I have an outgoing generation gaming laptop with 5800H and 3070. Came with 2x8GB DDR4 3200C22 ram. Never thought much about it.

Then I heard about bank, not to be confused with rank. I knew about rank from long ago, where I saw Skylake-era systems with otherwise similar speed ram differ in performance by 25% in memory bandwidth limited prime finding usage. At the same speed/timings, 2 rank was generally much better and could out-perform much faster single rank memory. Unfortunately with increasing ram density, it is not common to find 2R modules at 8GB capacity these days, and even 16GB modules have gone to 1R in the past year or so.

But that is rank, and I'm here to talk about bank. The problem with it seems to be along similar lines. The ram in my laptop is in a configuration called 1Rx16. The x16 part is the new problem. As densities continue to increase, the number of chips per module decreases. I'm not up to speed on the technical definitions of bank and rank, but I can see they kinda correlate to the number of chips per module, and more can be better.

1Rx16 has 4 chips
1Rx8 has 8 chips
2Rx8 has 16 chips

It is technically wrong, but I imagine the number of chips a bit like sub memory channels. More of them means you can possibly do more at once. It doesn't increase the peak bandwidth, but it could improve the usable bandwidth.

Swinging back to this forum's topic, I did before and after testing. Out went 1Rx16. In went 2Rx8 at the same speed that I eventually managed to hunt down from Kingston. Crucial didn't give that level of choice. I did Prime95 8192k FFT throughput benchmarks to ensure it was well in memory limited region. The difference? 115% faster. I was expecting something, but that was a lot! Actual numbers were 64.1 iter/s before, 137.9 after.

It's an outlier but a nice once regardless. For another compute case, y-cruncher at 25m, 1b, 2.5b sizes were between +35% and +41% faster. Aida64 synthetic bandwidths were +9% read, +39% write, +19% copy, no change to latency.

Gaming was more mixed, as I think I picked what I'd consider practical quality biased settings which were more GPU limited than CPU limited, even at 1080p. CoD MWII (extreme) +4%, FFXIV EW (maximum) +7%, SOTTR (highest) +12%, Watchdogs Legion (very high) +34%.

Oh, I believe this is also a problem going into DDR5. It is my understanding that DDR5 8GB modules are probably 1Rx16 also. You'd have to go to 16GB modules to get 1Rx8, and even higher capacities if you want to try for 2R. Big caches are good at avoiding this pain.

If I can find a used set of 1Rx8 cheap I kinda want to test that to see exactly how it sits between the two extremes.

I know I'm not the best at technical writing. The following link goes to an Asus page where they describe it also. Maybe they do a better job of it than I did? Note they only go from 1Rx16 to 1Rx8 so it may be a smaller jump than I observe.
https://rog.asus.com/ch-en/articles/g-series-gaming-laptops/ram-chips-are-changing-2021-rog-laptops

I wanted to look further at this in a desktop environment so managed to get some relatively cheap 1Rx16 modules for testing on a desktop. This would be a wall of text if I wrote everything, so I'll just go over the Prime95 ram-limited scenario here, which could similarly apply to bigger LLR tests. I'm looking at making a more digestible version of all the data later on.

Test system has 11700k with practically unlimited power limit. This is a dual channel 4 slot system. So tests with two modules are dual channel 1 DPC, and tests with 4 modules are dual channel 2 DPC. I did not test 2 modules in one channel 2 DPC.

So that I'm testing the ram configuration, and not other factors like ram speed/timings, this testing was run at JEDEC 2133 timings. I did note the 1Rx8 modules had a much lower tRFC than the other modules but for this test I'm not sure it made a significant impact.

I normalised the results to two 1Rx8 modules, configured in dual channel. This is likely the most common configuration unless you have high ram quantity.

104% 2x 2Rx8
104% 4x 1Rx8
103% 4x 2Rx8
100% 2x 1Rx8
98% 4x 1Rx16
74% 2x 1Rx16

I've known for a while that having effectively two rank per channel, however you reach it, performed better. But the benefit here is a bit lower than I was expecting at up to 4%. During Skylake era, I had see improvement of 25% from having 2 rank. I'll need to repeat this at a ram speed of 3200 to see how that impacts things, as that will be more representative of how it would normally be used.

The nasty 1Rx16 memory is significantly behind, but if you somehow have 4 of these it mostly catches up with the other configurations.

It would appear that we have discussed this before. Linking back to my results from then:
http://www.primegrid.com/forum_thread.php?id=9443&nowrap=true#145976
____________

The effect of rank has been known for some time, but I'm expanding it to multiple 2R modules per channel, as well as the lower end 1Rx16 modules which I hadn't encountered until more recently.