Page 1:DDR3 Performance: What Makes Memory Perform Better?
Page 2:DRAM Sets Tested
Page 3:Testing Platform And Benchmark Suite
Page 4:Testing Methodology
Page 5:AMD Setup
Page 6:AMD Test Results
Page 7:Intel Setup
Page 8:Intel Test Results
Page 9:Comparative Analysis
Page 10:Top Performers
What makes one memory platform better than the other? In this article we go through benchmark testing and analysis to give you an idea of what you can expect from different DRAM sets running on Intel and AMD so you can get the best DDR3 performance.
DDR3 Performance: What Makes Memory Perform Better?
STOP! I know what you're thinking. And no, this isn't another typical DRAM review, so don't go running off to load up on caffeine. Sure, it will share some things you see in straightforward reviews — the intro, a look at the sets of DRAM, a look at the hardware test bed(s) and a few benchmarks. But the road will be far from straight and narrow.
I spend a fair amount of time in the Tom's Hardware forums, and a number of common problems and questions often arise about DRAM. Often, people don't understanding why the DRAM runs as it does, why one gets worse performance on one platform than another or between similar sets of DRAM, how to set up DRAM to run at its full specifications, whether one kind of DRAM works with another, and so on. I also know the DRAM manufacturers get flooded with the same questions, so I approached them with the idea of a review, but one where I would examine the performance difference on an AMD vs. Intel basis, and answered some of the questions I frequently see in the forums.
A 32GB set of 2400 DRAM would be ideal for the testing, since it can easily be downclocked and possibly overclocked, and we could scale the DRAM down to 16GB, or even 8GB for comparative testing. I approached several DRAM manufacturers and received a healthy dose of samples in response from (in alphabetical order): ADATA, AMD Radeon Memory, Corsair, G.Skill, Kingston, Mushkin Enhanced and Team Group.
DRAM Sets Tested
I requested a 32GB 4x8GB sample of DDR3 2400 DRAM, and left it to the vendors to decide what set to provide. The results came back with two sets of 32GB, 2400 DRAM at CL10; five sets of 32GB 2400 DRAM at CL11; and two of the companies didn't have 32GB sets of 2400, but in place provided two of their 2x8GB 2400/11 sets of DRAM. I also included a 32GB set of my own 2400/10 DRAM (Trident X).
Testing Platform And Benchmark Suite
For the testing, I primarily used my existing motherboard and CPU, AMD- and Intel-based combos. Other than the motherboard and CPU, I used the same components in each. I used a Hyper 212 EVO CPU cooler, a G.Skill Phoenix Pro 256GB SSD, a SeaSonic 750 PSU, an Asus 290X Matrix GPU and running Win7 Ultimate.
The AMD CPU was the 8370 paired with the Asus Crosshair V Formula Z motherboard. On the Intel side, I used a 4770K in an Asus Z87 Maximus VI Hero.
Testing Software And Testing Methodology
For the testing, I used a variety of staple programs:
- Aida 64 Engineer Edition
- PassMark Performance Test
- Si-Sandra Support-Engineer 2015 (and special thanks to Si-Software for providing a pre-release copy of the 2015 version)
- Geekbench 3
I selected Aida to test for latency and for read/write memory tests, PassMark for its Memory Mark score as well as its Database score, Si-Sandra for Aggregate Memory Latency, GeekBench 3 for both single and multicored overall scores, and WinRAR to see how memory amounts and frequencies affect overall performance.
I ran each test five times, removing the high and low scores and averaging the remaining three scores.
I enabled XMP (eXtreme Memory Profile), which set all the DRAM correctly to their specification (2400 and manufacturer's specified timings) with the exception of the ADATA and AMD sets, where two sets (2x8GB) of DRAM were provided to fulfill the 32GB. This often happens when people mix sets of DRAM, and as here, even with identical sets of DRAM. Often, one can make minor voltage and/or timing adjustments to get the sets or sticks to play, but not always. This is why DRAM manufacturers offer such a wide variety of sets of DRAM in 1, 2, 3, 4, 6 and 8 stick sets; the DRAM in a given package is all tested to work together. DRAM today, especially in the higher frequencies and tighter timings, is made to such high tolerances that many factors can affect the ability of sticks to play nicely together.
