| Hynix 2*32 – 3600 18-22-22 tRFC-525 | 17592 +3% |
GB showed a significant linear dependency with a gap of more than 7% between the bad and good bins. 3200 setup showed the worst result, while 3600 with bad timings performed about 4% worse than the stock good bin Samsung B-Die. Overclocked Samsung B-Die, as expected, performed the best.
3DMark Time Spy
| RAM setup | CPU multi-core score |
| Sams B-die 3600 16-16-16-34 tRFC-288 | 15870 +10% |
| Sams B-die 3866 14-14-14-30 tRFC-288 | 16320 +13% |
| Bad bin RAM 3200 16-18-18 tRFC-525 | 14430 – baseline |
| Hynix medium 3600 16-19-19 tRFC-525 | 15380 +6.6% |
| Hynix 2*32 – 3600 18-22-22 tRFC-525 | 14850 +2.9% |
Direct linear dependency in Time Spy as well, with even bigger gaps. 3200 CL16 is remarkably bad in this test with a gap of 10%.
Total War: Three Kingdoms
| RAM setup | 1080p – Low graphics | 1440p – High graphics |
| Sams B-die 3600 16-16-16-34 tRFC-288 | 416 +3.5% | 156.5 |
| Sams B-die 3866 14-14-14-30 tRFC-288 | 427 +6.2% | 156 |
| Bad bin RAM 3200 16-18-18 tRFC-525 | 402 – baseline | 154 |
| Hynix medium 3600 16-19-19 tRFC-525 | 415 +3.3% | 156 |
| Hynix 2*32 – 3600 18-22-22 tRFC-525 | 407 +1.3% | 155 |
TW has also shown a linear dependency, but the gaps were narrow, 3200 is still the worst though.
Shadow of the Tomb Raider
| RAM setup | 1080p – Low graphics – No RTX | 1440p – High graphics – High RTX |
| Sams B-die 3600 16-16-16-34 tRFC-288 | 275 +7% | 123 |
| Sams B-die 3866 14-14-14-30 tRFC-288 | 294 +14.4% | 124 |
| Bad bin RAM 3200 16-18-18 tRFC-525 | 257 – baseline | 123 |
| Hynix medium 3600 16-19-19 tRFC-525 | 264 +2.7% | 123 |
| Hynix 2*32 – 3600 18-22-22 tRFC-525 | 261 +1.6% | 123 |
SOTR is particularly significant, because it’s not a synthetic test but a real world application. While there’s no difference in the hi-res test due to the GPU being the bottleneck, at 1080p we see some drastic framerate drops of up to 14%. Even if you don’t overclock but compare a decent Samsung B-die bin to a mediocre 3600 bin, you’re still looking at 4% difference in framerates.
CS:GO
| RAM setup | 1080p – Low graphics |
| Sams B-die 3600 16-16-16-34 tRFC-288 | 715 +1.7% |
| Sams B-die 3866 14-14-14-30 tRFC-288 | 719 +2.3% |
| Bad bin RAM 3200 16-18-18 tRFC-525 | 713 +1.4% |
| Hynix medium 3600 16-19-19 tRFC-525 | 708 +0.7% |
| Hynix 2*32 – 3600 18-22-22 tRFC-525 | 703 – baseline |
About a 2% marginal variance across the board.
Red Dead Redemption 2
| RAM setup | 1080p – Low graphics | 1440p – Balanced graphics |
| Sams B-die 3600 16-16-16-34 tRFC-288 | 174 +1.8% | 117 |
| Sams B-die 3866 14-14-14-30 tRFC-288 | 177 +3.5% | 118 |
| Bad bin RAM 3200 16-18-18 tRFC-525 | 171 – baseline | 117 |
| Hynix medium 3600 16-19-19 tRFC-525 | 172 +0.6% | 118 |
| Hynix 2*32 – 3600 18-22-22 tRFC-525 | 172 +0.6% | 116.5 |
Again, a very marginal still linear dependency in Red Dead. Good RAM bins have an advantage.
