HyperX HX426C15FBK4/32 Review

It is now time for us to test what sort of overclocking numbers one can achieve with our kit. As usual, in this task we are accompanied by a Haswell-E based setup as alternatives simply do not yet exist.

 CPU  Intel Core i7-5960X
 (running at 4 GHz CPU and 3.5 GHz cache clocks)
 Motherboard  MSI X99A Xpower (BIOS version T.BB)
 Memory  Kingston HyperX Fury HX426C15FBK4/32
 Graphics card  ASUS HD4350
 Storage  Kingston SSDNow V300 60GB
 Power supply  Seasonic Platinum 1200W
 Operating System  Windows 7 64-bit Service Pack 1

There are many ways to test stability of the system but our method of choice is HCI Memtest as it is the toughest memory stresstest that we are aware of. As we are dealing with a 32GB kit on a CPU that can handle 16-threaded load, we use sixteen 1500MB instances and call things stable if we see all of them run past 150% without showing a single error. Each pass of such a test takes about 45 minutes to complete.

We limit our testing procedure to the DDR4-2666, -2800, -3000 and -3200 modes, being limited from reaching higher frequencies by the capabilities of memory controller inside our CPU. Once the memory frequency is set, we seek for optimal combinations of primary timings, for which we minimise the voltage in 0.01V steps.

HX426C15FBK4-32_results

The overclocking properties of Hynix-based memory modules can generally be described as follows. If CAS latency is kept without change, then there is a strong, almost linear dependency between stable frequency and minimal stable voltage. This dependency can be seen on the diagram above.

Even though CL18 gives a further frequency and/or voltage improvement above CL16, its practical relevance is currently low due to stable frequency being capped at 1600MHz by our test platform. For this reason, CL18 results are omitted from our testing.

What comes to the rest of primary timings (tRCD, tRP and tRAS): their minimal stable values depend primarily on the memory frequency with voltage playing a secondary role in the borderline areas.

The overall picture can be observed on the diagram below:

1333MHz
(DDR4-2666)
1400MHz
(DDR4-2800)
1500MHz
(DDR4-3000)
1600MHz
(DDR4-3200)
 CL10
not possible
not possible
not possible
not possible
 CL11
11-12-14-14@1.48v
not possible
not possible
not possible
 CL12
12-13-14-14@1.35v
12-13-15-16@1.41v
not possible
not possible
 CL13
13-13-14-14@1.26v
13-13-15-16@1.32v
13-14-16-16@1.40v
not possible
 CL14
14-13-14-14@1.20v
14-14-15-16@1.24v
14-14-16-16@1.31v
14-15-17-18@1.40v*
 CL15
not relevant
15-14-15-16@1.18v
15-14-16-16@1.24v
15-15-17-18@1.30v*
 CL16
not relevant
not relevant
16-15-16-16@1.19v
16-15-17-18@1.27v*
*only stable in dual-channel

Our sample has no problems with clearing its spec. In fact, the rated 1.20V were sufficient to run DDR4-2666 CL14 or DDR4-2800 CL15, the latter still with headroom to spare.

Increasing the voltage to 1.30-1.32V we were able to achieve 1400MHz CL13 and 1500MHz CL14, both of which is something that only the truly high end DDR4 kits can currently offer. With additional 0.1V, a further drop of CAS latency was also possible. At the end of the day, our best result with the kit ended up being 1500MHz at 13-14-16-16, which required 1.40V for full stability. We also had DDR4-3000 with CL12 within reach, but the kit could simply not cope with 1.5V+ that were required to make it happen.

We could not fully stabilise our platform at 1600MHz in quad channel. However, as any combination of two modules would easily run DDR4-3200 in dual channel, we believe that our test platform should take the blame for this failure. Still, running two worst modules in dual channel in the worst slots B and D, we get a rather good estimate of what the kit could do, had our platform been able to run DDR4-3200 in quad.

To show the performance gain achieved by memory overclocking, we set up a quick demonstration using XTU, CineBench R15 and AIDA64 Cache&Memory test. Note that we also took care of the secondary timings on the overclocked results.

HX426C15FBK4-32_performanceDEF HX426C15FBK4-32_performanceOC

As we argued many times in the past, memory overclocking on LGA2011-3 is mainly a thing of prestige as the memory controller of Haswell-E processors simply cannot fully utilize all the available bandwidth.

Leave a comment

 

  1. Thanks for testing this ram not even Kingston list the latency of this ram and i wanted to know for purchase decision

    thanks