AMD Memory Overclocking
Overclocking Tips and Tricks
Spoiler: AMD DDR5 Tweaking
When it comes to overclocking, things get a little bit complicated trying to understand all the technical terms and how they fit together. With the new AM5 socket, AMD has switched to DDR5 exclusively, leaving DDR4 behind. Those familiar with the ideal configuration from the previous platform can still apply some of the same overclocking principles here as well. Those new to the memory overclocking scene or who need a refresher only need to follow a few rules to get started.
First is a quick refresher for AMD and DDR4. The ideal configuration for Zen 2 (3000 series) and Zen 3 (5000 series) based processors will always be to match the internal Infinity Fabric (IF) to the memory controller, which matches the frequency of the system memory as well. This is often referred to as a 1:1:1 configuration. Once the system memory exceeds the capabilities of the memory controller, a secondary system is in place to allow the memory to still function and operate by breaking this ratio. In doing so, the memory controller frequency is now cut in half, operating in a 2:1 configuration. This introduces a major latency penalty that negatively affects games that are memory sensitive with a measurable frame rate loss in some instances. AMD marketing uses DDR4-3600 as the "sweet spot" for Ryzen 3000 series, with DDR4-3800 best suiting Ryzen 5000 series processors. This is where cost vs. performance intersects, giving good results by using just an X.M.P (A-XMP, DOCP, EOCP) DDR4-3600 profile and the motherboard auto settings. For the most part, this is a set and forget type of situation using single or dual-rank memory in the system for a total of two DIMMs. In the example above, DDR4-3600 in a 1:1:1 configuration would be 1800 MHz for all three. Generally speaking, ignoring outliers, 2000 MHz IF and DDR4-4000 RAM is the upper limit for this ratio when using an AMD Ryzen 5000 processor.
With the basics out of the way, we can talk about DDR5 and what to expect now that this AMD (AM5) platform uses it exclusively. The AMD engineering team has changed the formula slightly from the ideal DDR4 configuration. The new "sweet spot" is now DDR5-6000, but with a twist. The Infinity Fabric (FCLK) is now independent and no longer is required to match the system memory for best results. In many instances, a FCLK of 2000 MHz is the default value with a range "up to" 2133 MHz. It is highly suggested to leave this value at 2000 MHz, as higher frequencies often will outright refuse to work with a system hard-lock, requiring you to clear the CMOS to recover. Even though some motherboard manufacturers have cited 2133 MHz as a possibility in media review guides, it may not be until the next CPU generation when we achieve these numbers consistently, without the aid of binned CPUs.
During the testing phase of different AMD motherboards and from observation while working on these reviews, it seems that for now DDR5-6400 and beyond will be out of reach for many users in a memory ratio of 1:1. AMD AGESA/BIOS updates have increased compatibility. The standard setup of a 2000 MHz 1:1 ratio works well through DDR5-6000. While a wide range of AM5 B650/X670/X870 motherboards support DDR5-6400+, this frequency and above can be problematic without some manual intervention for motherboard settings. Often lowering the FCLK and/or changing the memory ratio to 2:1 (Gear 2) will help alleviate some configurations that are struggling to get fully stable.
Caution is advised with raising DRAM voltage over the rated EXPO/XMP profile. Direct airflow may be necessary for long-term stability. This extends to the CPU as well. Raising the integrated memory controller voltage above AMD specifications may cause irreparable damage. Please proceed with care and do research before attempting this. Do not copy and paste values without understanding the impact first, especially if simply taken from screenshots posted on Discord or Reddit.
After previously reviewing the EXPO ULL memory, we found it to be an excellent starting point for future AMD overclocking. With that in mind, we decided to apply the same configuration to this memory kit. For this test, we kept our voltage limit at a maximum of 1.45 V, which is generally considered safe for both beginners and enthusiasts. It's crucial for newcomers to understand that each memory IC can respond differently to voltage changes. SK hynix ICs are typically more forgiving, whereas Samsung and Micron ICs have a lower threshold and can be more prone to permanent damage if voltage is pushed too far. Always verify the brand of your ICs before increasing DRAM voltages to avoid any mishaps.
Because this is SK hynix A-Die, we are able to push the CAS latency down to CL28. While CL26 did boot with 1.55 V, we decided that wasn't worth the headache to validate and not ideal for long-term use. Instead, since this is already starting at 6400 MT/s, we opted to go for CL28 6400 MT/s with a FCLK of 2133 MHz for the maximum impact. Using the
TPU DRAM Latency Calculator we changed the tRFC2 and tRFCSB values to keep the same ns ratio at 6400 MT/s using the aggressive EXPO ULL tRFC2/sb values. Finally, we did some changes to the sub-timings.
The gains are quite noticeable, which is good for those who are willing to overclock the memory beyond the EXPO profile. However, there is still performance left on the table. For those willing to invest the time to tweak each timing to its absolute minimum, there's a slight performance bump to be had in certain games.
DDR5-6400 FCLK-2133 CL28 y-cruncher 2.5B HWBot Submission
Baldur's Gate 3 Results
Counter-Strike 2 Results
Memory Calibration
A completely unnecessary thing to do is flashing the "ULL" OC onto the secondary EXPO profile. It was an adventurous move, and definitely not one for the faint-hearted or those who might need to RMA their memory later. Once you've taken the plunge like me, it simplifies future use; just enable EXPO profile 2, and you're good to go. There's even a personal hidden Easter egg in there for those keen enough to find it.