DDR5 vs. DDR4: How Much Performance Will You Gain From Today’s Newest RAM? – PCMag
With Intel’s “Alder Lake” 12th Generation desktop processors, we have the debut of a new memory standard in consumer PCs: DDR5, which is used by many of the motherboards that support the new chips. One thing we know about DDR5 is that it’s “faster” than DDR4 by the most commonly marketed numbers, but those numbers are just a measure of theoretical peak data rate. (For a primer on the essentials of DDR5, see What Is DDR5? Everything You Need to Know About the Latest PC Memory Standard.)
Real-world applications, however, often use a rapid succession of small transfers to and from memory, where the time it takes to open and close memory cells is at least as important as how quickly the data is transferred once those cells are open. In measuring memory speed, the two parameters we’re dealing with are bandwidth and latency, and early DDR5, so far, isn’t exactly blessed in the latency department.
Why is that? For starters, hardly any high-end (overclocked), performance-grade DDR5 has come available yet. (Indeed, DDR5 of any kind, period, was scarce at this writing in late November 2021. Scalpers were even getting in on the proceedings.) And while we expect industry-standard DDR5 to eventually reach data rates as high as DDR5-6400, what’s currently available is DDR5-4800, with a sluggish-sounding CAS latency rating of 40. (Higher-data-rate DDR5 RAM has been announced, but good luck finding any at this writing!) With DDR4 CAS latencies below 20 in high-performing DDR4 memory, that “CAS 40” might sound concerning.
Lower numbers, all things being equal, are better when you’re talking about latency figures for memory. Of course, the latency situation for DDR5 isn’t as dire as that big number appears at first blush: Latency is specified in clock cycles, so that anything that cycles twice as fast requires half as much time per clock cycle, as outlined in our overview of how to buy the right RAM. But that just means that DDR5-4800 CAS 40 can only be as quick as DDR4-2400 CAS 20, which would in turn still be laggier than the slowest DDR4-2400 at CAS 18. You’d think that JEDEC (the industry group(Opens in a new window) that standardizes DRAM) would have at least waited for its memory makers to get DDR5-4800 down to CAS 36 just to stay on track with its previous efforts. But there’s more to it than that!
How We Tested: DDR5 vs. DDR4
Each 64-bit rank of DDR5 memory (a rank being a subset of memory chips on a memory module) is divided into two 32-bit ranks, with the latency benefits described in that same overview article being applied in a slightly different way. (Again, check out that piece for a much more detailed background discussion of memory ranks.) With this many factors in play, we really didn’t know what to expect when we tossed several modules into our test systems for a modern-DDR throwdown.
DDR5 below, DDR4 above: All the RAM we used in our tests.
(Photo: Thomas Soderstrom)
For our DDR5-versus-DDR4 faceoff, we configured two testbed PCs in sequence using the Intel Alder Lake Core i9-12900KF (the graphics-free version of the Core i9-12900K flagship Alder Lake CPU) on two different Z690 motherboards (one for DDR5 testing, one for DDR4 testing). The rest of the testbed components, apart from the motherboards and memory modules, were used in common. Here is the component pick list we used…
All of the DDR4 and DDR5 memory kits tested were 32GB kits, comprising two 16GB modules. Crucial supplied us with a kit of the industry-standard DDR5-4800 for today’s technology evaluation, and Mushkin stepped in with several of its XMP-enabled (overclockable) DDR4 kits to allow for testing across multiple data rates. While Crucial’s DDR5(Opens in a new window) operates with 40-39-39-77 latency values, Mushkin’s(Opens in a new window) timings look better as frequencies decrease from DDR4-4400 (19-26-26-46) to DDR4-4000 (18-22-22-42) and DDR4-3600 (16-19-19-39) in its three Redline kits. All modules are organized as single (64-bit) rank, despite the way each DDR5 rank is divided.
Our test modules, along with the DDR5-capable Asrock Z690 Taichi
(Photo: Thomas Soderstrom)
The Z690 motherboard platform, the only one supporting Alder Lake so far, encompasses support for both DDR4 and DDR5 memory. Unfortunately, no motherboard manufacturer (so far, anyway) has been willing to put both DDR4 and DDR5 slots on the same motherboard. This might have something to do with the fact that these different technologies can’t be used simultaneously, but we’ve seen several brands take this risk in the past with earlier memory generations (with DDR2/DDR3 combo boards, for example).
Motherboard sandwich: Asrock’s Z690 Extreme WiFi 6E (DDR4) atop its Z690 Taichi (DDR5)
(Photo: Thomas Soderstrom)
Lacking that option, we had to experiment with two near-aligned boards. Asrock sent two of its most closely matched DDR4 and DDR5 boards: the Z690 Extreme WiFi 6E(Opens in a new window) (for the DDR4 testing), and the Z690 Taichi(Opens in a new window) (for DDR5).
Testing DDR5 vs. DDR4: Down and Dirty With the Performance Differences
Since the JEDEC DDR4-3200 standard is a BIOS option for the DDR4-supporting Z690 Extreme WiFi 6E, we selected it to form our performance baseline in the tests that follow. Apart from that, a custom-configured DDR4-3200 14-16-16-30 setting fills the gap between it and Mushkin’s DDR4-3600 CAS 16 XMP setting.
