Results!
As we continue to use the same test platform and software from our earliest DDR4 round-up, it’s important to remember that DDR4-2400 runs at 9x the memory controller frequency, and that DDR4-3000 uses that same ratio with an overclocked BCLK. Using a 4:3 ratio between the memory controller and CPU core-side BCLK and x2 for DDR data rate, the full multiple for DDR4-2400 is 9 x 4:3 x 100 x 2, while the full multiple of DDR4-3000 is 9 x 4:3 x 125 x 2. Without extreme tuning tactics, the highest stable memory multiplier for Haswell-E on most motherboards is 10, allowing DDR4-2666 at 100MHz BCLK and DDR4-3333 at 125MHz BCLK.
The highest sustainable BCLK on this CPU is 132MHz, which boosts DDR4-3000 to a data rate of 3168 MT/s. Going beyond that requires selection of the next stable memory multiplier, followed by clocking down from DDR4-3333 to the highest stable data rate.
Klevv's Cras DDR4-3000 kit clocks past Kingston’s HyperX Predator, yet doesn’t quite reach the highs of G.Skill’s DDR4-3000 C15. It also doesn't scale past 3168 MT/s, which means there's no need to try the next-highest (DDR4-3333) ratio. Proving that its rated timings are conservative, Klevv's Cras was even stable at 3144 MT/s and reduced 16-16-16-32 latencies.
XMP timings hurt Klevv's Cras a little in Sandra's Bandwidth benchmark, but it easily catches up to Kingston’s HyperX Predator when timings are optimized.
Lower is better for latency, and we again see where a slight loss at XMP defaults gives way to a performance tie after timing optimization.
Grid 2 is particularly harsh on Klevv’s XMP timings, and a restart didn’t fix this benchmark. Optimized timings did though, allowing the Cras DDR4-3000 to catch its Kingston rival. Other benchmarks were barely affected and, come to think of it, we don’t know anyone who could see the difference between 219 and 232 FPS.
The only disconnect is price. For $150 more than its closest rival (Kingston's Predator), Klevv's Cras adds slightly fancier heat spreaders, eight LEDs and a piece of acrylic to show off the LEDs. It also overclocks a little better, but not by enough to catch G.Skill’s cheaply-clad Ripjaws 4.
