Conclusion

The transition between Intel's 4th Gen Core "Haswell" and 6th Gen "Skylake" (DDR3 to DDR4) wasn't nearly as clean as the one we're seeing now. This is because DRAM prices were low back in 2016, and despite Skylake supporting both memory types, most motherboards shipped with only the newer DDR4 slots, forcing people to upgrade memory if they wanted a decent motherboard. 2021 is a whole different beast with nearly every component much pricier for the reasons we're talking about every day. Giving Alder Lake not just backwards compatibility with DDR4, but also making sure motherboard vendors will launch a pretty decent lineup of overclockable Z690 chipset-based boards with DDR4 slots was a much needed move by Intel. Our launch-day reviews confirm that that Alder Lake is a worthy upgrade from platforms several years old, which means you're probably still holding on to a decent DDR4 kit. That's why we wrote this review, to answer one of the most pressing question these days: Is choosing DDR5 a worthy performance upgrade? Even with pricing taken into account? After an exhaustive amount of testing with over 300 individual DDR4 test scores, we have some very interesting results.

There is no doubt that choosing DDR5 gives you a faster overall experience than DDR4, but sticking to DDR4 isn't that big a performance loss, and more importantly, DDR4 does not significantly change the overall performance outlook of the Core i9-12900K against its competitors from the previous generation or the AMD camp.

DDR5 comes with a massive uplift in data rates. Our reference point for DDR5 is data from our main review, which uses 32 GB (2x 16 GB) of DDR5-6000—a DDR5 speed that's considered "high-end." This was compared to four kinds of DDR4 configurations—32 GB of DDR4-3200 and DDR4-3600, and 16 GB at DDR4-4000 and DDR4-4400. The reason we split the memory configurations is that our 2x 16 GB Hynix dual-rank memory kit couldn't handle 4000 MHz or higher on the Alder Lake platform, so we had to switch to good old single-rank Samsung B-Die, which could handle the increased speeds. We also had to enable Gear 2 for the latter runs because similar to Rocket Lake, the integrated memory controller of Alder Lake can't run higher than DDR4-3600 at Gear 1, at least not on my CPU sample. The silicon lottery does play a role here, as does manual adjustment of the VCCSA voltage (also known as "System Agent").

Averaged across all CPU tests, the unlikely hero of this review has to be the humble DDR4-3200, which ended up just 3.7% slower than the DDR5-6000 reference. Just considering "6000" vs. "3200," I have to admit I was surprised to see such a small difference. In synthetics and rendering tests that aren't too memory bandwidth-sensitive and benefit more from lower timings, DDR4-3200 closely trails the DDR5 reference, beating all other DDR4 configurations. The DDR4-4400 and DDR4-4000 speeds are beaten by as much as 2%, which isn't much numerically, but still a result worth explaining. There are several things happening that put DDR4-4400 at a disadvantage. First, DDR4-3200 has significantly better timings of 14-18-18-36 1T, compared to 19-26-26-46 1T, and second, it is utilizing Gear 1 mode, which runs the memory controller at twice the command rate of Gear 2. We also had to switch memory kits from 2x 16 GB to 2x 8 GB, which besides the obvious change of total system memory capacity means we moved from dual-rank to single-rank memory. While 32 GB vs. 16 GB shouldn't have any significant effect on our benchmarks, dual-rank does offer slightly higher performance than single-rank.

Taking a closer look at individual results, we can even identify several application workloads that run better on DDR4 than DDR5: V-Ray Rendering, Unreal Engine 4 Bake Lighting, Microsoft Word, Text Recognition OCR, Virtualization, MySQL Database, 7-Zip Decompression, and MP3 Encoding. That's eight out of 37 tests, or roughly 20%. That having been said, in many other tests, things are too close to call, and I doubt any human would be able to notice the difference between DDR4 and DDR5 in a blind comparison test. Then there's tests that REALLY scale well with DDR5. For example, in WinRAR, the difference between DDR5 and the fastest DDR4 is 14%, and 30% (!) to the slowest DDR4, which isn't "slow" in any way, but the performance result of DDR4-4000. At first, I did consider excluding tests Thread Director scheduled onto the wrong core type, but then realized that even workloads run on E-cores are affected by changes in the system's memory configuration, so these should be considered valid results for the scope of this article.

While overall, DDR4-3600 is a tiny fraction slower than DDR4-3200 in applications, there are several bandwidth-loving tests in which it's able to get ahead. The DDR4-3600 configuration is at the sweet spot of just the right frequency, timings, and price point, and it can still run on Gear 1—this shows in our game tests. In the academically relevant 720p gaming resolution that highlights CPU-level bottlenecks, DDR4-3600 is the fastest DDR4 configuration, and just 2% behind the DDR5 reference. DDR4-3200 remains ahead of DDR4-4000 and DDR4-4400 at 720p.

Real-world gaming begins at 1080p, and here, we see DDR4-3600 about 0.9% ahead of DDR4-3200 and about 1.6% faster than DDR4-4400. Games the CPU bottlenecks more, such as CS:GO, throw up results similar to 720p. At WQHD 1440p, a resolution that begins to put noteworthy load on our GeForce RTX 3080, we see quite a few instances where DDR4-3600 gets ahead of even the DDR5-6000 reference: Battlefield V, Borderlands 3, and Metro Exodus. These games benefit from the inflection of just the right bandwidth, just the right memory controller frequency, and much tighter timings than DDR5. At 4K Ultra HD, the DDR5-6000 reference is able to get ahead by a wafer-thin margin in most tests. This is where its sheer bandwidth is able to offer some benefit.

Overall, the choice between DDR4 and DDR5 is an interesting one. Intel typically spans one socket over two microarchitectures, and we're hearing rumors that the future "Raptor Lake" chips retain DDR4 support. This means your DDR4-based Z690 motherboard isn't a "dead-end" platform—no more than DDR5 for the LGA1700 socket. Our results tell you that the choice between DDR4 and DDR5 applies not just to those with reasonably good DDR4 kits they want to hold on to, but also those building a machine from scratch and wanting to save some money by opting for the much cheaper DDR4 memory kits in the market these days. Over the last weeks, memory prices have come down a lot, and you'll be able to find decent DDR4 memory at around $150 for 32 GB. 32 GB of DDR5 starts at roughly twice that for a DDR4-4800 kit and goes to quadruple that for high-end kits like the DDR5-6000 we've been using.

However, there are many things to consider when choosing to take the DDR4 route. You'll need reasonably fast memory able to sustain Gear 1 mode and tight timings—we recommend something similar to our DDR4-3600 configuration. Intel memory controllers aren't very picky about dual-rank modules, at least while you stay in Gear 1. In our performance results, Dual-Rank DDR4 Gear 1 at DDR4-3200/3600 is able to stay ahead of their DDR4-4000/4400 single-rank Gear 2 counterparts. In conclusion, you could avoid paying the hefty DDR5 early-adopter tax if you're willing to live with a single-digit percentage performance hit in applications, and a negligible hit in gaming performance. On the other hand, if you want the absolute highest performance, DDR5 is what you want, but be ready to pay for it. I am happy to report that running DDR5 is trivial, at least on my G.SKILL Trident Z memory kit. Just enable XMP and you're done. Overclocking and tweaking was also trouble-free, and straightforward, which is an impressive feat for a technology that's as new as DDR5.