Test Results: Dual-Channel vs. Single-Channel RAM Platforms - Synthetic
Let's start with synthetic multi-channel platform tests and then move into real-world stuff. All of these tests were defined, explained, and detailed in their use-case scenarios and methodologies on the previous page. Please read that page before asking questions or making claims about the below tests. It is very well thought-out and probably addresses your question directly.
We saw a fairly substantial performance difference between single- and dual-channel configurations with Euler 3D. The dual-channel configuration performed approximately 17.7% better than the single-channel configuration. This advantage will carry over in simulation applications, especially when dealing with CFD or similarly heavy-duty computations, simulations, and compilers. Engineers and scientists will theoretically benefit here (see: Signal Integrity, parametric analysis, crosstalk analysis, electromagnetics simulation, etc.).
MaxxMem measures memory copy, memory read, memory write, memory latency, and memory bandwidth performance. The MemCP, MemRD, and MemWR stats are measured-out in MB/s (bytes, not bits); latency is measured in nanoseconds (ns); bandwidth is measured in GB/s, though I've converted it to MB/s for simplified charts.
We saw substantial performance differences between single- and dual-channel configurations in this testing. Memory copy, read, and write tests were heavily-advantaged with the dual-channel configuration, performing 30.79%, 26.94%, and 14.52% better in each, respectively. The total memory bandwidth also rested 21.3% higher than single-channel -- again, pretty substantial. Latency with dual-channel was consistently higher, perhaps due to additional overhead.
WinRAR File Compression
With our WinRAR archive benchmark, we saw single-channel take 460s (7m 40s) to compress the 9GB data archive, meanwhile the dual-channel platform performed ~2.87% faster at 447s (7m 27s). This difference could potentially be fairly substantial in specific real-world environments (think: enterprise) where heavy, constant file archival is being performed. Web & db servers are an excellent real-world example where you'd want that 2.87% speed boost.
Handbrake Video Transcoding
I defined our Handbrake settings in the previous page, so check that if you're curious about what's actually being tested. We saw a consistent 4.4% advantage in the favor of dual-channel platforms while transcoding files with Handbrake. The single-channel platform took 209s (3m 29s) vs. the dual-channel platform's 200s (3m 20s). Extrapolating this to a larger, more complex transcoding task, the impact could be relatively noteworthy. I don't think most users do that level of heavy-duty passes on their video ripping / transcoding, though.
Shogun 2 Benchmark Load Time & FPS
I really didn't expect a lot here, but that's sort of what I set out to illustrate. These tests are totally uninteresting. The load time was about 0.78% faster with dual-channel platforms, but this is well within margin of error. For point of clarity, the results averaged were:
Single-Channel: 51.82, 51.78, 51.98, 51.83.
Dual-Channel: 51.76, 51.45, 50.41, 52.17.
Like I said, easily within normal system fluctuations. I also tested Skyrim's load time with several high-fidelity mods loaded, which should have theoretically hammered RAM and I/O for file retrieval, but saw effectively zero advantage between dual- and single-channel performance.
The FPS results are even less interesting. A 0.21% delta between the channel performance is within margin of error, once again, and can be effectively thought of as 0 noticeable difference.
Cinebench OpenGL Performance
Cinebench, like Shogun, wasn't really added to show a delta as much as it was added to illustrate a point: That multi-channel memory platforms have very little impact on specific tasks, like gaming and some types of live rendering. The delta between single- and dual-channel configurations was 0.25% in favor of single-channel.
Adobe Premiere Encoding Pass
Here's where it gets a bit more interesting. Our single-channel encoding pass results averaged out to around 236s (3m 56s); dual-channel averaged out at 229s (3m 49s), for a delta of about 3.01% in favor of dual-channel memory configs. Not massive, but for people who dedicate tremendous time doing rendering, it could be a big difference. Still, that's only a 2-minute gain per hour of rendering downtime. I often have systems rendering for 20 hours per day during conventions, so that could be upwards of 40 minutes saved, which starts to get significant.
Adobe After Effects Live RAM Preview Framerate Performance
We saw nearly a 6% difference (5.94%) between RAM previews with single- and dual-channel RAM in Adobe After Effects, favoring dual-channel configurations. This starts getting be somewhat noticeable. The average FPS of live playback with the single-channel platform was 14.227; the average FPS of live playback with the dual-channel platform was 15.098. Larger differences might be spotted under some specific test conditions, but it ultimately depends on what you're doing in AE. It certainly wasn't real-time (60FPS), but a single frame per second can feel like a big difference when you're staring at this stuff for days on end.
The Verdict: What Do These Results Mean; Is Dual-Channel "Worth It?"
That was a helluvalot of testing and methodology discussion, but I hope it was for a good cause - I wanted to ensure there were no questions about how we performed these tests. More importantly, I wanted to ensure that others can replicate and add to my results. If you're running this yourself for some specific game or program that I didn't test, please feel free to list your specs and software tested below!
Despite all that I thought I knew leading up to our MSI meeting last July, dual-channel just isn't necessary for the vast majority of the consumer market. Anyone doing serious simulation (CFD, parametric analysis) will heavily benefit from dual-channel configurations (~17.7% advantage). Users who push a lot of copy tasks through memory will also theoretically see benefits, depending on what software is controlling the tasking. Video editors and professionals will see noteworthy advantages in stream (RAM) previews and will see marginal advantages in render time. It is probably worth having in this instance -- in the very least, I'd always go dual-channel for editing / encoding if only for future advancements.
Gamers, mainstream users, and office users shouldn't care. Actually, at the end of the day, the same rule applies to everyone, simulation pro or not: It's density and frequency that matters, not channeling. Quad- and better channels theoretically have a more profound impact, but this is in-step with the increased density of kits that are targeted for quad-channel platforms. If you want to push speed, density and frequency should be at the top of your list. Generally, when you're spending that kind of money, you're going with a multi-channel kit of two or more anyway, but the point still stands.
I'd love to test the real-world impact of dual-vs.-single-channel memory configs on a server platform, but that starts exiting my realm of expertise and would require extensive research to feel confident in. If any of you are knowledgeable in the virtualization or server spaces, please let us know below if you think we'd see a bigger impact in those worlds.
As for whether it's "worth it" to get a kit of two, the answer is generally going to be yes -- but primarily because it's rare not to find a good deal with two sticks. If you're on a budget or an ultra-budget and are trying to spare every $5 or $10 you can, then perhaps grab a single stick of RAM. It feels so wrong saying that, but we have to trust the results of this test, and the results say that it simply doesn't matter for those types of users. Anyone building a ~$500 or cheaper system shouldn't spend the time of day being concerned about 2x4GB vs. 1x8GB as long as the price works out in their favor. Price is the biggest factor here, and with recent fluctuations, you're just going to have to check the market when you're buying.
This was a big undertaking. We normally don't go through such depth to detail test methodology. Please provide some feedback below if you'd like to see similar depth in the future. If you think it's too much and would rather we get to the results faster, let us know about that, too.
- Steve "Lelldorianx" Burke.