Video - RAM: Testing Dual-Channel vs. Single-Channel Performance 

The above is the video component, wherein we show off our After Effects RAM preview FPS when dual-channel is used, along with most the other tests. The RAM featured is Kingston's HyperX 10th Anniversary RAM, which is a special version of their normal HyperX memory. More on this below.

Addressing Terminology: There Is No Such Thing as "Dual-Channel RAM"

First of all, there's no such thing as "dual-channel memory." I want to get that cleared-up early. Memory channeling exists at the platform level, so a dual-channel chipset or IMC (Integrated Memory Controller, as in modern CPUs) may exist, but the memory itself does not have a special bit or chip that controls this. It is up to the motherboard and supporting platform to offer multiple channels.

Things get a little different once you enter quad-channels from a density perspective, but we won't get into that here. The short of it is that higher channel configurations (like with IB-E) give you the opportunity to go to higher densities, like 64 gigabits, which has a pretty large impact on performance potential. Again, that's out of the scope of our dual-channel / single-channel article.

What Are We Testing?

We're testing the performance of two sticks of 2x4GB (more on the platform specs & methodology below) in single- and dual-channel configurations. Please note that all tests were conducted with a discrete GPU and will not use the IGP present in IvyBridge.

Further, note that channel configurations could (not tested) and frequency will have a significant impact upon APUs and IGPs, so these test results are strictly targeted at systems that don't rely heavily upon an integrated graphics chip. This is because APUs and IGPs do not have on-card memory, as a video card does, and thus must access system memory for their graphics processing; in this instance, DDR3 RAM is (1) physically farther from the GPU component of the CPU and (2) significantly slower than the GDDR5 memory found on video cards. This in mind, you'd want every advantage you can get with an IGP.

Again, this test strictly looks at memory performance when the IGP is not utilized.

Dual-Channel Architecture: How Dual-Channel RAM Platforms Work

If you already understand the basic, top-level concepts of multichannel platforms as it pertains to memory, you can skip this part and jump into the methodology section immediately following. For those unsure of exactly what "dual-channel" or "multichannel" means, read on.

A single stick of RAM will operate on a single 64-bit data channel, meaning it can push data down a single pipe that is 64-bits in total width. The channel effectively runs between the memory controller or chipset and the memory socket; in the case of modern architectures, the memory controller is often integrated with the CPU, rather than acting as standalone board component.

By utilizing multi-channel platforms -- something available on every modern build -- we multiply the effective channel width by the count of channels available. "Effective" is key. In the case of dual-channel configurations, we've now got 2x64-bit channels available to the memory. This means we've doubled the data traces running in the memory bus, and now have an effective 128-bit channel, which in turn doubles maximum theoretical bandwidth. This is why I made it a point to say that dual-channel platforms are what exist, not memory -- in the case of dual-channel, the board will host 128 physical traces to handle data communication between the IMC and RAM. This is compared against 64 for single-channel platforms and 256 for quad-channel platforms.

In this image, we see our physical data traces (wires) running between the memory and the IMC. You'll notice that there's room for 128 in the below image, as opposed to the 64 above. D0-D63 represent the first channel, D64-D127 represent our second channel. Modules can process 64 bits of data at any given time, and so dual-channel platforms will read and write to two modules simultaneously (saturating the 128-bit wide bus).

The memory will sort of ping-pong data down the channels, effectively doubling its potential "speed." This should not be confused with double-datarate memory (DDR), which operates independent of the channel configuration.

In order to use RAM in a dual-channel configuration, the memory must be socketed into matching memory banks and should be identical in spec. Technically, some boards will allow different spec memory in dual-channel configurations, but you'll be throttled to the slowest module and may experience instability. If you're running four sticks with two different models of memory, just stick the matching modules into a bank (bank 0 contains brand A; bank 1 contains brand B).

Up until very recently, you'd have to use bank 0 (slots 1 & 3, assuming we start counting at '1') to properly use dual-channel with only two sticks present. Slots 2 & 4 (bank 1) would then be filled upon upgrade.

RAM Test Methodology Pt. 1 - Synthetic & Real-World Tests Needed 

Testing any component competently isn't a trivial feat. I used to do portables test engineering at a large computer manufacturer, where we devised some of the very first USB3.0 test cases and worked with Intel's first consumer-available SSDs (X25). This was somewhere around the Nehalem reign, so the architecture was relatively similar to what we deal with today. Those tests weren't easy to figure out, but with the understanding of how each product works and what software can stress its bottlenecks (without bottlenecking elsewhere - like the CPU, GPU, etc.), it's achievable.