Bandwidth and Latency Tests
For starters, here is a summary of the memory-speed and synthetic application tests we ran…
And onward to the first set of results…
Remember that more bandwidth is better, and more latency is worse. The DDR5 kit falls as far behind in latency as it gets ahead in bandwidth. That interplay of results was expected, but we’re still hoping for a few surprises in actual applications. So, onward to those…
Application and Gaming Tests
The synthetic application tests that we listed above played out thus…
There isn’t much to see in the 3DMark Time Spy test or various PCMark 10 testing. The only outliers are the DDR5 memory topping the DDR4 field in the 3DMark Time Spy CPU test, and falling a fair bit behind in PCMark 10’s Photo Editing simulation. The differences among the rest of the results were within the kind of margin-of-error ranges you see in variances from test run to test run with the same hardware.
So much for synthetic tests! Let’s see how this shakes out in real-world applications. Here’s a rundown of those tests we ran, and the settings, where appropriate….
…and the results are summarized below. In the chart below we clustered several of the timed tests into a single “Timed Tests” tab to reduce complexity.
The UL Procyon benchmark, which runs a consistent, repeatable workload on a licensed install of Adobe Premiere Pro, shows that more bandwidth can help matters, with DDR5-4800 taking the win there. The results of this test still indicate some gains for reduced latency when moving from CAS 22 to CAS 14 at DDR4-3200, however.
Next up: F1 2021, the AAA game we ran. The results there illustrate that the game’s performance becomes increasingly memory-bandwidth-dependent as background images become more complex. (Choosing the game’s Belgium track, with a wet surface, brought out the differentiation we desired.) The DDR5 topped the field by a smidge at 1,920 by 1,080 and the Medium setting. Better memory latency still had some impact, however, as evidenced by the standard DDR5-4800 falling slightly behind the DDR4-4000 XMP at other tested resolutions and settings. But it was super-close.
Recommended by Our Editors
In contrast, the 7-Zip Beta compression/decompression benchmark shows tremendous gains in file-compression performance with the DDR5 kit. (See the 7-Zip results on their own tab.) So, in addition to the built-in 7-Zip benchmark, we added an actual 7.6GB file-compression time (a second 7-Zip result) to our timed benchmarks.
As for the other tests: Nothing meaningful in Cinebench R23, while DDR5 wins at the HandBrake bench. But most interesting of these results is 7-Zip, with that huge file compressing more than 20% faster on DDR5 than even the DDR4-4000 result. And, of course, it wipes the floor with DDR4-3200. Further fun with the data includes the Corona benchmark executing its highest rays-per-second with DDR4-4000.
Power Consumption Testing and Performance Summary
The Asrock motherboards we used for our testing may be similar, but they’re not identical, and the DDR5 Z690 Taichi board we received is slightly better-featured than the DDR4 board. The fact that the power difference narrows when the memory is loaded indicates that much of that 8-watt difference between DDR5-4800 and DDR4-3200 lies within the board’s features. The easiest way we thought of to create a better estimate was to subtract the difference between idle and full load power, leaving 293 watts for the DDR5 and 296 watts for DDR4, presenting a best-case power-savings estimate of 3 watts for DDR5. Not a big difference in the grand scheme of a system, but we’ll take it.
The second tab in the chart above aggregates selected benchmarks that represent real-world applications, such as the App Startup subtest from PCMark 10, the Premiere Pro result from UL Procyon, all the apps from our timed workloads chart, and F1 2021. The combined figures are an average of those results. Thanks to its huge lead in 7-Zip file compression, the DDR5-4800 wins over DDR4-4000 by 1.5% overall. We also experimented with removing the timed benchmarks from this data set (not charted here), and in that scenario, the DDR4-4000 kit would have led by 1.3% without that additional data.
What that tells us: The answer to the question “Which is better, DDR5 or DDR4?” really comes down to which tests you include (or don’t include). But whatever your selector set is, the results remain close, at least with today’s modules.
Verdict: It’s No Blowout (But DDR5 Picks Up Where DDR4 Stops)
Putting aside the one application we found where DDR5 truly excels (7-Zip…after all, who compresses files all day?), DDR5-4800 trades blows with DDR4-4000 quite well. And that’s pitting non-overclocked DDR5 against a DDR4 kit’s overclocked profile. (Remember that JEDEC standardization ended at DDR4-3200.) Moreover, DDR5 starts at the same density that DDR4 ends, at 16Gb (sixteen gigabits) per IC (16GB per eight-chip rank). Even in pricing, DDR5-4800 is eventually expected to end up going for the same 60% price premium as DDR4-4000, when compared to standard DDR4-3200.
(Photo: Thomas Soderstrom)
That’s out in the future, though. Those buying today will find DDR5 in short supply, with markups of roughly 80% on top of the previous “add 60% for DDR5” estimates, where you can find kits in stock at all. (We’re talking retail prices, too; don’t even look at the speculative pricing on eBay.)
That’s a tough pill to swallow for the minimal gains we see so far, even when you factor in the extended upgrade life that DDR5 will offer as that new PC build ages. (DDR5 is the future, after all, for subsequent Intel platforms.) For most of us considering a new Intel-based desktop-PC build today, the decision will be whether to build a new system now, using some DDR4 modules already in hand, or to wait for DDR5 pricing to normalize. For the moment, the price premium is too big to justify going all in on DDR5 just for DDR5’s sake, unless you’re flush with cash. Fortunately, a fair number of Z690 boards have DDR4 support; see our guide to the initial wave of boards for some shopping candidates.
Get Our Best Stories!
Sign up for What’s New Now to get our top stories delivered to your inbox every morning.