Moody IMCs, beware! – Corsair Dominator Titanium RGB DDR5-8000 2x 24 GB Review with Overclocking and Teardown

1 - Unboxing and design 2 - SPD information and heatsink test 3 - Measurements, teardown and PCB analysis 4 - XMP compatibility and overclocking (1/2) 5 - XMP compatibility and overclocking (1/2) 6 - Benchmarks – XMP* 7 - Benchmarks – OC 8 - RGB lighting and conclusion

XMP compatibility

We normally test DDR5 XMP kits on the Asus Maximus Z790 Hero. I was initially skeptical whether DDR5-8000 would run stable here, but I have already successfully tested 7800 Mbps several times with other modules and the Z790 Hero is also listed as one of a few motherboards in the compatibility matrix of the Corsair Dominator Titanium RGB 8000 modules. Since I also use a specially selected CPU with particularly good IMC (at least according to my tests) and, according to various OC forum postings, SK Hynix’s new 24 Gbit M-Die ICs are supposed to be even more clock-friendly than their 16 Gbit predecessors, I was in good spirits at the start.

Well, the XMP profile with DDR5-8000 can be loaded and also booted into Windows, but that’s where the stability ends. DDR5-7200 is the maximum that can be operated stably on my Z790 Hero motherboard with my CPU. With manual tweaking, a better board or an even better CPU, I’m sure more is possible, but the fact that the Z790 Hero or any other 4-DIMM motherboard should be able to run this 8000 kit is, in my opinion, a probable copy-paste error in the specifications. By the way, I tested with BIOS 1501, which was the latest official release BIOS at the time of testing. BIOS 1303, which is listed in Intel’s official XMP list of the fastest 2x 24GB kits on the Hero, was no better.

cdtr_8k_bios_1

cdtr_8k_bios_2

cdtr_8k_bios_3

Now I also have the Z790 Apex motherboard – the white one, not Encore – which is also listed as compatible by Corsair. So I also flashed the new BIOS here, loaded the XMP profile via “XMP II” and tested it. With “XMP II”, all timings and voltages should actually be loaded, while with “XMP I” only the primary timings and voltages are loaded – at least that’s how I’ve always known it. In this BIOS version (1501), however, tWR and VDD2 are unfortunately ignored. It remains to be seen to what extent Asus has deliberately decided to do this. Of course, the kit is still tested with tWR 120, as intended by the manufacturer, in order to avoid any performance advantages or disadvantages.

So, is the kit stable on the Asus Maximus Z790 Apex at 8000 Mbps? Yes and no, sometimes. The second XMP profile with DDR5-8200 is also stable, but only after approx. 50% of all memory trainings. But what is a memory training?

To pick up all readers: When booting, the CPU and RAM negotiate various parameters completely automatically to ensure stable operation. This process takes approx. 10 seconds on Intel mainboards and is not carried out by default at every start. Instead, in most cases the parameters from the last system start are reused. If the “MRC Fastboot” parameter in the BIOS is set to Disabled, the mainboard is forced to perform a new training session each time the system is started. If the mainboard has a postcode display, you can also see when the mainboard is currently performing a memory training. I have put this together for you in the side-by-side (disabled on the left, enabled (default) on the right):

With MRC Fastboot disabled, I have now tested the kit several times for stability. And yes, sometimes the kit trains stably, so it can also complete stress tests in Windows without errors. But if you reboot the system afterwards, the result often looks completely different. It doesn’t matter whether the tWR 120 is used as in the XMP profile or the 54 (Auto), which the mainboard sets automatically. VDD2 or MC2 Voltage is set by the board to 1.4 V instead of 1.2 V as specified in the XMP profile, but this does not matter either. Even manually tweaking the VDD2 (MC) voltage up to 1.5 V or adjusting the DQVRefUp/Down settings does not enable reproducible stable operation. I have also tested a 14700KF with likewise selected IMC and active cooling of the RAM modules, but to no avail. Here are a few screenshots:

cdtr_8k_xmp_apex_instability_randomness (2)

cdtr_8k_xmp_apex_instability_randomness (3)

cdtr_8k_xmp_apex_instability_randomness (4)

cdtr_8k_xmp_apex_instability_randomness (5)

cdtr_8k_xmp_apex_instability_randomness (6)

cdtr_8k_xmp_apex_instability_randomness (7)

cdtr_8k_xmp_apex_instability_randomness (8)

cdtr_8k_xmp_apex_instability_randomness (9)

cdtr_8k_xmp_apex_instability_randomness (10)

cdtr_8k_xmp_apex_instability_randomness

Now, of course, you can discuss again to what extent a RAM kit with DDR5-8000 XMP profile “has” to be stable. Intel does not officially guarantee a clock rate beyond DDR5-5600 JEDEC. Whether the expectation of “Plug and Play” should even exist is open to debate. To make matters worse, Intel’s memory training is sometimes stable and sometimes not. These “mood swings”, as I simply call them, make it much more difficult for users to set up their system and test it for stability. Even if you have the courage to try out the timing and voltage settings in the BIOS yourself, carry out RAM stress tests and make various other sacrifices for a system running at maximum capacity: You can’t rely on a one-off RAM training session with successful stability tests either (anymore)!

But if you reduce the clock rate by just 200 Mbps to DDR5-7800, all these problems are gone at once. Even after the third reboot in a row with MRC Fastboot Disabled, the kit passes all tests without any errors. I then sent the Dominator Titanium modules to the benchmarks with XMP timings tCL 38, tRCD 48, tRP 48, tRAS 98, tRFC2 878, tRFC_sb 758, tWR 120, 1.4 V VDD, 1.4 V VDDQ, but only DDR5-7800. i’m happy to accept 2.5% less clock rate and thus potential bandwidth and performance in exchange for the fact that the system is really stable, the benchmark results are reproducible and I don’t have to worry about the integrity of the operating system.

Incidentally, this phenomenon of “mood swings” in the limit range has always been present with Intel’s DDR5 CPUs, completely independent of the added silicon lottery in IMC hardware quality. So you first need a good IMC and then it must also have a good day (training) so that high frequencies run stably. This is exactly why individual screenshots in relevant OC forums are actually worthless. After all, how often has the user had their system retrained in order to achieve the necessary stability for this screenshot? You will never know for sure.

7800_xmp_downtune_stability (2)

7800_xmp_downtune_stability (3)

7800_xmp_downtune_stability

Of course, even several screenshots, like the one I took today, are not watertight proof. You could also play tricks with system time and system uptime to profile your alleged overclocking skills online. Even an hour-long video that records the entire system boot would not be reliable, because you would have to reboot (train) several times and could delete the video during the time when the screen is black. Not to mention more complex video editing solutions – unfortunately, we’ve already seen something like this in the OC scene. And even then, you would still have to watch this 12-hour video.

What I’m getting at: DDR5 stability on Intel CPUs is difficult to prove online and if in doubt, you should first distrust every screenshot. In the end, all you can do is trust the poster you trust or know personally. Incidentally, some RAM manufacturers are also among these posters of alleged stability screenshots.

Intel should eliminate these “mood swings” of the IMC for the next CPU generation with DDR5 at the latest, if not earlier with a BIOS or MRC (Memory Reference Code) update. Another simple approach would be to display the negotiated training results to the user. Once the system has proven to be stable, it would be possible to read out these training results and enter them in the BIOS. However, this would of course require Intel to make the various training parameters transparent and be a little more open about the internals of its DDR5 memory controller.