Conclusions
In general, there’s a significant performance boost when going from 3200 to 3600. So, if you’re in the market for some cheap ram, aim for at least 3600, we don’t recommend getting the 3200 option unless it’s Samsung B-die CL14.
If you can afford a good bin Samsung B-die memory, it is clearly the winner here, plus you can overclock it to go even further. However, typical Hynix ram modules with slightly worse timings like 3600 CL16-19-19 work just fine. Yes they are about 3% slower than the default Samsung B-dies 3600 CL16-16-16, and they can’t be overclocked, but they are almost half the price. If you look at it that way, and you can invest that money into a CPU or GPU – that could be a good decision. Moreover, if you game at 4K resolutions, you will not see any difference whatsoever, even if you opt for the 3200 option.
However if you’re looking for the best of the best, a nicely binned Samsung B-die is the king, no doubt about it.
Test 10 – 3600 CL16 or 3200 CL14
A typical question for people looking to buy Samsung B-die memory kits: should they use 3600 with CL16-16-16 or 3200 with CL14-14-14? These are common configs sold online, with 3200 being slightly cheaper. We’ve already seen in a previous test that 3200 produces worse results, but we tested it with bad timings, so now let’s test a properly binned Samsung B-die.
G.Skill Royal 2*16 DR 3600 16-16-16-36 tRFC 288 Samsung B-Die
Vs.
G.Skill Royal 2*16 DR 3200 14-14-14-34 tRFC 288 Samsung B-Die
AIDA 64
| RAM setup | Read | Write | Copy | Latency |
| 2*16 DR 3600 16-16-16-36 tRFC 288 | 52850 +11% | 50850 +9.1% | 48100 +7.8% | 61.6 +5% |
| 2*16 DR 3200 14-14-14-34 tRFC 288 | 47600 | 46600 | 44600 | 64.8 |
AIDA is very sensitive to the raw frequency, 3600 shows a 5% improvement in latency and about 10% in read/write/copy tests. Huge difference to say the truth.
Cinebench R20
| RAM setup | CPU multi-core score |
| 2*16 DR 3600 16-16-16-36 tRFC 288 | 11120 +0.8% |
| 2*16 DR 3200 14-14-14-34 tRFC 288 | 11033 |
Less than 1% difference, 3600 better.
Geekbench
| RAM setup | CPU multi-core score |
| 2*16 DR 3600 16-16-16-36 tRFC 288 | 18335 +2.4% |
| 2*16 DR 3200 14-14-14-34 tRFC 288 | 17909 |
2.4% difference, 3600 still better.
3DMark Time Spy
| RAM setup | CPU multi-core score |
| 2*16 DR 3600 16-16-16-36 tRFC 288 | 15870 +3.5% |
| 2*16 DR 3200 14-14-14-34 tRFC 288 | 15340 |
3DMark is similar with a 3.5% difference.
Total War: Three Kingdoms
| RAM setup | 1080p – Low graphics | 1440p – High graphics |
| 2*16 DR 3600 16-16-16-36 tRFC 288 | 416 +0.5% | 156.5 |
| 2*16 DR 3200 14-14-14-34 tRFC 288 | 414 | 156 |
Only half a percent difference for Total War.
Shadow of the Tomb Raider
| RAM setup | 1080p – Low graphics – No RTX | 1440p – High graphics – High RTX |
| 2*16 DR 3600 16-16-16-36 tRFC 288 | 275 +1.5% | 123 |
| 2*16 DR 3200 14-14-14-34 tRFC 288 | 271 | 123 |
TR, 3600 still better.
CS:GO
| RAM setup | 1080p – Low graphics |
| 2*16 DR 3600 16-16-16-36 tRFC 288 | 715 +1.1% |
| 2*16 DR 3200 14-14-14-34 tRFC 288 | 707 |
And a similar result for CS:GO.
Red Dead Redemption 2
| RAM setup | 1080p – Low graphics | 1440p – Balanced graphics |
| 2*16 DR 3600 16-16-16-36 tRFC 288 | 174 | 117 |
| 2*16 DR 3200 14-14-14-34 tRFC 288 | 174 | 117 |
Read Dead shows no preference for either.
