For many years, it’s been assumed by a lot of people that buying high-end memory, or rather high frequency/low latency memory, is a waste of money. Obviously, however, there’s much more to it than that. In GPU-bound situations, what applies to the CPU will apply to memory as well. As such, if there’s merit in upgrading your CPU, then surely there’s merit in buying the more high-end memory as well.
Unlike with a CPU, you can’t readily tell what kind of memory a system has (you can see capacity or speed, but nothing about sub timings, latency, etc. from your Windows system info). However, you’d be surprised just how much faster a system with highly tuned or configured memory is as opposed to one where nothing has been tuned or configured accordingly.
In this feature, the point is to highlight this difference, not in productivity terms, but for games. Naturally, lower-end systems benefit more from high-end memory but by the same token, this holds true as well for the high-end machines at the opposite end of the spectrum – particularly for those who game at 1080p and prefer high frame rates (144Hz monitors etc) rather than the high pixel density and low scan rates. Ideally, you want as high a resolution as possible with the highest frame rate as possible, in which case you’d still need a powerful GPU in addition to a beefy CPU and equally fast memory to feed both.
To better show the differences that memory makes in a gaming system, though, we chose to use the ill-received Intel Core i3 7350K CPU. Based on the Kabylake-S architecture, this CPU has two physical cores, 4MiB of L3 cache, supports hyper-threading, and of course is for the Z170/270 LGA 1151 platform. It’s a better solution than what the Core i3 6420 was providing in terms of performance, and certainly beats those dual-core Pentiums Intel was offering not too long ago. Those CPUs are simply too slow for any meaningful gaming or for use in today’s operating systems, where they quickly run out of steam at anything resembling a modern title.
For the motherboard we chose a small, but capable Z270 based Micro-ATX board. The AORUS X270MX – Gaming5 is a no-nonsense board that strips away a lot of the luxuries, so it can come in at an attractive price point. That said, it will still support DDR4 3600MHz+ memory so if you have a combination of this board, the 7350K CPU, and a memory kit such as the Trident Z which manufacturer G.Skill provided for this editorial, then you’ve a lot of performance to unlock (assuming you haven’t already).
Obviously on a better motherboard the memory will scale even farther, but the odds of one buying a 7350K and something like an ASUS Maximus IX Apex, for example, are low, simply because the board costs a lot more than the CPU, which is rather difficult to justify. As such we chose this board of choice for the Trident Z. This is a Samsung B-die based kit, which means it’s not only capable of high frequencies, but low latencies as well.
It should be noted that merely having memory with these ICs will not do anything for performance, contrary to some claims. They are no different to what you’ll get from a similarly spec’d kit using XFR ICs from Hynix. The magic is in what the Samsung ICs allow you to do when you get down to the tuning options. They have a very high voltage tolerance along with a wide frequency and memory timings band.
If you buy high-end memory from G.Skill or any other vendor, it’s imperative that you manually tune this memory to get the most out of it. If you don’t, then you’ve literally wasted your money and might as well have bought a generic kit instead. The memory merely being B-die does not mean anything unless you roll your sleeves up and get down to the actual tuning.
CPU: Intel Core i3 7350K (courtesy of GIGABYTE)
Motherboard: Aorus Z270MX – Gaming5 (courtesy of GIGABYTE)
GPU: EVGA GTX 980 TI K|NGP|N Edition (courtesy of EVGA)
HDD: Corsair Neutron XTi 480GB (courtesy of Corsair)
PSU: Corsair AX1500i (courtesy of Corsair)
OS: Windows 10 x64
All tests performed at 1920 x 1080
(F4 BIOS/ FW 382.05)
Testing was done in an easy-to-follow manner so that you can try this for yourself should you have similar components. If not, your chosen memory kit and motherboard should exhibit similar behaviour, and so the same performance figures are a relevant guideline for a wide variety of systems regardless of which brand or SKU you’ve bought (especially if it uses the same ICs). What sets the G.Skill Trident Z kit apart is that this level of performance is guaranteed, especially since at no point did we operate the memory at its rated 3,866MHz frequency, mostly due to the limitations of the motherboard. Not an issue in and of itself, because we were still able to reach frequencies as high as 4,266MHz C19 on the ASUS Maximus IX Apex, using the same kit and a 7700K CPU (see the screenshot below) for complete 24/7 stability.
For this test on the Gaming5, and for the sake of consistency, we disabled turbo for the CPU, so that the 7350K remained at its base 4.2GHz clock throughout, except in the overclocked scenario. Here we used a fairly high 4.8GHz lock. This required a fair bit of voltage at roughly 1.35V. Still safe for continuous use but be mindful of the fact that 7350K CPUs generally clock poorly compared with the Core i5s and i7s.
Even though the default XMP frequency for this memory kit is 3,866MHz, we loaded it anyway, but changed the CPU multiplier to 32 so we could end up with 3,200MHz. For the rest of the settings, we manually tuned the DRAM going as low as 12-12-12-34 at 3000MHz, with a nominal voltage of 1.45V (the power of B-die ICs) on the low side. On the high side, we stopped at 3,600MHz C13-13-13-36 again employing the same 1.45V. Do note that the DRAM is perfectly safe at this operating voltage and those into competitive overclocking often use similar memory at voltages as high as 1.9V without any issue. So, don’t be afraid, but at the same time always exercise caution and don’t exceed 1.5V for your daily use. The beefy heatsinks on the G.Skill Trident Z modules will keep the DRAM cool enough (and wouldn’t be an issue even without the heatsinks) to ensure long life and reliable operation.
