Testing Method & Test Setup
There are many platforms that will easily run our memory as DDR3-2133, but Intel's Ivy Bridge is only one that's capable of producing full stability figures north of 1300MHz. Speaking of full stability – we define that as ability to pass eight 1500MB instances of HCI Memtest without a single error at least up to 150% each.
||ASUS Maximus V Gene (BIOS 0086)
||Intel Core i7-3770K
||XFX 8600 GT
||Kingston HyperX T1 KHX21C11T1K2/16X
||Samsung 40 GB
||Windows 7, 64 bit SP1
Having started testing, we immediately encountered a problem. To have our testing procedure as close to real-life as possible,
we only want to alter the four primariy timings. But on these Kingstons, keeping all the subtimings
on Auto made it impossible to even do 800MHz, so we had to make manual adjustments.
First, we found that using Ref Cycle Time (tRFC) at 300 is a must. Plus, in the range from 800 to 1200MHz it's best to set both Read to Pre (tRTP) and Write to Read Delay (tWTR) to either 9 or 10. Above 1200MHz, tRTP and tWTR were defaulting to 10 and did not cause any problems.
Of course, loosening those subtimings to such extent had a negative impact on performance, but at least that allowed us to have predictable overclocking experience. Besides, on the other 8Gb sticks we recently tested (the 2400C10 TridentX) tRFC, tRTP and tWTR were automatically set loose all the time.
As always with Hynix or Samsung ICs, second and third primary timings (tRCD and tRP) are eventually going to cap the stable frequency regardless of other timings or voltage.
Our four-day memtest marathon has also shown that, given no limitation by tRCD and tRTP, voltage allows Hynix MFR to scale nicely and achieve impressive frequencies. For example, 1.75V made our kit stable above DDR3-2600 which is higher than any 8GB modules you can currently buy. Not that such frequency is practical for a day-to-day use, but it shows how
much potential there is yet to be discovered.
The scaling doesn't actually stop at 1.75V, but we didn't extend our testing beyond that mark as we are not sure what effects such sort of voltage might have in long-term use. Instead, we made a couple of high-frequency SuperPi 32M runs and a DDR3-2937 validation shot
, which were all limited by the capabilities of our memory controller.
Comparing these results to what we had on Samsung-based TridentX, it's not hard to see that the only advantage of Hynix MFR lies in ability to work above 1'230 MHz. As such memory frequencies are not conveniently achievable, given choice between the two, we would recommend going for the Samsung as a more reliable option.
Discuss this article in the forums