Conclusions
Synthetic tests showed around 3% consistently better performance for 3600 memory over 3200. However, gaming didn’t, some games showed zero difference, while others showed around 1%.
So, if you have the ability, go ahead and purchase 3600 CL16-16-16, since technically it was better. However, if there is a really good deal on 3200 CL14, it will work as well, besides, you can actually overclock it to match the 3600 version, just give it a bit of extra voltage, like 1.37v or 1.38v.
Test 11 – AUTO timings vs. optimized
The following test is very important. Many people keep their memory on auto timings, they set the frequency to the rated speed and leave it at that. Rarely they go into the primary timings menu, let alone the secondary.
In our case, for our G.Skill Samsung B-die memory rated at 3600 MT/s CL16, our X570S Master motherboard on auto has decided that the best timings to use are 3600 CL26-25-25-58. This is a drastic difference.
Let’s check what happens if you keep your memory on auto and don’t manage to go into the BIOS to set your timings either manually or with XMP. By the way, XMP will be tested separately – spoiler alert: it’s not looking good either.
G.Skill Royal 2*16 DR 3600 Manual 16-16-16-36 + tight secondary
Vs.
G.Skill Royal 2*16 DR 3600 AUTO 26-25-25-58 + AUTO secondary
AIDA 64
| RAM setup | Read | Write | Copy | Latency |
| 2*16 DR 3600 Manual 16-16-16-36 | 52850 +5.6% | 50850 +6.5% | 48100 +11.3% | 61.6 +17% |
| 2*16 DR 3600 AUTO 26-25-25-58 | 50050 | 47730 | 43230 | 72 |
While read, write, copy tests vary between 5% and 11%, the latency is off by 17%, which is the worst result we’ve seen so far.
Cinebench R20
| RAM setup | CPU multi-core score |
| 2*16 DR 3600 Manual 16-16-16-36 | 11120 +0.9% |
| 2*16 DR 3600 AUTO 26-25-25-58 | 11020 |
CB provides a surprisingly good result for auto timings, being only 0.9% worse
Geekbench
| RAM setup | CPU multi-core score |
| 2*16 DR 3600 Manual 16-16-16-36 | 18335 +7.8% |
| 2*16 DR 3600 AUTO 26-25-25-58 | 17006 |
Significant result in GB, 7.8% difference, auto looks really bad.
3DMark Time Spy
| RAM setup | CPU multi-core score |
| 2*16 DR 3600 Manual 16-16-16-36 | 15870 +12.3% |
| 2*16 DR 3600 AUTO 26-25-25-58 | 14130 |
3DMark, one of the most balanced tests of all, shows a tremendous 12.3% difference in performance.
Total War: Three Kingdoms
| RAM setup | 1080p – Low graphics | 1440p – High graphics |
| 2*16 DR 3600 Manual 16-16-16-36 | 416 +5.3% | 156.5 +1.6% |
| 2*16 DR 3600 AUTO 26-25-25-58 | 395 | 154 |
TW3K Even the hi-res test shows a difference of 1.6%, the low-res even bigger of 5.3%. This is a big difference for a gaming test.
Shadow of the Tomb Raider
| RAM setup | 1080p – Low graphics – No RTX | 1440p – High graphics – High RTX |
| 2*16 DR 3600 Manual 16-16-16-36 | 275 +5.4% | 123 |
| 2*16 DR 3600 AUTO 26-25-25-58 | 261 | 123 |
Hi-res shows no difference, while the low-res test is 5.4% better with optimized timings, very similar to Total War. We see a definitive performance hit here.
CS:GO
| RAM setup | 1080p – Low graphics |
| 2*16 DR 3600 Manual 16-16-16-36 | 715 +1.7% |
| 2*16 DR 3600 AUTO 26-25-25-58 | 703 |
1.7% difference, and the same tendency.
