Memory Frequency Scaling in SFF Systems: An Investigation with SO-DIMMs and Coffee Lake
by Ganesh T S on November 28, 2018 8:00 AM ESTOverclocking has generally been the domain of enthusiasts with desktop rigs. Though more recently we have seen even SFF PCs joining the bandwagon – Intel's Hades Canyon NUC, for example, supports overclocking the CPU as well as the GPU. However, increasing the CPU frequency beyond the official specifications is not the only way to extract more performance from a computing system. Memory-bound workloads can benefit from memory hierarchies with increased bandwidth and/or lower latencies.
We last looked at DDR4 memory scaling effects on SFF PCs when we experimented with different SO-DIMMs in the Skull Canyon NUC (NUC6i7KYK) based on the Skylake platform. Current SFF PCs are based on Coffee Lake, which brings in more cores while keeping power efficiency in mind. Compared to the Skylake memory controller's official limit of 2133 MT/s, the Coffee Lake memory controller ships with DDR4-2666 supported out of the box. In this article, we explore the effects of varying DDR4 SO-DIMM frequencies and timings on a SFF PC with a standard Coffee Lake desktop CPU.
Introduction
Since the introduction of DDR4 support in the Skylake platform, we have seen expanded support for overclocked memory kits on both the desktop and notebook segments. On standard non-overclocked systems, the DDR4 memory controller in Coffee Lake desktop CPUs operates at 2666 MT/s, while the U-series CPUs have a 2400 MT/s interface. DDR4 DIMMs operating as high as 4266 MT/s are available for desktop systems with full-sized memory slots. On the SO-DIMM side, we have seen various vendors introduce kits operating between 2133 MT/s and 3200 MT/s. While 2400 MT/s has become the de-facto SO-DIMM frequency for current systems, the usage of desktop CPUs in SFF PCs such as the ASRock DeskMini have ensured that the higher frequency SO-DIMMs also have adoption.
At the time of our Skull Canyon NUC memory scaling experiments, the number of SFF PCs supporting overclocked SO-DIMMs was very few. Since then, we have had numerous SFF systems utilizing desktop CPUs go through our test bench – the Zotac ZBOX MAGNUS EN1080K and the ASRock DeskMini GTX bring prime examples. For today's article we chose the latter for our experiments; the BIOS includes XMP support, making it very easy to get the SO-DIMM kits running at their higher-rated frequencies.
With memory prices still running quite high, we shifted from the 2x16GB configurations used in our Skull Canyon memory scaling piece to 2x8GB configuration for the Coffee Lake experiments. To this end, we procured some of Team Group's T-Force Vulcan SO-DIMMs, which are available in both 2400 MT/s and 2666 MT/s speed grades. Meanwhile to look at higher memory speeds, we also picked up one of Kingston's HyperX Impact 3200 kits.
The rest of this review deals with the quantitative measurement of the effectiveness of different types of DRAM in the ASRock DeskMini GTX (Z370). In order to do this, we processed various benchmarks while keeping everything other than the DRAM SO-DIMMs constant. Each configuration was booted to BIOS multiple times to ensure that the SPD information was properly parsed and the optimal frequency / timing parameters chosen. Once the OS was booted, we also checked with multiple hardware monitoring tools that the parameters indicated by the BIOS for the DRAM SO-DIMMs were indeed what the OS was also seeing.
ASRock DeskMini Z370 GTX1060 Specifications | |
Processor | Intel Core i7-8700 Coffee Lake-S, 6C/12T, 3.2 GHz (Turbo 4.6 GHz), 14nm++, 12 MB L2, 65W TDP |
Memory - Option 1 | Team Group T-Force Vulcan TLRD416G2400HC15BDC-S01 DDR4 SO-DIMM 15-17-17-35 @ 2400 MT/s 2x8 GB |
Memory - Option 2 | Team Group T-Force Vulcan TLRD416G2666HC18F-SBK DDR4 SO-DIMM 18-18-18-38 @ 2666 MT/s 2x8 GB |
Memory - Option 3 | Kingston HyperX KHX3200C20S4 DDR4 SO-DIMM 17-19-19-39 @ 2933 MT/s 2x8 GB |
Memory - Option 4 | Kingston HyperX KHX3200C20S4 DDR4 SO-DIMM 20-22-22-42 @ 3066 MT/s 2x8 GB |
Graphics | NVIDIA GeForce GTX 1060 (6GB GDDR5) |
Disk Drive(s) | Team Group Cardea Zero TM8FP2240G0C111 (240 GB; M.2 Type 2280 PCIe 3.0 x4 NVMe; Toshiba 15nm; MLC) |
Networking | Intel Dual Band Wireless-AC 3165 (1x1 802.11ac - 433 Mbps) 1x Intel I219V Gigabit LAN |
Audio | 3.5mm Headphone Jack Capable of 5.1/7.1 digital output with HD audio bitstreaming (HDMI) |
Miscellaneous I/O Ports | 2x USB 2.0 5x USB 3.0 (Type-A) 1x USB 3.0 (Type-C) |
Operating System | Retail unit is barebones, but we installed Windows 10 Pro x64 |
Pricing (As configured) | $1100 - $1138 |
Full Specifications | ASRock DeskMini Z370 GTX1060 Specifications |
In the next section, we will first take a look at the specifications of the three SO-DIMM kits / four configurations that were evaluated in the ASRock DeskMini GTX Z370, along with the AIDA64 Memory Bench for each. Following this, we present the results from our updated test suite for SFF PCs and workstations - SYSmark 2018, Futuremark / UL benchmarks, SPECworkstation 3 etc. Prior to our concluding remarks, we take a look at a few miscellaneous aspects such as pricing.
