4 dimms better than 2 for non-TR Ryzen? Am I reading this article correctly?

[nstgc]

Maybe it's just too early in the morning, but I was reading this article on Ryzen and it seems to suggest four dimms is better than two. This flies in the face of everything I've known since I've been building computers (starting in 2002) where one dimm per channel is optimal. I double checked the CPU used and it's a Ryzen 7. It seems more likely that I'm reading it wrong.

https://www.tomshardware.com/reviews/amd-ryzen-3000-best-memory-timings,6310-2.html

 

[BugFreak]

I'll start by saying I know nothing about Ryzen. That being said I haven't trusted Toms for a long time and most articles I see there seem to be more clickbait than a worthy read. Maybe just me though.

 

[trents]

You will not see any real world difference and four modules puts more strain on the IMC.

 

[nstgc]

...and four modules puts more strain on the IMC.

Which translates to a worse overclock?

 

[trents]

Yes, it can have a negative impact on overclocking.

 

[Woomack]

It doesn't really matter as in both cases will be a similar result.
Example for 32GB:
2x16GB will make max ~4200 and will be limited by high-density IC
4x8GB will make max ~4200, maybe some more but not much, and will be limited by the motherboard or IMC (usually 2 slots on 4-slot mobos are overclocking better because of the mobo architecture)

The optimal clock is ~3600-3800 and tight timings where both setups will be about the same.

Performance-wise, both will be similar too.

Now the difference can be if you use 2x32GB vs 4x16GB:
2x32GB will be limited by the motherboard and sub-timings table - won't make more than ~3800 and it will be at quite relaxed timings
4x16GB will make 4000-4200 at quite tight timings, can be limited by the motherboard or IMC

Again, optimal will be ~3600-3800 but much tighter timings of 16GB modules will give a higher performance in general.

 

[mackerel]

Skimming the linked article, the section on 4 ranks over 2 I'd agree with, however it has to be taken in perspective.

Assuming the same speed and timings, I would take 2 rank per channel over 1 rank per channel for real world usage. In my Prime95-like use cases, which to save ED from pointing it out again, may not represent everyone's use case, the benefit can be tens of %. Note capacity is irrelevant except indirectly from the number and type of modules used.

It is harder to say at what point does raw speed overtake the rank benefit. 3200 dual rank modules gave me better performance than 3600/4000 single rank modules on both Intel and Zen 2.

I don't have data for going beyond 2 rank per channel.

With my benching hat on, it gets more complicated and you might want to optimise for the benchmark. I'll bow to the knowledge of others in this area as going fast and tight isn't an area I've gone aggressively into.

 

[Woomack]

In older generations, there was a difference between single and dual rank operations, even in real-world usage. For maybe 5 years it doesn't really matter (besides single benchmarks) if memory is single or dual rank. Also, every modern 8GB module is a single rank and every 16GB in mass sales is dual rank. There are new 32GB dual rank modules but 16GB single rank parts are not available in stores yet.
In most cases limiting factor is the memory controller, not memory itself or its design. As long as memory passes 3600 clock then it is "optimal". Even timings (regardless of what people on the forums say) don't matter much. There are differences around 0-1FPS in games going from CL18 to CL14 on new AMD or Intel but up to 5 FPS going from 2133 to 3600 clock. There are exceptions of course but this is what we can see in various games or tests, also in OCF reviews.

 

[mackerel]

With the "for my uses" disclaimer, the rank thing certainly isn't limited to older architectures. Coffee Lake is affected. Zen 2 is affected.

From testing in November last year, iter/s for Prime95 5120k FFT on stock Ryzen 3600 (higher is better):
Corsair Vengeance Pro 16GB modules (2R) 3200: 247 (252 if IF increased to 1800)
Kingston HyperX 8GB modules (1R) 4000@3600XMP: 206
Kingston HyperX 8GB modules (1R) 4000@4000XMP: 192
Crucial Ballistix 4GB modules (1R) 2666: 174 (177 if IF increased to 1800)

So in this particular example, 3200 2R modules are about 20% faster than 3600 1R modules. Now they're different modules, so I can't rule out that timings also matter. Certainly in everything I've ran, the Kingston is worse at XMP4000 than it is on XMP3600 profile. I think they had to slack off the timings so much for 4000 it makes it worse than tighter 3600.

 

[Woomack]

I said in general, I know what "your uses" are I mean that in the past games were also affected by ranks, now not really so everything that left is something like Prime95 so more like dedicated software and if you care about it then can get dual rank memory, if not then it doesn't matter.
I wonder how it will change when DDR5 will be released but so far it's delayed and next 2 generations will be probably still on DDR4. There are new 32GB modules on the market but also dual rank so performance should be similar at the same settings (but sub-timings are more relaxed so performance will be lower than on 16GB/2R modules).