Red Dead Redemption 2
| RAM setup | 1080p – Low graphics | 1440p – Balanced graphics |
| 2*16 DR 3600 Manual 16-16-16-36 | 174 +2.4% | 117 |
| 2*16 DR 3600 AUTO 26-25-25-58 | 170 | 118 |
Red Dead doesn’t care much about the RAM, but 170 is the lowest result we’ve got across all setups. Even Red Dead suffers slightly from downgraded memory timings.
Conclusions
The conclusions are shocking. We’ve tested many setups, both cheap and expensive RAM modules with relatively mediocre variance in the results, but after testing a nicely binned memory module with automatically set timings by the motherboard, we received a huge gap. Synth tests fluctuated around 8%, 12%, and 17%, while game tests varied by around 5%. It is remarkable to see how badly you can downgrade your entire system by simply setting your timings to auto. So once again, please use the Ryzen DRAM calculator. If you don’t want to mess with secondary timings, at least set the primary manually. Or maybe use the XMP? Let’s find out.
Test 12 – XMP vs. Manual timings
Should we use the Ryzen DRAM calculator to set the secondary timings, or is the XMP just fine?
In general, XMP sets subtimings to very loose values to accommodate various bins of memory. We know that XMP profile storage has very limited capacity for data, and in many cases, XMP sets only the primary timings, keeping the others on auto. Let’s test XMP vs. manual settings.
For this test, we used 2*32 GB Ballistix modules. One config had the XMP profile enabled, and the other had manually set timings – no fancy overclocks, just a sensible optimization of primary and secondary timings with stock voltages.
Ballistix 2*32 DR 3600 16-16-16-34 tRFC 525 Manual primary + secondary
Vs.
Ballistix 2*32 DR 3600 16-18-18-38 XMP
AIDA 64
| RAM setup | Read | Write | Copy | Latency |
| 2*32 DR 3600 16-16-16-34 optimized 2nd | 51800 | 50100 | 46900 | 63.9 +1.4% |
| 2*32 DR 3600 XMP 16-18-18-38 | 54900 +6% | 54000 +7.8% | 52000 +10.9% | 64.8 |
Right from the start, XMP shows us better read/write/copy speeds of 6%, 8%, and even 11%, but worse latency by 1.4%. We don’t know the reason, but AIDA seems to favor XMP.
Cinebench R20
| RAM setup | CPU multi-core score |
| 2*32 DR 3600 16-16-16-34 optimized 2nd | 11000 +10.9% |
| 2*32 DR 3600 XMP 16-18-18-38 | 9920 |
A powerful drop in performance for the XMP config in CB. We’re looking at 11% here. Surprisingly low score, XMP does magic, this time the other way around delivering the worst result we’ve got so far in this test.
Geekbench
| RAM setup | CPU multi-core score |
| 2*32 DR 3600 16-16-16-34 optimized 2nd | 17920 +8.5% |
| 2*32 DR 3600 XMP 16-18-18-38 | 16520 |
Geekbench is also bad, 8.5% not in favor of the XMP.
3DMark Time Spy
| RAM setup | CPU multi-core score |
| 2*32 DR 3600 16-16-16-34 optimized 2nd | 15300 +6.8% |
| 2*32 DR 3600 XMP 16-18-18-38 | 14330 |
3DMark also bad, non XMP is almost 7% better. Let’s hope that game tests show better results.
Total War: Three Kingdoms
| RAM setup | 1080p – Low graphics | 1440p – High graphics |
| 2*32 DR 3600 16-16-16-34 optimized 2nd | 417 +2.7% | 156.4 |
| 2*32 DR 3600 XMP 16-18-18-38 | 406 | 156 |
TW 1440p almost no difference. 1080p 2.7% worse result for XMP, at least it’s not as bad as the 10% drop in synth tests.
Shadow of the Tomb Raider
| RAM setup | 1080p – Low graphics – No RTX | 1440p – High graphics – High RTX |
| 2*32 DR 3600 16-16-16-34 optimized 2nd | 278 +2.6% | 123 |
| 2*32 DR 3600 XMP 16-18-18-38 | 271 | 123 |
Not as bad either, only 2.6% difference.