The Asus Crosshair V Formula Z uses DOCP (DRAM Over Clock Profile) — which is basically a derivative or version of XMP tailored somewhat to the frequency of the DRAM, as opposed to XMP, which pulls and directly uses the actual timings found in the SPD as programmed in by the DRAM manufacturers. AMD also utilizes other overclock profiling found on other AMD motherboards, like AMP (AMD Memory Profiles) and EOCP (Easy Over Clocking Profiles). The one exception was to use AMP for the AMD Radeon Gaming DRAM which set up both sets of DRAM to spec settings of 11-12-12-31, 1.65. Under DOCP the DRAM was set to the same as the rest of the sets: 11-13-13-35, 1.65, indicating that with these sets, DOCP simply sets the DRAM to what it is programmed to believe the best timings are for 2400 DRAM. This occurred with both the CL10 sets and the CL11 sets, and all of the sets ran fine at these settings. However, as the specification timings are tighter on all sets, I set the timings manually to the DRAM manufacturers' specifications.
Before getting into the testing results, I had mentioned looking at DRAM set up problems on both platforms and there were indeed a few.
The AMD FX CPUs are rated by AMD to run up to 1866 DRAM at 1 DIMM per channel and up to 1600 with four DIMMs. This can be deceptive at both the upper and lower ends of their spectrum. I've seen and run a number of the upper-level CPUs with higher-frequency DRAM and seen many of the lower tier FXs that couldn't run the 1866 DRAM. This was one reason I chose 2400 DRAM as the test point. I fully expected to have to run the bulk of the testing at 2133, as I have seen very few FX CPUs that could handle 2400 DRAM, especially fully populated at 32GB. Pleased that I appear to have hit the lottery with my 8370, I believe that the Asus Crosshair V Formula Z may be another factor in the test system's exceptional stability. Higher priced than other AM3+ motherboards, I feel it's money well-spent.
Mixing Sets Of DRAM
Mixing sets of DRAM can be problematic, and this occurred with both the AMD Radeon DRAM and the ADATA sets. (Again, each sent two 2x8GB individual sets of DRAM, the first two 'sets' I tested with). Generally, the approaches taken to try and get mixed DRAM to play nicely can consist of timings and/or voltage adjustments. In the case of the AMD DRAM, it booted up fine and ran under the AMP settings until I tried to run Prime95, which promptly crashed. Raising the CPU/NB voltage (which runs the memory controller) by a slight + 0.06 increase stabilized the system. The ADATA sets didn't want to boot at all, but after a similar increase to the CPU/NB (+0.05), it booted and ran. Again, Prime95 was the nemesis, resulting in a crash. With this group of DIMMs, I raised the DRAM voltage to 1.7, and the CPU/NB another + 0.04, to achieve stability at 1.31. (Looking onward to the Intel build, I have little doubt that an actual four-stick packaged set of DRAM would have run without the additional adjustments.)
Without going into a bunch of boring details, I also tried mixing two sticks each from a variety of combinations of the DRAM available, with mixed results. I managed to get some to work together with voltage adjustments, two with timings adjustments, some with a combination of the two — and some combos just wouldn't work at all. In at least one case I tried with two Corsair sticks and two Team sticks, and they just wouldn't play at all — period! I then tried with the two remaining Corsair sticks and the same two Team sticks. That combination did play nicely. So it really goes to show that mixing DRAM is really a crapshoot; you never know if mixing sets/sticks of DRAM will play together.
Failure To Boot
Two of the sets wouldn't boot under DOCP at 2400 with all four sticks installed. I took another common tack with those sets and raised the CPU multiplier to 21.5 (raising the base frequency of the CPU from 4.0 to 4.3). Additionally, when this slight OC was applied to the ADATA DRAM, it allowed the DRAM voltage to be lowered back to 1.65, though the CPU/NB voltage still had to be maintained at a lowered level of 1.26.
Failure Under Stress
There were three cases of this in Prime95, and all three were solved with slight CPU voltage increases. So, it wasn't really caused by DRAM faults.
AMD Test Results
On the AMD side there were far fewer problems than I expected, and all of those problems were very likely to happen with various AMD CPUs/motherboards and DRAM.
WinRAR File Compression
In the testing, using WinRAR, I compressed a video file that was 6.97GB in size. The time differential was minor across the sets. As would be expected, two of the CL10 sets came out on top, and were three of the top four sets. I also did the same with the DRAM at 1600/9.
While charts are nice, I'm more of a numbers guy. So to summarize a bit in the chart above, the time took compressing the file ranged from a top score of 4:39 by the Team Xtreem sticks to the slowest of 4:58. With the DRAM running at 1600, the range was from a fast time of 5:11 to a slow time of 5:21. Moving to 16GB of DRAM at 2400, the best time was more than a minute slower, at 5:38, and the worst was 5:47. Then, at the 8GB level, still at 2400, the fastest was 5:47 and the slowest was 6:01.
