1Rx16 to 2Rx8, filling in 1Rx8 testing?

[mackerel]

I only recently became aware how bad 1Rx16 ram is, so I debated replacing my laptop ram which is affected. However most new 16GB SODIMMs were 1Rx8 and I was unsure if that was worth the bother. Then I found Kingston listed 2Rx8 modules on their site. Sold! Previously I had seen big changes going from 1R to 2R, and that was worth it in itself. Asus' position on 1Rx16 modules where they try to downplay the difference: https://rog.asus.com/ch-en/articles/g-series-gaming-laptops/ram-chips-are-changing-2021-rog-laptops

Laptop is 5800H (Zen 3 APU, 8 core) with 3070 laptop version (130W).

Before: 2x Samsung 3200C22 1Rx16
After: 2x Kingston (Samsung chip) 3200C22 2Rx8

Games at 1080p
Final Fantasy XIV Endwalker Benchmark (maximum): +7%
Watchdogs Legion (very high): +34%
Call of Duty Modern Warfare II (extreme): +4%
Shadow of the Tomb Raider (highest): +12%

For gaming I decided to use highest reasonable settings with native rendering. Differences could be bigger had I used lower settings.

FFXIV EW also has a loading time result. I didn't focus on this so much but there could be 2-5% speedup in that metric with the new ram.

Compute
Y-cruncher 0.7.10 build 9513
25m +41%
1b +35%
2.5b +41%

Prime95 30.8b17 Throughput Benchmark, 1 worker, 8192k FFT: +115%!!!

I was expecting an increase with Prime95, but this is far beyond expectations. I had previously seen on desktop systems +25% difference going from 1R to 2R, all else being equal. Are 1Rx16 modules that much worse than 1Rx8?

The system has a 5800H CPU, which has great Zen 3 architecture, but unlike its desktop counterparts only has half the L3 cache. Feeding it has always been a concern and this faster ram could be just what's needed to get more out of it.

Synthetic
Aida64 6.25.5400
Mem read +9%
Mem write +38%
Mem copy +19%
Latency - no change

Note this is an older version of Aida64 from before Zen 3 was released, as I never bothered to update my licence. They caution the latency test does not support the CPU, but the result didn't change regardless. As for bandwidths, we see a general increase here. Reads only showed 9% uplift. Writes are a massive 38% higher, with copy kinda in between.

I was never able to correlate Aida64 bandwidth tests to the compute results shown earlier, and I suspect it is a combination of bank and rank interleave that allows the compute software to perform much better.



Anyway, I feel like I have a gap in my testing as I didn't have 1Rx8 as middle ground. I didn't want to buy a set of 1Rx8 for my laptop, but did find some 1Rx16 modules on ebay for desktop. I ordered 4 of them so I can try 2DPC as well, which wouldn't be possible on laptop. The modules are 4GB so I can't test high ram usage. I already have 4 of each 1R (presumed x8) and 2R modules.

Of the tests above, it means I have to drop Watch Dogs Legion and Y-cruncher 2.5b for 8GB configurations, but everything else should still run. I'm just wondering if there are other cases I should add if I'm going to do this? I'm debating adding Cinebench but I believe it was never very ram sensitive. Still, confirming that doesn't hurt.

To recap, the main test configurations will be: 1Rx16, 1Rx8, 2Rx8
I will also test those as 1DPC and 2DPC

I could also test the above single channel and dual channel, but I'm not sure this really adds much value. If you're not running dual channel, you probably don't care about performance.
Hybrid testing is also a maybe. By that I mean, unbalanced channels, or 2DPC with mismatched types. An example was my previous laptop. It had a single SR module fitted which was not accessible, and a hatch with a user accessible module where I put an older DR module. It worked fine, but I have no idea how overall performance was compared to both SR and both DR.

Edit: this phenomenon seems to have been first reported nearly 2 years ago. However I'm not aware of anyone doing testing of the configurations I propose so it should still be an interesting data point. https://www.overclockers.com/forums/threads/1rx8-vs-1rx16.799245/

 

[mackerel]

Ok, I've done a bunch of testing. I will chart it later which should make it easier to see. For now I'll give a written description.

The main test compares the following scenarios:
2x 1Rx16 ram (1 DPC) - reference condition for comparison, should be worst
4x 1Rx16 ram (2 DPC) - probably better, but how much?
2x 1Rx8 ram (1 DPC) - typical configuration, should be better than reference
4x 1Rx8 ram (2 DPC) - good configuration, how much better?
2x 2Rx8 ram (1 DPC) - good configuration, how much better?
4x 2Rx8 ram (2 DPC) - should still be good, but where?

Got that?

Going in I had some ideas of what to expect. Maybe I should have written this before I did the testing, but it is included above.

Test system was 11700k + 1080 Ti, games tested at 1080p
Ram was tested at JEDEC 2133C15 settings, kinda. More on that later.

Summary of gaming results: (all results relative to 1Rx16 ram)

FFXIV Endwalker Benchmark, Max settings: 4x 1Rx8 or 2x 2Rx8 did best at 4% over reference, with other configurations close behind at 3% above reference. Was kinda hoping for more but it is visible beyond measurement tolerances.