As you’d expect, the regular 3,200MHz setting delivers the lowest performance which would still be far ahead of DDR4 2133MHz, but nowhere near where the other configurations are. Games scale independently from one another so there are situations where one setting is better than another for a particular game. By and large, though, the rule is that performance is best with the highest frequency possible at the lowest timings possible.
Batman: Arkham Knight
The better the memory subsystem, the better the performance here, and it scales in a predictable way with memory performance in AIDA64 (not shown). Moving from the 3200MHz XMP setting to 3,333MHz C13 yields 7.8% or 11fps difference in performance. Remember, if you want to play at that 144Hz, just overclocking the CPU alone may not get you there (4.8GHz with 3,200MHz memory is slower than a 4.2GHz CPU clock with 3,333MHz C13 memory). This title has some relatively heavy reliance on memory subsystem performance and this holds true for other late UE3.X engine based games.
Far Cry Primal
Based on the Dunia engine (much like the upcoming 2018 Far Cry 5), this is another one that has some respectable gains from memory tuning and frequencies. Less so than Batman, but still enough to warrant the effort with 7fps (5,8%) gain from the 3200XMP to 3,333MHz C13. CPU clock cycles are of major importance here as well. When mixed together with the right memory settings, there’s a significant performance boost from 119fps to 134fps for a total 12% gain in performance, with the CPU and DRAM tuning/overclocking delivering 6% of performance over the baseline figure respectively.
Grand Theft Auto V
A popular title still, if not the most popular in this suit, Grand Theft Auto V can be very taxing on low-end systems and these results show it. The CPU is limiting performance and while the memory subsystem tuning helps, it doesn’t offer much at first. It’s only when you use the memory tuning in combination with the CPU overclock that you get performance near the 100fps mark. This benchmark is actually memory sensitive, but it needs the right CPU for that to show. Once the CPU limitation is lifted, then you’re able to see the huge gains that memory tuning provides. With the CPU out of the equation, the difference between the worst and best score is 4.7% or just over 3.5fps which isn’t much of anything at all. With the CPU at 4.8GHz the memory tuning and overclocking makes itself visible, and the difference suddenly grows to 14% or about 12fps – our biggest performance difference yet. For a game which at first seemed to not scale much with memory tuning, this is a significant boost in performance. Overall, the difference between the lowest score and the highest is almost 20fps, and most of it due to memory, but for you to see that you need an overclocked CPU.
One of two DX12 titles, Hitman is built with the Glacier engine, which supports all the typical rendering tricks and advanced features we’ve come to expect from modern titles. A beautiful looking game artistically, even if it’s unequal in many areas technically. It’s still a relevant game with a fair balance between the GPU and CPU performance, though. From the worst 3200XMP score to the best C13 3333MHz score, performance scales by 5.2% (3.6fps) versus 7.8% (6.8fps) for isolated CPU overclocking. Much like GTA V, when the game is not entirely CPU-bound, the memory starts to play a slightly bigger role. At 4.8GHz, the differences between the slowest memory configuration and the fastest is 6.3% (5.9fps). Putting it all together is literally an addition of the memory and CPU overclock for a 14.6% boost in performance or 12.7fps.
Rise of The Tomb Raider
Here’s another DX12 title, but one which behaves more like GTA V than Hitman. The difference between the slowest and best score when considering memory exclusively is a mere 3.5fps (2.6%), and between 4.2 and 4.8GHz nearly 10fps or 6.9%. With the CPU at 4.8GHz and the memory at 3,333MHZ C13, the performance jumps up an additional 6.1% (9fps). In total, performance discrepancy between the worst and the best is 18.5fps or 13.56%
Total War: Attila
An older title, but one that’s still able to bring many machines to a crawl (see AMD Ryzen about this). The performance in low overall, but for the type of game it isn’t too much of a problem. Ideally here you want a true four-core CPU or better, but the 7350K still benefits from some tuning. CPU frequency alone accounts for 2.5fps or 5.11%, while the memory performance delta between the worst and best at 4.2GHz is 2.3fps or 5.09%
This particular game found favour with the 3,466MHz setting for some reason – probably an anomaly but one that should be reported, anyway. Given that the differences between the other memory speeds and configurations is so low, it’s likely that even with the overclock, the CPU is not able to keep up with the game and as such is hampering performance throughout, regardless of configuration. Whatever the reason, pure performance gain from the worst to the best is 13.9% or 6.3fps
What you’ll notice is that the best performing memory setting wasn’t the highest frequency or lowest timings, but a combination of both. C12 3,000 wasn’t good enough, but neither was 3,600MHz C15. The sweet spot was 3,333MHz C13. There are a great many reasons for this, but in part it’s how the CPU memory controller deals with memory transactions. At a certain point, more bandwidth yields diminishing returns, and lower latencies are required. Memory performance scaling, or rather why it scales that way, is extremely complicated and in fact is the most challenging thing to deal with for both Intel and board vendors alike. The trade-off between latency and frequency is a moving target. In future, we’ll do a more detailed memory breakdown analysis so you can readily see how memory timings and frequency play off each other, and just what kind of differences it makes to the system as a whole, even outside of gaming – especially on a high-end workstation/content creation machine where saving a few seconds adds up to minutes and hours eventually.
Until then, we’d like to thank G.Skill for providing us with the Trident Z F4-3866C18D-16GTZ. The kit is available locally and can be sourced from Wootware directly.