CS:GO
| RAM setup | 1080p – Low graphics |
| 2*32 DR 3600 16-16-16-34 optimized 2nd | 713 |
| 2*32 DR 3600 XMP 16-18-18-38 | 718 +0.7% |
This time around the XMP is better, but only by 0.7%.
Red Dead Redemption 2
| RAM setup | 1080p – Low graphics | 1440p – Balanced graphics |
| 2*32 DR 3600 16-16-16-34 optimized 2nd | 174 +0.9% | 118 |
| 2*32 DR 3600 XMP 16-18-18-38 | 172.5 | 118 |
Read Dead as always is rock stable, with a slight 0.9% advantage of the non-XMP manual settings.
Conclusions
In general the game tests aren’t bad. By choosing XMP you’re looking at up to 3% worse performance in low resolution benchmarks. AIDA is unique and dubious, while read/write/copy tests show a preference for XMP, the latency clearly indicates some odd things happening in the background.
Other synthetic tests like CB, GB, 3DMark show alarming numbers. Simply by not setting your sub-timings manually you’re facing a drop in performance of up to 10%.
You may go and purchase some good and expensive RAM, but if you don’t plan to set your primary and secondary timings manually and just use the XMP, this pricey ram will be a waste of money. It will produce results on par with some mediocre kits whose owners managed to log into BIOS and optimize the settings.
Combined charts
AIDA Read is very sensitive to frequency, a very strict linear dependency. Other parameters, like rank setup, decoupling of the FCLK and timings, play a lesser role.
The same applies to AIDA Write, raw frequency matters the most.
Some of the rankings shifted, but we still see a direct correlation with frequency.
Higher frequency is still on top, however, when it comes to latency, we observe that timings also matter. An overclocked and optimized Samsung B-Die wins by a significant margin.
Very little variance in CB, but two results stand out: decoupled memory and XMP show a significant downgrade in speeds.
Geekbench shows a preference for overclocked and optimized setups. Both frequency and timings matter. Samsung B-die 3866 with tight timings is at the top.
Time Spy is very similar to Geekbench, where smart, optimized overclocks are on top. Samsung B-die 3866 CL14 is still the best.
Little variance in the results and the Samsung B-die 3866 CL14 remains on top. It is also observed that this game responds well to frequency, even with some questionable decoupled setups like the Ballistix MAX 4666.
Samsung B-die 3866 CL14 still wins. The game shows a balanced preference for both frequency and timings.
Very little variance in the results, so this test isn’t particularly informative. One thing is certain though: Samsung B-die 3866 CL14 is still on top.
Similar to CS:GO – all setups performed well, Samsung B-die is still on top.
Final words
Price, availability, and performance are major factors here. We won’t give you any advice on price and availability since those fluctuate. But as far as performance goes, here is our advice.
There is no clear answer; different RAM will suit different scenarios, especially when overclocking. Our personal preference may not be in line with your specific use cases.
For a 2*16 setup, our choice is a nicely binned DR Samsung B-die. It has superior timings, it does not overclock as good as the Micron SR memory in terms of frequency, but since current-gen Intel and AMD systems don’t favor clocks over the 4000 MT/s anyway and decouple them, our choice is Samsung B-die (in our case a pair of G-Skills) amazon link .
For a 4*16 setup, our choice is to avoid it; 2*32 makes more sense. But if you still want a 4*16 because it gives up to 3% better performance in some tests, or maybe due to price/availability concerns, then go with four sticks of Samsung B-die. Even though they can’t be overclocked, they still produce a better balanced overall score in our opinion.
For 2*32 or 4*32 setups, there’s only one choice, and it is Micron chips with Ballistix memory modules. All other 32 GB per stick options on the market are primarily manufactured by Hynix and have substantially inferior timings like 18-22-22. The problem is that Micron discontinued their Ballistix product lines and prices on eBbay have skyrocketed – amazon link.