Overclocking, Optimization And Performance
A data rate of DDR4-3808 appears to be the limit of this CPUs memory controller, at least in concert with this motherboard and firmware (which was selected for its DRAM overclocking capability) and a measured 1.35V DIMM slot. That measured output comes at the board's 1.33V setting. Most "enthusiast" motherboards incorporate a +20mV to +30mV memory bias, and there's a fair possibility that highest data-rate modules require these "cheats" just to reach rated settings.
Motherboard and/or CPU limitations are the easiest explanation for how Patriot's 16GB DDR4-3400 reached the same 3808 MHz maximum data rate as G.Skill's low-density DDR4-4000. Higher density usually reduces a memory IC's maximum frequency.
Given its high density, 3400 MHz looks like a fairly solid maximum data rate for G.Skill's 2x 16GB Trident Z DDR4-3000. The same organization (16GB per module at 1GB per IC) had previously limited Corsair's 128GB 8-DIMM kit to a mere 2603 MHz data rate.
|Lowest Stable Timings at 1.35V (Max) on Gigabyte Z170X-Gaming G1 (BIOS F5i)|
|G.Skill Trident Z 32GB F4-3000C15D-32GTZ||15-15-15-30 (1T)||13-13-13-28 (1T)||11-11-11-28 (1T)|
|Patriot Viper 4 16GB PV416G340C6K||16-16-16-32 (1T)||13-13-13-28 (1T)||11-11-11-28 (1T)|
|G.Skill Trident Z 8GB F4-4000C19D-8GTZ||15-15-15-30 (1T)||13-13-13-28 (1T)||10-11-11-28 (1T)|
Surprisingly, the 32GB Trident Z DDR4-3000 kit was stable at tighter DDR4-3200 kits compared to the Viper 4 16GB DDR4-3400. Given the DDR4-3400s greater overclocking capability, we were expecting the opposite result.
|Autodesk 3ds Max 2013||Version 15.0 x64: Space Flyby Mentalray, 248 Frames, 1440x1080|
|WinRAR||Version 5.0: THG-Workload (1.3 GB) to RAR, command line switches "winrar a -r -m3"|
|Grid 2||Steam Version, In-Game Test, High Quality, No AA|
|Battlefield 4||DirectX 11, 100-sec. Fraps "Tashgar", High Quality|
|SiSoftware Sandra||Version 2014.02.20.10, Memory Bandwidth, Cache & Memory Latency Benchmarks|
Another surprise pops up in the Sandra Memory Bandwidth test. We had previously noticed that four single-sided modules outperformed two, in spite of the LGA 1151 platform's dual-channel design. A pair of G.Skill's double-sided modules provide the same benefit in today's test. It appears then that four banks of memory provide the optimal Sandra Memory Bandwidth configuration for this platform, regardless of whether those banks are spread across four modules or two.
A look back at our recent 4x 8GB DRAM review provides further evidence to support the four-bank Sandra optimization hypothesis, with four-bank kits reaching 39 GB/s to 41 GB/s regardless of whether those banks are situated on two or four DIMMs.
The 2x 16GB (32GB) Trident Z kit appears far weaker in Sandra Latency than it had in Sandra Bandwidth, though it almost catches up to Patriot's 2x 8GB (16GB) kit at optimized DDR4-3200 timings.
Real World Benchmarks
Unfortunately, none of our current real-world benchmarks are fully capable of taking advantage of DDR4's vast bandwidth. Grid 2 for example is able to show big gains when switching from DDR4-2133 to DDR4-2666, but only at super-light GPU loads where the extra frame rate has no real-world benefit.
The 32GB DDR4-3000 kit consistently falls behind the 8GB DDR4-4000 kit in bandwidth-sensitive benchmarks, but the average performance gap is less than one half of one percent. In this case, the vast majority of users get more benefit from the slower memory's added capacity.
Why would anyone pay for more RAM than they're ever going to use? Doing so looks bad on a price-per-performance chart, yet we're certain most of our readers will eventually need more than 16GB.
More capacity is, after all, the point of G.Skill's Trident Z 2x 16GB (32GB) DDR4-3000 kit. And when we compare the performance/price metric to capacity, we finally see value leadership.