Some RAM technical questions! Single, Dual, Quad, Latency, Pagefile.

[yoadknux]

Hello,

I'm happy to post on this forum once again!

I am trying to learn more about RAM, a little bit of micro and more of macro. I have a X79 board that supports quad channel RAM, although my questions will not be specific to my hardware.

1. It is "common knowledge" that quad channel is faster than dual channel that is faster than single channel. What exactly does "faster" actually mean? From what I've noticed, when it comes down to RAM benchmark, going into quad channel boosts my "Multi Core" values in "userbenchmark" to 144% (or: 50.5 GB/s, MC Read 52.7, MC Write 50.3, MC Mixed 48.6). This is about 3x better than my "Single Core" result of 17.7 GB/s. Latency is 50 ns. Regarding this:
1a. Which is generally more relevant to everyday PC use, single core or multi-core RAM performance?
1b. On which tasks is single core more important, and in which is multi-core more important?
1c. Is it OK to call those transfer rates RAM bandwidth, or does bandwidth have a different meaning? Is it generally right to say that going into quad channel increases your RAM bandwidth?
1d. Is this latency (50ns) reasonable? What does latency usually affect? Is RAM latency usually the bottleneck, or are there other components that respond at a slower rate? because 80ns sounds like it's a really, really fast response....
2. When the OS or software use my RAM, do they generally care about the capacity of the individual sticks, or do they see only the total capacitance? Is it possible that for some tasks 2x8 could do better than 4x4?
3. Pagefile. Do some software require pagefile to actually operate correctly? If my RAM is never fully used, is PF really necessary?

Would love to hear what you guys have to say! Every time I post on this forum I learn something new.

Thanks!

 

[Johan45]

Quad channel offers higher bandwidth (transfer) but usually higher latency.
Latency will affect some software more than others. The Ryzen platform has a high latency but overall it doesn't affect the average performance too much.
Capacity of the sticks can impact performance slightly as dual rank will have a better result than singl but running 4x single rank sticks is nearly identical.
Yes some software/games will require a page file. I usually set mine to ~ 2GB

 

[yoadknux]

Quad channel offers higher bandwidth (transfer) but usually higher latency.
Latency will affect some software more than others. The Ryzen platform has a high latency but overall it doesn't affect the average performance too much.
Capacity of the sticks can impact performance slightly as dual rank will have a better result than singl but running 4x single rank sticks is nearly identical.
Yes some software/games will require a page file. I usually set mine to ~ 2GB

Thank you for the comment! I really appreciate it!
When talking about RAM bandwidth, does quad channel with say 2133MHz RAM has comparable bandwidth to a dual channel 3000MHz RAM? which configuration is generally better, can the channel configuration "compensate" for low frequency, or is the individual frequency more important?
What type of software is usually affected by latency?
Not sure I understood what you wrote about the capacity, so I'll be more specific, let's say I take an example of 1x8 vs 4x2, let's assume both run as single channel, will it be problematic for the 4x2 to perform the same as the 1x8 because the individual sticks are smaller?

 

[Woomack]

Everything is affected by latency more or less but in daily work it doesn't matter much as most work in reducing delays makes large cache. You can clearly see differences only in benchmarks.
If you have quad channel platform then use quad channel memory. If you have dual channel then you don't have big choice and have to use dual channel for higher performance. Dual channel on dual channel platforms has lower latency than dual/quad channel on quad channel platforms. It's because of how cache, memory controller and some other things work.
On X79 there is no big difference between something like 2666 and 3200. Difference between 2133 and 3000+ is already visible in some programs ( but also not all ).

Read some more about memory ranks. In short, dual rank = faster in applications that can use it. Typical desktop modules look like single sided ( chips on one side of memory module ) = single rank, double sided = dual rank. In servers are also quad+ ranks but performance isn't really better than in dual rank.
Dual rank is faster in some applications but also not all while it overclocks worse and less memory kits on the market are at higher frequency so even if single rank can be a bit slower in something then can be available at higher frequency.
Again differences between memory ranks are visible more in benchmarks than in daily work. Considering that differences in memory speed are not big then ranks are not important for 99% users.
In theory to fill all banks on 4 memory slots you have to use or 2x dual rank or 4x single rank memory. If you have dual channel motherboard then using 2x dual rank ( typical would be 2x16GB modules right now ) you can get about the same performance as on 4x single rank ( 4x8GB from new series or 4x4GB from older ). It's not equal but also not so much different.

 

[yoadknux]

Everything is affected by latency more or less but in daily work it doesn't matter much as most work in reducing delays makes large cache. You can clearly see differences only in benchmarks.
If you have quad channel platform then use quad channel memory. If you have dual channel then you don't have big choice and have to use dual channel for higher performance. Dual channel on dual channel platforms has lower latency than dual/quad channel on quad channel platforms. It's because of how cache, memory controller and some other things work.
On X79 there is no big difference between something like 2666 and 3200. Difference between 2133 and 3000+ is already visible in some programs ( but also not all ).

Read some more about memory ranks. In short, dual rank = faster in applications that can use it. Typical desktop modules look like single sided ( chips on one side of memory module ) = single rank, double sided = dual rank. In servers are also quad+ ranks but performance isn't really better than in dual rank.
Dual rank is faster in some applications but also not all while it overclocks worse and less memory kits on the market are at higher frequency so even if single rank can be a bit slower in something then can be available at higher frequency.
Again differences between memory ranks are visible more in benchmarks than in daily work. Considering that differences in memory speed are not big then ranks are not important for 99% users.
In theory to fill all banks on 4 memory slots you have to use or 2x dual rank or 4x single rank memory. If you have dual channel motherboard then using 2x dual rank ( typical would be 2x16GB modules right now ) you can get about the same performance as on 4x single rank ( 4x8GB from new series or 4x4GB from older ). It's not equal but also not so much different.

