Comparing DDR4 IC on Z170 platform – part 3 – 1.55 volts, safe and brave :)

This is the third post in DDR4 on Z170 series – for previous parts see part 1 and part 2, where I cover stock and safe settings, respectively. In this part I’m going to use settings which are beyond everyone’s safe range, however in my opinion this is just about max I could use 24/7 in my system with adequate cooling (at least 120mm blowing directly over RAM). This is quite brave, still not extreme though as we’ll see in the next parts. The settings from previous part have been optimized further, which is summarized in the table:

Changes

Let me quickly summarize what changes in the memory settings:

Memory	Original frequency	Original timings	Voltage	Optimized frequency	Optimized timings	New voltage	OC frequency	OC timings	OC voltage
G.Skill TridentZ 3600C17	1800 MHz (3600 effective)	17-18-18-38 2T	1.35V	1800 MHz (3600 effective)	15-15-15-28 1T	1.4V	1800 MHz (3600 effective)	13-13-13-28 1T	1.55V
G.Skill RipjawsV 3200C16	1600 MHz (3200 effective)	16-16-16-36 2T	1.35V	1733 MHz (3466 effective)	15-17-17-28 1T	1.4V	1800 MHz (3600 effective)	13-17-17-28 1T	1.55V
HyperX Savage 2666C13	1333 MHz (1666 effective)	14-15-15-35 2T	1.35V	1400 MHz (2800 effective)	12-14-14-28 1T	1.4V	1500 MHz (3000 effective)	13-15-15-28 1T	1.5V
				1466 MHz (2933 effective)	13-14-14-28 1T	1.45V	1533 MHz (3066 effective)	12-15-15-28 1T	1.55V

In addition to main timings and frequency change, I also tighten secondary and tertiary timings, so they are quite aggressive. This includes tCWL changed to 9 and – in case of B-die – TRFC to 260.

Test results are:

XTU

The progress continues in case of Samsung-based mems, both of them gain 15 points by moving from CL15 to CL13. Hynix break 1400 barrier, the less aggressive setting closing on to the tighter one.

Geekbench

General results in Geekbench 3 are the same with Samsung ahead of Hynix, however 3200 caught up on 3600 despite of advantage of tighter TRCD/TRP on B-die. It is also possible that I experienced bad RTL training on B-die while testing this. In Memory subtests HyperX are now far behind due to too low operating frequency and it is also reflected in the overall score.

AIDA

Latency test confirms theoretical improvement, both G.Skill mems improved 2 ns since previous test, while both HyperX gained “only” 1 ns.

Synthetic bandwidth tests also improve, B-die are almost maxed out so they don’t gain much, E-die almost catch them, while Hynix show significant boost, especially in first (less aggressive) setting – over 4 GB/s in each test.

SuperPi 32M

Closing test is SuperPi 32M, where both Samsungs gain additional 3 seconds. Results for both HyperX settings are very similar and they don’t differ much from previous results, which might suggest that we’re reaching performance limit of Hynix MFR on air.

Voltage scaling

Starting from this part, I will show collective results for XTU and 32M as real life benchmarks on one chart to visualize voltage scaling. I went the easy way with charts here by cutting out the part which is not yet described.

XTU

32M

This concludes tests with safe voltage. I was able to demonstrate possible gains in benchmarks, which were visible, but not as impressive when switching from XMP to optimized settings. Next part will include another voltage raise to 1.7V which is definitely beyond safe settings – remember that 1.65V was officially maximum recommended voltage to be used on previous generation of memory (DDR3). It will also be last part, where HyperX will be included as maximum of MFR on air will be reached.

Comparing DDR4 IC on Z170 platform – part 2 – 24/7 overclocking setting

This post is a second part in DDR4 on Z170 series. In previous part I covered the basic setup and baseline for tests. Here I will begin overclocking by raising memory voltage to 1.4-1.45 volts, which seems quite safe for 24/7 usage. In addition timing settings taken from XMP profile will be optimized and tightened.

Changes

Let me quickly summarize what changes in the memory settings:

Memory	Original frequency	Original timings	Voltage	Optimized frequency	Optimized timings	New voltage
G.Skill TridentZ 3600C17	1800 MHz (3600 effective)	17-18-18-38 2T	1.35V	1800 MHz (3600 effective)	15-15-15-28 1T	1.4V
G.Skill RipjawsV 3200C16	1600 MHz (3200 effective)	16-16-16-36 2T	1.35V	1733 MHz (3466 effective)	15-17-17-28 1T	1.4V
HyperX Savage 2666C13	1333 MHz (1666 effective)	14-15-15-35 2T	1.35V	1400 MHz (2800 effective)	12-14-14-28 1T	1.4V
				1466 MHz (2933 effective)	13-14-14-28 1T	1.45V

In addition to main timings change as mentioned in the table above, subtimings were also optimized, including tightening of TCWL to 13 cycles, TRFC to 280 cycles and TRTP to 8 cycles. These settings are far more aggressive than default, although they are not still the tightest settings possible. Here’s the screenshot from the OS with TridentZ installed (most of the subtimings are identical for other sticks):

On to the tests:

XTU

Compared to previous results, we can observe dramatic result improvement in all cases. XTU is very sensitive to memory setting and optimizations made here cause ~8% increase in achieved results. Judging from experience, this is a difference of around 200-300 MHz core clock. What’s also interesting is that RipjawsV kit using older E-die is almost able to match its bigger brother and optimized HyperX will be better than much faster G.Skills on stock (XMP) settings.