As you can see, there is little difference between the sets, though the time it takes rises with the lower frequencies. This is one reason when people multitask and use memory-intensive applications like those in video, imaging, CAD, GIS type work, VMs, etc. can see performance increases with higher frequencies. You can also see where it takes longer with reduced amounts of DRAM. As mentioned, these differences don't look drastic, but as an example, the difference between 2400 and 1600 DRAM here runs into about a 30-second difference in roughly 5 minutes, which is about 6 minutes per hour in time savings or about 45 minutes in a workday. Doing DRAM-intensive work all day, you would come close to doubling that savings by going from 8GB to 32GB.
Geekbench 3.2.2 Pro
Geekbench, using the overall scoring for both single- and multi-core scoring:
Again, I got mixed results. I was really hoping for something definitive, but the numbers were all over. The following results are basically overall recaps I used to evaluate the individual sets of sticks, but I thought some people might find the results interesting. And again, I promised to keep the chart count down.
Latency And Bandwidth
I used AIDA64 Engineer Edition to test DRAM latency as well as read/write scores. Again, the results varied, and I promised to keep charts to a minimum. So, to recap the scores:
Si-Sandra was used to get an Aggregated Memory Score in MB/s
|11.9 to 23.8 MB/s||17.62 to 17.89 MB/s|
The PassMark Performance Test was used to provide scores for a Composite Memory score from all the tests, as well as what they call a Database Operations score:
Next, we'll look at the scores from the Intel side of the testing and compare. Then, we will head into what I found to be the most interesting.
On the Intel side, there were basically no problems. Again, the two ADATA 2x8GB sets required a slight DRAM voltage boost (which can be expected with any mixed sets), and here, I set the Radeon DRAM up manually, and it required both a slight DRAM voltage increase and a bit of memory controller voltage (VTT) added. The other remaining sets all fired right up to spec by simply enabling XMP. I ran the 4770K at 4GHz with turbo boost on.
While there were no major problems with setting up the DRAM, when you do have problems, the approaches you can use basically mirror those used with the AMD sets of DRAM. A big difference is that when you need to adjust the memory controller voltage, Intel seems to change the term with every new chipset and, at times, uses different terms among different motherboards in the same chipset. It can be called DDRVTT, VTTCPU, CPUVTT, VCCIO, VTT and other names, and is even sometimes grouped in with the VCCSA (System Agent Voltage).
Intel Test Results
WinRAR File Compression
In the testing, using WinRAR, I compressed a 6.97GB video file. The time differential was minor across the sets. As would be expected, two of the CL10 sets came out on top and were three of the top four sets. I also did the same with the DRAM at 1600/9.
Getting back to numbers: With file compression, the sets ranged from a top score of 3:06 (by the G.Skill Trident X sticks) to the slowest of 3:31. With the DRAM running at 1600, the times ranged from a fast time of 4:01 to a slow time of 4:32. Moving to 16GB of DRAM at 2400, the best time was 3:31, and the worst was 3:51. Then, at the 8GB level, still at 2400, the times ranged from 4:16 to 4:39.
Geekbench 3.2.2 Pro
Geekbench, using the overall scoring for both single- and multi-core scoring:
Latency And Bandwidth
I used AIDA64 Engineer Edition to test DRAM latency as well read/write scores.
Si-Sandra was used to get an Aggregated Memory Score in GB/s.
|31.3 to 31.9 MB/s||20.41 to 21.23 MB/s|
PassMark Performance Test was used to provide scores for a composite memory score from all the tests, as well as what the company calls a Database Operations score:
Anyone who has glanced at the numbers can pretty much do an "analysis" on their own. Yes, we all know the Intel 4770K is a stronger CPU, both in single-threaded as well as multi-threaded operations. And that, in and of itself, leads to higher scores in Geekbench and plays in with all the benchmarks. I believe the big telling difference, though, is in the memory controller within the CPUs. What we see with the AMD CPU is older technology that is more oriented to lower frequencies and less total DRAM. I, for one, hope that the next generation of CPUs will use the advances the company has made and implemented in its newer APUs, which seem to like higher-frequency DRAM and handle it well. Intel handles the DRAM at lower latency and provides far higher reads and writes.
I could go into a detailed comparative analysis of the different numbers, but that isn't really the point of this. Rather, the point is primarily to give you an idea of what you can expect from either platform as far as DRAM, and why — if you are using one platform or the other — there are large differences in the numbers you will see when running benchmarks, and why it takes longer to perform an equivalent task on an AMD rig than on an Intel one.
