• We’re currently investigating an issue related to the forum theme and styling that is impacting page layout and visual formatting. The problem has been identified, and we are actively working on a resolution. There is no impact to user data or functionality, this is strictly a front-end display issue. We’ll post an update once the fix has been deployed. Thanks for your patience while we get this sorted.

Question Dual channel vs Quad latency and bandwidth.

J

jamesdsimone

Golden Member
I've been looking for information on bandwidth and latency but my searches have yielded an bunch of useless articles. I assume quad channel PC2133 DDR4 will give you about the same bandwidth as dual channel PC4266 DDR4? Will latency be the same or would one have an advantage?
 
You have to use OCCT's memory bandwidth benchmark to get a better perspective on that.

Here's a DDR4 benchmark: https://forums.anandtech.com/threads/post-your-maxxmem²-score.2608515/post-40891330

However, that is an older version. I can get you the latest score tomorrow.

This benchmark however does not play nice with quad channel RAM, in my experience. Usually gets a lower than expected score on Intel quad channel setups. OCCT is much more reliable.

However, 4266 MT/s is simply too fast a RAM setup. I don't think quad channel 2133 MT/s would be able to beat that. Multiple channels do not scale perfectly in terms of bandwidth.
 
jamesdsimone said:
I've been looking for information on bandwidth and latency but my searches have yielded an bunch of useless articles. I assume quad channel PC2133 DDR4 will give you about the same bandwidth as dual channel PC4266 DDR4? Will latency be the same or would one have an advantage?
Click to expand...

As a general rule in computing, increasing bandwidth through parallelism creates less than perfect scaling, as at least a little bit of time is spent playing "traffic cop." The algorithms that do that are very efficient these days, but it's still non-zero. So there are diminishing returns. (IOW, I would expect dual channel 4266 to provide greater real world bandwidth than quad-2133. But the same principle effects everything from RAID arrays to multi-core processing.)

For memory latency specifically, I don't think you pay much of a penalty for adding channels, at least going from 2 to 4. The traffic cop in that case has more to do, but the bottleneck seems to be the responsiveness of the actual chips - usually 12-15ns.* That seems to be plenty of time for it to work. However, there is presumably some number of channels you can add which will start to increase latency.

* Helpful Chart - The CL rating (the "latency" number that we see advertised) is the number of I/O bus clock cycles that the memory controller will have to wait for the memory chips to respond. Run the bus clock faster, individual cycles take less time but you have to wait more of them, because that chip isn't going to respond any faster than it can. In the chart linked, notice that as clock speed and CL timing increase, the CAS latency in nanoseconds remains pretty consistent. IOW, all the memory specs listed have about the same real-world latency. Up the bus clock? CL increases and you have the same latency. Add channels? Same latency. The chips DGAF, you're going to get the first bit of data you requested in ~14ns and like it.

Also, even 30 years ago, memory chip latency was still like 60ns. It's been stuck at its current ~12-ish ns for like a loooooong time. There are some very expensive and weird ram techs that get it down to ~6ns supposedly, but I've never seen those as a consumer. Without researching it much beyond that, I'd guess we're up against some fun laws of physics?

Finally, keep in mind that a quad channel board will tend to be more electrically complex and may have some limitations as a result. (It's not atypical for motherboards to be limited to slower bus clock speeds or higher CL waits when all four channels are in use.) So in real-world use, you may experience more of a penalty than you ought to, depending on the hardware implementation.
 
Another thing I'm curious about: Does the application program have to be written in a way to take advantage of dual or quad channel RAM, or is this always done by the cpu regardless of what software is running?

The reason I ask is that I've never noticed much difference when I set up the RAM modules to run in dual channel .vs. single. Maybe I did something wrong though.
 
Jimminy said:
The reason I ask is that I've never noticed much difference when I set up the RAM modules to run in dual channel .vs. single. Maybe I did something wrong though.
Click to expand...
You need to run something membw heavy. If you compare Geekbench subtests with single and dual channel, you will see that some of the subtest scores improve a lot with dual channel RAM.
 
Jimminy said:
Another thing I'm curious about: Does the application program have to be written in a way to take advantage of dual or quad channel RAM, or is this always done by the cpu regardless of what software is running?

The reason I ask is that I've never noticed much difference when I set up the RAM modules to run in dual channel .vs. single. Maybe I did something wrong though.
Click to expand...

It's not so much that the application has to be written differently; more like the amount of data you're trying to process at a given time needs to be too large to fit in the CPUs onboard cache. Because accessing main memory results in the CPU sitting idle so long, most programmers try to avoid that scenario. As a result, most desktop applications (Windows, Office application, Games, etc.) aren't really limited by memory bandwidth.

Some applications, like Photoshop, do routinely process 100+ MBs of data at a time, but they're also doing a lot of CPU work at the same time, so it would be hard to catch the application being bottlenecked by memory bandwidth.

Theoretically, a memory limited application would be something that is performing a computationally easy task on a very large amount of data. I worked supporting data science efforts at my last gig, and those applications (processing multi-GB datasets) ARE very much limited by memory bandwidth. Going back to pre-"Big Data" for a real-world memory-limited application, database servers were always a big one.
 
Games see a significant difference.

1720538576277.png

Personally, I wouldn't recommend anyone run single channel RAM because that's like running a car in first gear. Sooner or later the CPU will get starved for data from main memory.
 
igor_kavinski said:
Games see a significant difference.

View attachment 102657

Personally, I wouldn't recommend anyone run single channel RAM because that's like running a car in first gear. Sooner or later the CPU will get starved for data from main memory.
Click to expand...
TIL. Most of my games are… older.

Aren’t pretty much all desktop platforms at least dual channel now? (Although I assume they’ll fall back to single if you only install one DIMM.)
 
igor_kavinski said:
Personally, I wouldn't recommend anyone run single channel RAM because that's like running a car in first gear. Sooner or later the CPU will get starved for data from main memory.
Click to expand...
Dual channel is definitely faster than single channel, no question. I've seen videos where Quad channel is a lot faster slightly faster and the same as dual channel.
 
Last edited:
You must log in or register to reply here.

TRENDING THREADS

Back
Top