COD MWII "balanced" was used to bump up the frame rate into 3 figures. Normally I'd use "extreme" which still gave around high 70's fps. This one was close to measurement tolerances, with best at 2% above reference.

SOTTR "highest", practically everything was around 5% above reference.

GTAV I let it pick defaults 1st time then didn't touch it again. Like EW above, bet configurations were around 7% up, others a bit behind.


Compute/Rendering Results

Y-cruncher
25m: 4x 1Rx8 took the lead with 35% improvement. Worst of the rest was 4x 1Rx16 at 25% improvement.
1b: Similar pattern to above, but 39% and 26% respectively for best and worst.

Prime95: similar trend to Y-cruncher, 40% best, 31% worst.

Blender 3.4.0
This consisted of 3 sub-tests. The middle test uses around 8GB of ram, which is tricky for the test conditions I use where I had 8GB of ram, total. With OS, there's a shortage and I think this is reflected in the results. I'm not testing having not enough ram, so I have to disregard that subtest. The remaining two subtests, I'd say everything was within tolerance of each other so no significant difference.

Cinebench R23:
I didn't expect this to show a difference, and it didn't show a difference.


Synthetics:

Aida64 6.25.5400 (from April 2020 as I didn't renew my update licence)
This is complicated, but the 4x 1Rx8 and 2x 2Rx8 did best overall (read/write/copy/latency). For copy, the 4 module runs did better than the 2 module SR runs. Presumably having the extra bank groups or ranks helps out particularly there.


Was the ram equal?

All ram was set to JEDEC 2133 C15 primaries, with the rest on auto. But there was a significant difference in one of the timings. The 1Rx16 and 2Rx8 modules had a tRFC of 374. The 1Rx8 modules had a tRFC of 278. That might help give it a slight performance edge. I could manually back it off and retest, but I'm not motivated enough to spend that time on it.


Overall impressions of ram

If I had to pick the best two configurations, it would be a draw between using 4x 1Rx8 modules, and 2x 2Rx8 modules. These performed consistently best throughout.

I was unsure about the impact of 4x 2Rx8. It seems even more ranks are not beneficial here, with it performing a bit worse that the best. It is similar to 2x 1Rx8 overall, but has a slight lead in compute.

And last and actually least, how did those nasty 1Rx16 modules do? Using a pair of 1Rx16 was clearly the worst, but the margins observed here varied somewhat. I think I messed up the gaming tests as the 1080 Ti is just weak to show up the performance difference, as we're close to fully GPU limiting. Compared to the best, the differences were 4% in FFXIV EW benchmark, 2% in COD MWII, 5% in SOTTR, and 7% in GTAV. Not exactly game changing. Pun intended.

The gap was bigger in compute cases as already mentioned. What was interesting was that using 4x 1Rx16 modules recovered most of that gap. It was still 2nd worse, but much closer to the other configurations. This is where laptops suffer most I guess, since most can only fit 2 modules.

Since compute was the only area a significant difference was seen, I'll use that for the following ranking, averaging the results previously.
100% 4x 1Rx8 (could the tRFC difference mentioned earlier give this the edge over 2nd place?)
99% 2x 2Rx8
97% 4x 2Rx8
95% 2x 1Rx8
92% 4x 1Rx16
72% 2x 1Rx16

Thoughts on software used

Aida64 bandwidth measurements still don't seem to correlate to anything else. While it might still be an interesting measure in its own right, I'm not convinced of its value.

I knew Cinebench wasn't very ram sensitive in the past and that doesn't seem to have changed at all. Maybe in you have some insane core count CPU it might start to have some impact, but for now I'll continue to skip it for memory performance testing. Blender benchmark is a new one to me, but it seems to behave similarly to Cinebench.

The tests that showed biggest difference were Y-cruncher and Prime95 benchmark. They have different memory access patterns and size requirements but were in same ball park as each other. If you want to know your ram is performing well, I think these are the ones to go for.

And gaming... we kinda knew it already. GPU impacted testing might not be the best to show differences in CPU, and on from there, ram performance. You're going to need a high end GPU. While the 1080Ti I used might have counted as high end 6 years ago, now it is mid range performance even at 1080p. It just isn't pushing the CPU and ram enough to show a big difference.

 

[mackerel]

Finally made the charts for those who don't want to go through a wall of text. Results are relative to 2x 1Rx8 modules, which is a typical dual channel configuration at lower capacities. All results are normalised so that higher is better! Normally if you think for time based things, like latency, higher is worse. I've inverted it to make it more "at a glance" comparable.

Test system was 11700k + 1080 Ti.
Games tested at 1080p
Ram was tested at JEDEC 2133C15


FFXIV Endwalker benchmark has both average fps and load time. GTAV settings I just went with whatever the game picked after checking V-sync was off.




The version of Aida64 used is quite old and came out before the CPU, so it gives a warning when I run the latency test.


Blender benchmark has 3 sub tests, and I'm only showing two here. The other sub-test "junkshop" seems to use around 8GB of ram. Since the smaller sticks I'm using were 4GB, when using two of them I only had 8GB of ram in total and the performance suffered from lack of ram. Therefore those results are not shown.