You also can get an idea of how higher-frequency DRAM can show an increase in productivity on either platform when jumping from, say, 1600 to 2400, or moving up from 8GB to 16GB to 32GB of DRAM. To that end, the gains shown above may appear small, so I also ran a few tests on the sets that aren't on any list of benchmarks available.
I'm not a big fan of most benchmarks, in large part because most of them don't really use the DRAM. In other words, there are lots of gaming benchmarks available, but most gaming is centered entirely on the CPU or the GPU, and DRAM isn't much more than a conduit for data to flow through. I've long said that having more higher-frequency DRAM tends to show the strength of it more when multitasking or using memory-intensive applications. Earlier, I used WinRAR as a benchmark to provide examples of changes between 2400 and 1600 DRAM, as well as in different amounts (8GB, 16GB and 32GB) at 2400. As I expected, the completion time took longer with the lower frequency of 1600 using 32GB, as well as when using smaller amounts of DRAM. Taking it a step further, though not something one can truly and fully quantify, I experimented with running multiple applications and then running WinRAR to compress the same file used earlier. The "simulation" consisted of opening 10 tabs in Chrome to a page that changes, running a Malware Bytes scan and running Geekbench along with WinRAR. I came up with these numbers (again, it's the average of the mid three of five tests) on the Intel system"
|Original WinRAR Score||Multitasking Score|
When multitasking, you can see even bigger gains with the additional DRAM.
All eight of these DRAM sets are nearly equal, and with the testing that was done across two platforms — and considering this was intended to be more of an informative article than a typical hardware review — here are some things to take into consideration.
Top AMD Platform Performers
The top performer of all the sets on the AMD platform was Team Xtreem. In the overall testing, Team Xtreem edged out the other two CL10 sets, G.Skill Trident X and Corsair Vengeance Pro (in order), but not by much. The G.Skill Snipers were the best of the CL11 sets. The Team Group DRAM set was an eye-opener; I’m looking forward to working with more of their sets of DRAM.
Top Intel Platform Performers
On the Intel platform, the best performing set was the G.Skill Tridents X, again only by a small margin, topping the Corsair Vengeance Pro and Team Xtreem sets (in order). In the CL11 sets, the Snipers were at the top, barely edging out the Kingston Savage and Mushkin Blackline set.
AData And AMD
Both manufacturers guarantee their DRAM by the packaged set, which means both ADATA and AMD have a lot of faith in their products to provide us with two sets of 2x8GB (since they don’t have 4x8GB sets). I hope they will come out with actual 32GB sets in the higher frequencies since their sets performed very well and would no doubt would have been even better if they were matched 4 stick sets.
A Note On Packaging
Team Xtreem and Mushkin Blackline had the best packaging; the clamshells in a box ensure the sticks don’t get damaged, and look great, too. The packaging of the Kingston Savage sticks is rather unique; it has a great tray concept and it’s not a package you’d likely throw away.
The AMD Radeon Gamer Series has implemented a new approach to overclocking profiles with the implementation of AMP. Time will tell if it catches on with other manufacturers. The Trident X has a unique approach to high profile DRAM with its removable upper fin to fit under tight heat sinks.
This was a tough one, since it’s all subjective. I loved the look of the Trident X sticks, but as I worked with these others on a daily basis, I was drawn to the looks of the Vengeance Pro sticks, especially on the ROG motherboards. The initial look in a case, and seeing the name written along the top edge of the sticks, is also appealing – which you’ll also see on AMD’s Gamer Series. The ADATA XPG, Savage, and Muchkin Blackline sticks are nice, clean-looking sticks, each with their own twist to the heat-sink design. The Team Xtreem sticks with the reflective label that appears to change colors might well be most original-looking, though I’d like to see the labeling on both facing sides. And then there’s the Sniper series, with the embossed rifle, looking to shoot the others is (in the computer world) a long time classic.
I spent many hours testing these sets and wouldn't hesitate to recommend any of them, given the right circumstances. I’ll continue to test them with other mobo/CPU combinations to find the best combo scenarios where we can suggest various DRAM sets to Tom’s Hardware readers and forum members.
As you can see, I've given you some ideas of why the same DRAM will provide different results on different platforms. I also wanted to present some info that I hope would come in handy if and when you encounter problems when setting up DRAM.
As I mentioned earlier, I'm hoping to follow up with another article or two about DRAM and explore common questions and misconceptions. For instance, "Can I use any DDR3 1600 DRAM in my system?" I would love to hear from you with your thoughts, questions you might like explored in the follow-up article and, as always, any comments, criticism or critique.