25 Comments
View All Comments
cygnus1 - Wednesday, November 28, 2018 - link
How does anyone look at those memory benchmarks and justify buying anything other than the cheapest RAM that meets minimum spec?Yuriman - Wednesday, November 28, 2018 - link
Pretty much agree. Good to know, though.nwrigley - Wednesday, November 28, 2018 - link
Yep. The only difference for me is that I only buy Crucial. This comes from personal experience of AMAZING customer support from them.I had one of their sticks die on me once after 8 years of use. I called in and was shocked to talk to someone in the US. Since they didn't make the same RAM I had anymore, they offered to replace all 4 sticks so that I had a matching set, even though 3 of the 4 sticks were fine. And since the replacements were slower timings without heat spreaders, they offered to double the capacity to cover the difference. They upgraded me from 4x 1-gig sticks of DDR2 to 4x 2-gigs of DDR2 without me raising any fuss - this was all customer services idea. That made me a customer for life.
cygnus1 - Wednesday, November 28, 2018 - link
Yeah, I too am a big Crucial fan because of reasons like this, from my experience as well. Not quite as generous as your story, but never any kind of trouble getting support for their hardware.But these benchmarks really show that performance should not be even remotely near the top of the list of reasons to pick one RAM part over another. Brand/warranty/customer service is a real way to differentiate and justify a given price.
koaschten - Wednesday, November 28, 2018 - link
I found this handy graphic on reddit some time ago:https://i.imgur.com/lbPIkiW.png
Looking at the tested offerings, it is obvious why there was so little performance gain, the Latency/clock relations are just off the chart, for the 3066 CL20 literally.
koaschten - Wednesday, November 28, 2018 - link
source: https://www.reddit.com/r/intel/comments/9mlwbn/ram...yes, this is DIMM not SO-DIMM, but shows the differences nicely.
willis936 - Wednesday, November 28, 2018 - link
It is somewhat frustrating to see all of this work done on a case that doesn't make sense to examine first.If the original question is "When does memory performance matter to CPUs?" then the place to start is at the extreme, not somewhere in the middle. If it was found that an 8 core 4 GHz x86 processor with whatever cache architecture and two channels of memory was memory bandwidth or latency starved THEN it would make sense to start moving down the stack and identify when it is no longer a concern. The conclusion to draw from this is much less meaningful to most any reader. There are like five people on the planet choosing between more expensive and cheaper memory kits for SFF systems.
GreenReaper - Thursday, November 29, 2018 - link
Might make more sense with AMD APUs. You'd probably get a much better return from overclocking the memory than the CPU, given how bandwidth-starved they can be.The_Assimilator - Wednesday, November 28, 2018 - link
Whatever happened to ranking memory by its performance rating, to determine how objectively good it is? For anyone who doesn't know/remember, performance rating = (memory frequency / CAS latency), and higher = better.It's sad that in this day and age, my 2x8GB DDR3-1600 CL8 (with no RGB LEDs or unnecessary heatsinks) has a higher PR than any of these DDR4 kits. It's even sadder that today's reviews of memory that "overclock" it, just concentrate on pushing up the frequency instead of trying to tighten the CAS timings, because the latter is where you'll see the most benefit.
nevcairiel - Wednesday, November 28, 2018 - link
All you are calculating here is the actual latency, since CAS latency is expressed in cycles. What this doesn't account for is the actual memory speed (ie. bandwidth).Just using your formula, a 1600/8 and 3200/16 module should be equal, right? But one of those offers twice the raw memory throughput, at about similar absolute latency (ie. performance rating).
It is a good idea to keep in mind that latency and frequency interact, but not in a way you suggest. Many people look at things like 3200 CL16 and 3600 CL18 and would instinctively say that the second set has a higher latency, while in reality the actual latency is quite similar, and you get more bandwidth.