Testing 4000MHz RAM: content creation

In traditional fashion, let's take a look at some quick content creation benchmarks before we get into the gaming tests.

Cinebench R20 is an industry-standard test of CPU power, with both single-threaded and multi-threaded workloads, and mimics the rendering of a 3D scene in Cinema 4D. We also tested video transcoding, a common task for any video producer, using the excellent open source Handbrake tool. Our test involved transcoding one of our Patreon video files to x264 and x265 (HEVC) using the Production Standard preset and CRF 18 quality setting.

First, let's take a look at our results testing the 4000MHz kit at its standard CL19 timings - plus results from when we increase the voltage from 1.35V to 1.40V for a quick-and-dirty overclock to 4200MHz. You can see that the single-threaded results show some variance between runs, but no clear increase from the higher frequency, but the multi-threaded results do trend slowly upwards as frequency ramps up, with that increase more-or-less stopping after 3600MHz. Given that AMD previously identified 3600MHz as the point at which diminishing returns start to kick in, perhaps that's not too surprising.

The Handbrake tests show similar results, with very little variance each time we tick the frequency 200MHz higher, with 4200MHz providing just a two per cent uplift in the HEVC encode frame-rate compared to 3200MHz. The H.264 encode is similarly uneventful, with run-to-run variance basically drowning out any performance gains. I think it's safe to say that, at least on our 9900K-based test rig, content creators won't see any appreciable increases to performance in these sorts of tasks from the use of higher RAM speeds.

We also recorded power usage at the wall from these tests, which seemed to flip between ~195W and ~210W whenever we increased frequency by 200MHz - weird!

Now, let's see what happens when we throw tighter timings into the ring. The excellent DRAM Calculator for Ryzen can also be used to suggest timings on Intel-based systems, but sadly doesn't seem to support the 4000MHz speeds we're using. We selected the highest frequency it did support, 3600MHz, and provided the rest of the data it required. It suggested primary timings of 16-17-17-34, compared to our stock 19-23-23-45, and we dutifully entered that into the BIOS, leaving the secondary and tertiary timings at their ASUS-optimised defaults for now. We'll only be running these timings for a short period, so we pushed voltage up to 1.4V and recorded our results.

Tightening our timings pushed up our Cinebench results against the CL19 counterparts from 3200MHz to 3600MHz, but then didn't provide much of a boost beyond that. In Handbrake, we recorded a new high scores for both h.264 and h.265 tests at 4000MHz CL16, but the overall increases were only around one per cent.

Given what we've seen so far, we don't expect to see massive changes from one RAM manufacturer to another, but let's look at the same tests performed on the same systems with slower XMP speeds, from 3200MHz to 3600MHz, all at CL16 with "XMP I" set in the BIOS.

The RAM sticks we used for this test are all 2x8GB kits:

• HyperX Fury 3200MHz CL16 (purchased for this test)
• G.Skill Sniper X 3400MHz CL16 (our usual RAM for GPU testing)
• G.Skill Trident Z Royal 3600MHz CL16 (our usual RAM for CPU testing)

Perhaps unsurprisingly, we see very little difference between our 4000MHz RAM at 3600MHz CL16 and different RAM with a 3600MHz CL16 XMP setting. That's good, as it suggests our results here from our Corsair 4000MHz kit will be more broadly applicable.

Before we get into our game testing, let's have a quick look at how all of these kits perform in their various configurations in a standard synthetic benchmark for RAM bandwidth testing: AIDA64's memory tests. These tests include four RAM-specific results we're interested in - read, write and copy times - plus a measure of latency. This should give us an idea of how the different configurations differ in raw performance, showing us how much of a performance difference we could expect in cases where RAM, not CPU or GPU, is the limiting factor.

The results here are pretty straightforward, with read, write and copy speeds increasing by around 2000MB/s to 3000MB/s for each additional 200MHz of frequency. Stepping from CL19 to CL16 seems to provide another 3000MB/s in read speeds, but has a smaller effect (~1000MB/s) on write speeds. Latency is unsurprisingly mostly affected by timings, with figures in the high forties or low fifties at CL19 and the lowest to mid forties at CL16. Note that there was more run-to-run variance here, which may explain the very low latency at 4200MHz CL19 compared to the other CL19 results.

From 3200MHz CL19 to 4200MHz CL19, we see a 24 per cent increase in read speeds, a 31 per cent increase in write speeds and a 27 per cent increase in copy speeds. If we instead compare 3200MHz CL19 to 4000MHz C16, we get similar figures, between 25 and 30 per cent. These should be close to the theoretical maximum upticks we could expect in any workload from switching from 3200MHz to 4000MHz RAM, with actual performance improvements being dependent on a range of other factors and therefore likely to be much lower, as we already saw with the minute changes to our content creation results.

With that out of the way, let's move onto what we really care about - gaming - where we expect to see more noticeable performance improvements when swapping standard 3200MHz for higher-spec RAM.

Testing 4000MHz RAM: Are higher frequencies worth it?

• Introduction, hardware breakdown, test system
• Content creation benchmarks: Cinebench, Handbrake, AIDA64 [This Page]
• Gaming benchmarks: Ashes, Far Cry 5, Crysis 3
• Testing 4000MHz RAM: the Digital Foundry verdict