My point is that RAM isn't too dissimilar from Flash testing in terms of test concerns and methodology. For these RAM tests, I specifically wanted to focus on the performance of multichannel configurations vs. normal operating frequency ("single-channel," we'll call it).

From our Kingston HQ test & assembly facility tour. Learn about this machine here.

Applying both synthetic and real-world tests is important; without synthetic tests, we can't adequately isolate memory performance and make extrapolations / predictions for real-world tests. Being able to predict outcomes in the real-world is a keystone to scientific test methodology, so you'll almost always see synthetic benchmarks in our testing. That said, without real-world tests, it's tough to put things into perspective for users. Synthetic tests take some heat on occasion for being "unrealistic," but at the end of the day, a correctly-applied, correctly-built synthetic benchmark is paramount to performance analysis.

RAM frequency and channeling will have the biggest theoretical impact upon, obviously, memory-intensive applications. In this environment, those applications tend to be render, encoding, transcoding, simulation, and computation-heavy tasks (applying a filter in After Effects, for instance). I want to make clear that we are strictly testing multichannel performance between dual-channel and single-channel platforms and will not be testing triple-, quad-, or alotta-channel (that's a technical term) performance in this benchmark; we will also not be testing memory frequency herein, and so it will remain a defined constant in the test. I also want to make clear that the memory capacity remained constant throughout the entire test, as did the sticks tested. I will discuss how this was achieved below.

These tests were conducted with consumers in mind, but I will comment on the impact for developers and simulations briefly -- at least, as far as my professional experience will confidently allow.

Multichannel Performance Hypothesis

Before getting into specific tools and use-case scenarios, let's explore my hypothesis going into the lab.

Up until that MSI meeting I mentioned, it was my firm belief that dual-channel configurations should always be opted for in any system build that supported it. This theoretically doubles your memory's transfer capabilities, after all, so halving potential seemed unnecessary.

After my preliminary tests that indicated dual-channel performance might not be quite as substantial as I'd always thought, my considerations changed. Going into this benchmark, it was my hypothesis that:

• Dual-channel configurations would exhibit no noteworthy difference in gaming use-case scenarios.
• Dual-channel configurations would exhibit no noteworthy difference in system boot and daily I/O use cases.
• Dual-channel configurations would show significant advantage in livestream previews for intensive render applications.
• Dual-channel configurations would show significant advantage in render, encoding, and transcoding applications (for video and incompressible audio).
• Dual-channel configurations would show significant advantage in compiling applications (data archiving).

I validated some of these beliefs, but 'proved myself wrong' on a couple of them. We'll revisit each throughout the article.

Test Concerns: Capacity, Memory Saturation, and Synthetic Benchmarks 

Going into the testing, I read several similar benchmarks that were performed by users across the web; we also did some collective team research on professional benchmarks performed in-house by memory manufacturers, who were very helpful in supplying test methodology revisions and their own results (shout to Kingston, Corsair).

Although capacity (1x4GB vs. 2x4GB) should theoretically not present an issue in most of our tests, it would present a very clear limitation with After Effects and Premiere RAM previews and encoding. I also wanted to eliminate the doubt in anyone's mind. To keep capacity a constant and mitigate concerns of test invalidation, the tests were run in these configurations:

• 2x4GB; bank 0 (Dual).
• 2x4GB; mismatched banks (Single).

We used Kingston's 10th Anniversary HyperX RAM at 2400MHz, which is basically just a special-colored version of their normal HyperX RAM. I will define the RAM modules below.

By using mismatched slots with our motherboard, we're able to force single-channel operation for the second test setup. This means we can still run 8GB of RAM, so capacity is constant, and it also means we're still testing multi-channel performance vs. single-channel without eliminating a full stick.

So that one was solved.

Other concerns arose with memory saturation. We don't have to saturate all 8GB of RAM to test the speeds, but it is beneficial to saturate as much of the memory as possible to ensure it's pushed to the point where it actually benefits from faster read/write performance. Our only test that nearly filled the entire capacity was the After Effects RAM preview. The other ones saturated memory to a point we deemed to be sufficient for testing (often around half capacity, or 3-4GB).

Finally, synthetic benchmarks can sometimes show 30%, even 80% differences between the different configurations... but real-world testing could be 3-5%. We have explained each synthetic test's real-world applications in great depth below, so hopefully that'll help you understand when the results matter most and if they're relevant to you.

Please continue to page 2 for the test methodology, test tools, and test procedures; I defined each synthetic and real-world benchmark very carefully here, which will help in understanding the results.