Not As What It Seems

After looking at the full suite of benchmarks for both AMD and Intel, there is "underperformance" seen in the benchmarks compared to the other DDR5-8000 in the charts. This requires a page dedicated to the explanation of why that is. These very same performance shortcomings also presented themselves in the G.SKILL DDR5-8000 Trident Z5 Royal memory review as well, but what has changed from that review is a deeper understanding of the cause and also a solution.

The first thing to tackle is whether this may be a testing error. It is our due diligence to write honest and fair reviews. Therefore, to make sure it is not a test system error, or rather a set of bad data, system checks and benchmarks were tested for an A/B comparison with a different DDR5-8000 memory kit, just like in the previous review. It was concluded that the "issue" must reside in the memory itself, or rather the XMP profile. Fast-forward to this Kingston review, as the G.SKILL review already uncovered the main issue, and we can skip directly to the culprit being the refresh tRFC2 and tRFC-SB values in Kingston's 8000 MT/s XMP profile. For those unaware, the XMP profile applies the primary timings, operational voltage and a few secondary values as well, including tRFC.

These tRFC2 and tRFC-SB values in their simplest description, determine in clocks cycles the time that must elapse during a refresh cycle. Essentially, the DRAM cannot be accessed while the refresh of memory cells is ongoing. With tRFC-SB being used for the "Same Bank" refresh, in which the same rules apply. In short, the higher these values are, the "longer" time (clock cycles) is spent being inaccessible to the system. This interrupt must happen due to the physical properties of the memory cells. Also known as volatile storage, it will lose the data in the event of power loss. If not periodically refreshed to ensure that it remains accurate, data bits are corrupted. On the flip side, if you set the value too low, the memory doesn't have enough time to fully refresh, resulting in data loss as well.

By changing the tRFC2 (880) - tRFCSB (760) in the XMP profile to 640–520, benchmarks results came back as expected from previous DDR5-8000 (2x16GB) memory kits tested from KLEVV, Team Group and Patriot, which all holds similar results to each other. This is a 27% decrease in clock cycles wasted on cell refreshing. What has changed from the previous review is a new-found understanding, and a realization that 16 Gbit (2 GB) ICs has these lower tRFC values compared to 24 Gbits (3 GB) ICs. Translation; 16 GB vs 24 GB DIMMs have different values. This realization that the performance drop is not so much an issue with the memory itself, but what it is compared to. The real underlining issue is the comparison used in the benchmarks, which is the Patriot Xtreme 5 DDR5-8000 / 8200 48 GB (2x24 GB) memory kit. It is also the only one to use tRFC2 (640) - tRFCSB (620) in the XMP profile. Translation: Same values as 16 GB DIMMs and what was used for these benchmarks.

Unfortunately, this uncovers a nasty issue for mostly all DDR5-8000 and DDR5-8200 memory kits currently on the market using 24 GB DIMMs. Without manually setting these tRFC values, there is a clear regression in performance from brands like G.SKILL, Team Group and Kingston compared to their 16 GB counterparts inside the XMP/EXPO profile. The reason for these higher refresh times is because of the larger ICs. By following specifications set by either the DRAM manufacturer (SK Hynix in this example) or JEDEC organization, most of the XMP profiles are generally designed (programmed) using a linear progression model and scaling approach.

Below are the results from just changing the tRFC2 (880) - tRFCSB (760) values to 640–520.

Baldur's Gate 3 Results

Counter-Strike 2 Results

Remnant 2 Results