Thanks for the comment!

When you talked about large cache you talked about the CPU cache? Is this L1, L2 or L3? Let's say I compare my 4930k to 7700k to Ryzen 1700 - It's 12MB vs 8MB vs 16MB. Does this mean that generally for everyday tasks the Ryzen is the best, then the 4930k then the 7700k?

Thanks for the info about the rank. With my RAM, CPU-Z reports dual rank. I guess that explains why they're not great overclockers - Can't even do CL12 2400MHz (it boots and remains stable throughout Memtest and Prime95, but some applications crash after a while - not BSOD, just crash). Stuck at CL10 2133.

Still not sure I understand, Let's say in theory I have a motherboard with 4 RAM slots, is it impossible to use 4x Dual-Rank sticks?

Also, not sure if my previous question was already answered, let's put rank and configuration aside. Let's say I want 32GB Ram. I can do either 8x4 or 4x8. Both will be quad channel. Will one configuration be better than the other in terms of performance?

 

[Woomack]

Thanks for the comment!

When you talked about large cache you talked about the CPU cache? Is this L1, L2 or L3? Let's say I compare my 4930k to 7700k to Ryzen 1700 - It's 12MB vs 8MB vs 16MB. Does this mean that generally for everyday tasks the Ryzen is the best, then the 4930k then the 7700k?

Thanks for the info about the rank. With my RAM, CPU-Z reports dual rank. I guess that explains why they're not great overclockers - Can't even do CL12 2400MHz (it boots and remains stable throughout Memtest and Prime95, but some applications crash after a while - not BSOD, just crash). Stuck at CL10 2133.

Still not sure I understand, Let's say in theory I have a motherboard with 4 RAM slots, is it impossible to use 4x Dual-Rank sticks?

Also, not sure if my previous question was already answered, let's put rank and configuration aside. Let's say I want 32GB Ram. I can do either 8x4 or 4x8. Both will be quad channel. Will one configuration be better than the other in terms of performance?

New processors have fast and large cache comparing to older series. Except maybe AMD FX which had large but really slow cache. When I say cache then I mean in general L1+L2+L3. L1+L2 is the most important but you can't really compare different generations or CPU brands. There are too many factors which are affecting performance. What you have to know is that what you see in benchmarks is not only memory performance but the way how data is transfered between CPU and RAM ... and there is always cache.

Quad channel DDR3 is not really working much above 2400 regardless of brand. In best case you can set stable 2600. I guess you can set something like 2400 CL11 when you play with additional settings but it's not guaranteed. Samsung based memory from better series could make CL9/10 at 2400.

On desktop motherboards in all slots can be used single or dual rank memory modules. So if you have 4 slots then you can use 4x single rank or 4x dual rank.

4x8GB and 8x4GB will perform about the same. However there is higher chance on stability issues with 8 memory modules than 4. It's related to chipset and motherboard design.

Personally I wouldn't invest more in X79 unless you really don't need faster CPU and other features that can be found on new motherboards while memory price isn't so high.
New dual channel platforms perform almost as good as quad channel DDR3. Soon we will see early samples of DDR5 ( I guess no chipset for next 2-3 years ).

 

[yoadknux]

New processors have fast and large cache comparing to older series. Except maybe AMD FX which had large but really slow cache. When I say cache then I mean in general L1+L2+L3. L1+L2 is the most important but you can't really compare different generations or CPU brands. There are too many factors which are affecting performance. What you have to know is that what you see in benchmarks is not only memory performance but the way how data is transfered between CPU and RAM ... and there is always cache.

Quad channel DDR3 is not really working much above 2400 regardless of brand. In best case you can set stable 2600. I guess you can set something like 2400 CL11 when you play with additional settings but it's not guaranteed. Samsung based memory from better series could make CL9/10 at 2400.

On desktop motherboards in all slots can be used single or dual rank memory modules. So if you have 4 slots then you can use 4x single rank or 4x dual rank.

4x8GB and 8x4GB will perform about the same. However there is higher chance on stability issues with 8 memory modules than 4. It's related to chipset and motherboard design.

Personally I wouldn't invest more in X79 unless you really don't need faster CPU and other features that can be found on new motherboards while memory price isn't so high.
New dual channel platforms perform almost as good as quad channel DDR3. Soon we will see early samples of DDR5 ( I guess no chipset for next 2-3 years ).

A lot of information in this post, just noticed the reviews on your sig too. Damn, a real expert is answering my questions! I feel honored!

So cache size plays a big role but also cpu architecture. Got it.

I play around with my X79 just because I enjoy the scientific aspect of it. I use it for gaming and desktop usage, so there's no big CPU demand. In terms of money spent, investing in a better SSD and GPU will probably give me better desktop+gaming results than buying a Z370+8700k+16GB DDR4.

Did I read that sentence right? DDR3 Quad channel is still on par with DDR4 dual channel? Is this due to the high bandwidth? because you've previously mentioned quad channel setups have higher latency.

 

[Woomack]

Typical memory bandwidth of quad channel DDR3 is between 40 and 55GB/s. After OC can reach about 70GB/s. Typical dual channel bandwidth on new processors/memory controllers is 40GB/s and after OC up to 60GB/s. At the same time quad channel latency is about 60-70ns while dual channel 40-45ns. Of course all depends on many other factors.
In memory bandwidth and latency tests ( the best for that is AIDA64 ) you can see that on some platforms, overclocking cache gives better results than overclocking memory. This also depends on how high memory is set and some other factors. Cache and memory are working together. If you test stability then you should load both, memory and cache and not every diagnostic software can do that.