Overclocking, Bandwidth And Latency
My Haswell-based processor's memory controller lost stability around 1.37V. But the highest voltage required by any DDR4 I’ve tested to date has been 1.35V. Sticking with the 1.35V as the limit, I got out my voltmeter and found this voltage at the motherboard’s 1.330V setting.
Best Stable Timings
With the help of some added voltage, I was able to push both 32GB samples to DDR4-2400 CAS 12. That’s a little more latency than the kit I use for motherboard testing, but we also expect that under the load of higher capacities. Remember that the Adata kit rated for DDR4-2400 was also rated at a significantly slower CAS 16.
Crucial’s bare modules also reached a remarkably low CAS 10 at DDR4-2133. That’s one-third quicker than its rated timings.
Adata’s 32GB DDR4-2400 reached 3000MT/s at CAS 15, but no more. Slowing it down to CAS 16 didn’t help, though playing with secondary and tertiary timings may have. Generally speaking, DDR4-2400 XMP values have tighter secondary and tertiary timings than DDR4-3000-rated parts, and that was also true when compared to G.Skill’s 16GB DDR4-3000.
Crucial’s DDR4-2400 reached DDR4-2838 by increasing the 125MHz BCLK to 129MHz while using a 22x multiplier. By comparison, G.Skill’s smaller modules achieved a 3168MT/s data rate at a 132MHz BCLK and 24x multiplier.
Remember what I said about tighter secondary and tertiary timings stopping Adata’s DDR4-2400 at DDR4-3000? Those same tight timings give it a boost in Sandra's Memory Bandwidth module, pushing it will beyond the performance level of G.Skill’s lower-capacity modules that overclock more aggressively.
Crucial’s DDR4-2133 wins the DDR4-2400 bandwidth race, likely for the same reason. Slower-rated data rates allow memory manufacturers to choose tighter secondary and tertiary timings.
Latency is the inverse of frequency, so fewer cycles allow quicker response times. G.Skill’s lower-capacity modules do well by DDR4 standards, but we’re still seeing 20% latency penalties compared to high-end DDR3.