Geekbench

In Geekbench results are similar – overclocked HyperX will catch stock G.Skills, but Samsung-based RAM will run away when optimized. Overall result is 200-300 points better than on stock.

AIDA

Starting with latency, we can see that the results are improving a lot. XMP settings on G.Skills gave latency measurements of around 43.5 ns and B-die based TridentZ gain 5 ns here, which is more than 10%. E-die also allows for improvement, but it’s substantially smaller, 2.5 ns and around 5%. Optimized HyperX again catch stock G.Skill.

The trend continues here as well, and we can see increases of 3-4 GB/s in memory bandwidth of G.Skill mems. This also shows how loose the default timings are, but we cannot expect that it will directly translate into 10% performance improvement. Unfortunately, operating frequency of HyperX mems is too low to let it spread its wings.

SuperPi 32M

SuperPi 32M shows some real performance improvement. B-die got down to 6:29 from 6:34, E-die – to 6:31 to 6:37 and MFR – to 6:33 from 6:40. Every second counts when fighting for optimal performance, but what we see here is that the test runs about 2% shorter.

This concludes first round of tests, where safe voltage of 1.4-1.45V is used. I was able to demonstrate how loose XMP timings are and what are possible gains in benchmarks. The real overclocking will start in next part, when I will raise voltage even more to 1.55V which in my opinion is just about maximum you could consider for 24/7 with proper cooling in your system.

Comparing DDR4 IC on Z170 platform – part 1 – intro and stock settings

Earlier this year I researched information for an article which will be published on overclock.pl anytime soon. The article is going to cover comparison of performance of DDR4 modules in synthetic benchmarks and their overclockability. For that purpose, I will use following benches in their default settings as per HWBOT rules:

• SuperPi 32M
• XTU – version 6.0.28
• Geekbench – version 3.4.1 64-bit
• AIDA – version 5.7.0

The test platform consists of air cooled setup of:

Component	Model	Settings
Motherboard	Asus Maximus VIII Hero
CPU	Intel Core i5 6600K	100 x 48 = 4800 MHz 1.28 vcore Cache multi: 45
Cooling	CPU: Scythe Mugen 3B + Arctic 120 mm RAM + MOBO: Enermax TwisterStorm
PSU	Corsair RM850
SSD	Samsung 850 EVO 128 GB

OS used is Windows 7 SP1 64-bit. All settings are constant apart from RAM. Presented results will be average of 5 runs.

Memory kits used in the test and DDR4 ICs on them:

• G.Skill TridentZ 3600C17 (2×8 GB) – Samsung B-die
• G.Skill RipjawsV 3200C16 (2×4 GB) – Samsung E-die
• HyperX Savage 2666C13 (2×4 GB) – Hynix MFR

Results on default settings

The results here are achieved using XMP profile in RAM SPD – XMP profile is loaded in BIOS, system is booted and off we go 🙂

XTU

XTU is most used benchmark for 2d ont HWBOT in last months, mostly due to fact that it’s bundled with Intel motherboards. It’s interesting to see that HyperX win here with much theoretically faster RipjawsV. This will lead into conclusion that G.Skill XMP profile contain quite loose timings to allow booting high speed on low volts on most motherboards.

Geekbench

HyperX lose a lot here due to lower working frequency, which is clearly visible in “Memory” tests. This confirms the hypothesis that frequency is very important for DDR4

AIDA

Due to the fact that different scales are being used data is displayed in 2 charts:

AIDA results confirm G.Skill advantage due to higher working frequency. Interesting thing is that latencies on both are similar, but B-die kit has higher operating frequency which translates into better bandwidth.

SuperPi 32M

SuperPi 32M results support previous observations – memory performance depends a lot on the frequency, and loose timings in G.Skill XMP profile make the chart a bit flat.

 

This concludes the test setup and baselining on stock settings. Next parts should be more interesting when we start overclocking and observe some patterns. I plan to divide it into several parts:

• safe OC for 24/7 – with 1.45 volts
• bench safe – with 1.55-1.7 volts range
• full out – with 1.75 volts and above
• summary to see scaling results

It was a lot of fun for me to prepare the results, but if you have some comments or questions, or just want to point my poor efficiency, feel free to drop me a note.