Comparing 2 sticks to 4 sticks in RAM performance
Comparing 2 sticks to 4 sticks in RAM performance
The setup suggests using 64GB with two 32GB units, possibly offering better value than similar options. Success depends on module density and availability. Early DDR5 chips are expected to become obsolete over time. For top performance, I’d prefer 2R modules instead of 1R, especially when running one per channel. I still have 8GB 2R DDR4 3200 modules that work better than any other 1R option I’ve used. Choosing 32GB units wasn’t about needing more storage but ensuring they’re 2R modules for reliability. I didn’t look for 2R kits because I assumed 16GB would be 1R and didn’t want to waste time searching for a specific model.
They're even more unusual than expected. These kits have appeared more frequently over the past few months, not from the beginning of DDR5. In reality, they are 32GB modules but with half of each chip disabled. I purchased one every month or two because it was $40 used, and it performed identically to a dual-rank 16Gb Hynix A die—just with double the capacity. No memory maker has produced 8Gb DDR5 ICs suitable for native dual-rank 16GB sticks, so achieving this requires disabling half the chip. This isn't uncommon; Crucial did it for their high-end DDR4 5000 Ballistix kits, Corsair does it for the 16GB 6400 CL36 sticks (which are actually 24GB with a third of the capacity), but it's strange they do this on cheaper kits like the 5200, 5600, and 6000 CL40. The only time I've noticed the 5.33.01 version number suggests this modification, even though adding a second rank nearly doubles the price. It doesn't make much difference for DDR5 either. I've experimented extensively with these modules, and the differences between XMP vs. tuned or similar settings are minimal—only about 0.1 seconds in PYPrime, which is the best measurable gap I've seen.
This might be a yield issue. I haven’t tested DDR5 yet because it would mean buying SR modules, but the signs suggested it was similar to DDR4. The advantage would likely appear only in memory-heavy tasks, and higher bandwidths in DDR5 might limit that. My main scenario was compute-heavy workloads like Prime95. Y-cruncher could work too if you have a fast CPU. It seemed more effective in the early DDR4 days, but my tests near the end showed less gain. I can’t re-test everything since I don’t have the hardware now. My recent tests on Rocket Lake only reached 7-8%, compared to 25% on Skylake. Many factors are involved. Software changes or improvements in CPU caches might be responsible for the shift.
Maybe, but then again why wouldn't they just return these memory chips to SK Hynix, and why would it only be Corsair who's doing it? Every answer I get about these kit raises more questions. Not really. DDR5 has so many more bank groups, so single rank DDR5 practically performs like dual rank DDR4. Dual rank DDR5 was more like quad rank DDR4, where there definitely is a performance uplift from it, but its relatively minimal and not worth the extra costs. Eh, I did do some Y-Cruncher testing and couldn't get any definitive difference between dual rank and single rank. Granted, that was on an 8600G so that might have something to do with it, though given it dropped 2.5 seconds going from 6000 to 8400MT/s I'd expect there to be at least some difference. I'll retest later tonight on my 13900K to see if there's a difference (it'll be with 24GB sticks instead as I don't want to pull my 16GB sticks out of a running rig, so the dual rank kit will have the tRFC loosened out to make the comparison apples to apples) as if there's any chip where it could have a difference, its the 13900K.
I considered it could be the chip manufacturer performing the cuts. This might be just a guess, but there could be some performance gains from that approach. I noticed 4R/channel in DDR4 tends to perform slightly worse in terms of speed compared to 2, though still better than a full 1. The percentage change in timing is unclear—testing larger variants might make it more apparent. Smaller variants might suffer more from cache effects. The 8600G chip may not be strong enough against RAM to show this yet.
There are some benefits, generally enabling greater voltage adjustment and increased clock speeds. However, a) it features a fixed PMIC and b) it operates at a dual-rank level, preventing it from reaching those extremes. This likely isn't the main reason for the outcome. Ideally, higher ranks should support better data interleaving and thus marginally improved performance per cycle. The gain here is modest compared to dual-rank configurations using DDR4. It probably falls within acceptable variation. Based on theory, I recall Wendel mentioning this point (I can locate his comments if needed).
For the benchmark results, the 6000–8400 range took 79.39s to 82.57s on the Y-Cruncher 2.5B with the 8600G. The variation wasn’t substantial, and the dual-rank differences were minimal. I recently completed tests with the 13900K, using both XMP and custom settings. All runs were at 6000 CL30, which was the fastest setting for my processor under dual-rank conditions.
XMP Dual Rank vs. Single Rank:
- Y-Cruncher: 1 second
- PYPrime: 0.4 seconds
- 7zip: ~2000 MIPS drop
These gaps are noticeable but unlikely to impact gameplay noticeably. With XMP, PYPrime is faster, and manual tuning reduces the difference to about 0.35s in Y-Cruncher and 0.15s in PYPrime. The performance loss for 7zip stayed close to its baseline.
Overall, while there’s a measurable edge, it’s not dramatic enough to be a decisive factor. On Intel systems, you can easily surpass dual-rank gains by opting for higher clocked models. On AMD, the argument for dual-rank may hold, but the performance boost is usually negligible unless memory capacity is a priority. The extra cost for more RAM would likely be better invested elsewhere.
My findings come from my own testing. I posted it here but can’t locate it, so here’s a link to the same: https://medium.com/@porinapew/ram-perfor...cd668ac715 I’m open to any other results you have from different setups. At those RAM settings (speeds and JEDEC timings), performance generally followed this order for my 11700k dual-channel 4 slots: 4x 1Rx8, 2x 2Rx8, 4x 2Rx8, 2x 1Rx8, 4x 1Rx16, 2x 1Rx16. Some variations appeared in certain trials, likely due to differing access patterns. I was using 2133 JEDEC timings, which represents the slowest DDR4 speed you typically see. This aligns with the highest speeds supported by those modules. Your data is intriguing—the difference is now very small. I’d be happy to contribute if it means buying more RAM, though I don’t need it for anything else. Personally, I’m glad I chose 2x 32GB to cover the 2R modules, which I never confirmed. CPU-Z doesn’t show rankings for this configuration. At least I won’t have to participate in future “Is 32GB enough for gaming” discussions.
I noticed a possible problem with the testing approach since varying stick types likely affect JEDEC subtimings. Different ICs have distinct specifications, so older sticks often have lower tRFC according to JEDEC standards. This explains why the 4x1R setup performs better than the 2x2R despite both being labeled equivalent. For my tests I had to adjust for this, using higher tRFC with 24GB sticks to ensure fairer results. I aimed to verify consistent training of DDR4 2133 at the same tRFC across devices. I own two dual-rank DDR4 kits; if you're interested, I could check your setup on my 5900x. Thaiphoon Burner can display ranks, though it behaves oddly with DDR5. The simplest confirmation would be checking that the sticks are dual-sided—this is obvious because a 32GB IC needed for single-rank 32GB sticks isn’t yet available.
The ram setup is quite different. I remember one non-primary timing wasn't the same, though I don't know which one. I still have all the RAM components, so I can review them later when I have time. My aim was to compare 1Rx16, 1Rx8 and 2Rx8 performance, so I managed to find a match that fits my expectations and hoped JEDEC would be close enough. The official Thaiphoon Burner website was down the last time I checked. I discovered an old copy on an older system, but it didn't recognize anything. Because of the DDR5 configuration, it won't be easy to inspect it visually. I assumed a 32GB configuration would correspond to two 2R modules and decided to stick with that.