• Welcome to Overclockers Forums! Join us to reply in threads, receive reduced ads, and to customize your site experience!

DDR4 Memory - Bandwidth, Latency, Quad vs Dual Discussion

Overclockers is supported by our readers. When you click a link to make a purchase, we may earn a commission. Learn More.

mackerel

Member
Joined
Mar 7, 2008
*********These posts were cut off from another thread and continued here to prevent further off topic discussion in that thread *********


Faster cores are certainly possible, but speed is not the sole design factor. Efficiency goes out of the window as we overclockers already know. It would take ever more exotic materials and the exotic pricing that comes with it. Also the scaling for clock doesn't have the same scope that going wide with more cores does. More cores is pretty much copy paste, increasing cost to make. The optimal point is a balance between clock and cores. Outside of niches, most probably wouldn't benefit from say either 100 cores at 100 MHz, or a single core at say 20GHz.

I'm not saying I don't want more cores, but I do question who exactly are AMD, and now also Intel, selling these high core count (16+) CPUs to in the consumer space? A lot of individual software simply doesn't need or can scale well to massive numbers of threads, so the value for those cases is running lots of different software at the same time. It makes sense in the "pro" or server markets, but consumer? Even for those cases where high core counts are worthwhile, I think we're better served with multi-socket than trying to cram it all into one socket. Spread the power spatially, and also more sockets means more ram bandwidth, helping with the pressure there.

As for 8086k, partially agree it doesn't make sense over 8700k if you run it stock, but it does seem on average it gets you a bin more OC headroom. Might not be worth it for everyone, but it is there. I only got an average one, doing 5.2 easily, cooler than my 1700 does 4.0.
 
Last edited by a moderator:
More RAM bandwidth would be great with RAM prices from 3 years ago, not so much now. Software will eventually come around to more multi core optimization, but that will require a lot more people to be using those chips in the first place. Intel needs to be involved because AMD just couldn't drive the movement on their own. I would expect several years of more cores than uses before software catches up.

Honestly, until/unless the average consumer's uses find a new need, I think we're at the point that computers are fast enough for the vast majority. In fact, massive overkill seems to be in effect. People are using phones more and more and some of the novelty of a powerhouse home computer may have worn off. Given my needs, I can see my rig meeting those needs for the next decade. That will put me near 70 years old if I even live that long. I'd like to build a Ryzen rig just because but I wouldn't expect it to check my email or post here any faster than my daughter's FX 6350 rig with an OCZ SSD does.
 
I think we passed the "good enough" phase for the masses years ago. A lot of people buy computers as commodities, where price is about the only thing they look at.

As for ram bandwidth, pricing doesn't really matter in that. 16GB is going to cost about the same, regardless if it's is 4x4, 2x8 or 1x16.
 
As for ram bandwidth, pricing doesn't really matter in that. 16GB is going to cost about the same, regardless if it's is 4x4, 2x8 or 1x16.

Ahh, I see what you're saying (I think). Better utilization of bandwidth through wider pipelines? Or A and B can travel through C and D instead of just C. If I'm making myself clear, or understanding clearly.
 
RAM bandwidth is not the issue. Memory controllers can't handle that anyway. Theoretical max memory bandwidth is usually 30-40% higher than what we actually get. Other thing is that only high multi-threaded environment can use really high memory bandwidth so most chipsets still support two channels as it's cheaper and easier to manufacture.

Personally I see no special difference in DDR4 price in last years. My first 4x4GB 2666 kit cost as much as my last one 2x8GB 4133. There were moments when prices were lower or higher but in general it's not so big difference. For many users capacity has changed. When most needed 8GB then now most are looking for 16GB. It's clear it will cost more.
Graphics cards cost much more and that was really significant difference but RAM not so much.
You have to add to the calculation also high Samsung IC prices. Some kits simply cost more because of used IC but don't worry, soon most will be on cheaper Micron and Hynix ;)

@mackerel
I meant max frequency in a typical environment that could be sold to masses. Max in lab environment is 9 or 11GHz or something but so far it's not possible to make CPU which wouldn't be ridiculously expensive, could use ambient cooling and could be introduced to the mass market at more than ~5GHz.

Even though I don't like the idea of the 8086K as it's just pure marketing, then the main reason why I got one is the same as you said so higher chance on a good chip. In my case I needed high performance CPU which could overclock at stock or lowered voltage and wouldn't overheat in a small ITX case with small cooler. It runs at 4.5GHz 1.1V what is just perfect for 24/7 work (can make some more but it's not really required).
 
For ram bandwidth scaling, we have a similar problem as with cores. You can increase either speed or number of channels. However in this case, more channels don't need software to support it and it just works. The simplistic way I look at it, is that for a given total amount of core performance, it should be balanced with a certain amount of ram bandwidth. The amount of ram bandwidth as a proportion to core has been falling as it simply hasn't kept up. Things like bigger caches that come with more cores help mitigate around that, but it'll still be there in the background. I'd rate dual channel as barely adequate for fast Intel quad cores. Now we're at 6 fast cores, with 8 fast cores on the horizon, I think bringing back support for 3 or 4 channel, at least at the higher consumer end, would help a lot. Lower models can remain on dual. I don't know if the two can be combined onto the same mobo implementation/socket though. If there might be some termination differences between those configurations. Previously to get this core count, you'd need HEDT platforms with quad channel so that wasn't so much a problem. Also at much higher core counts, they came with lower clocks in general, helping to take the edge off the worst.
 
I'd rate dual channel as barely adequate for fast Intel quad cores. Now we're at 6 fast cores, with 8 fast cores on the horizon, I think bringing back support for 3 or 4 channel, at least at the higher consumer end, would help a lot.
Are you talking solely in your world where this matters or the rest where bandwidth really isn't a big concern?

That said, we know AMD CPUs respond to faster RAM, but, that is due to how the CPU was designed. Intel doesn't have that issue, in particular on the mainstream CPUs and to some extent, HEDT.

I still haven't seen a need for 'typical' enthusiast computer use for more than dual channel.
 
Memory performance is not only about the bandwidth and what we see is based on cache-CPU-ram performance and every delay on the way. If CPU has slower IMC then memory bandwidth won't help much. If cache is slow then is the same story.
Intel CL has faster IMC per channel than SL-X. Also, the cache is faster. We can reach the point when dual channel at higher frequency reaches the performance of a typical quad channel setup.
I just doubt that quad channel will be required on anything below 12 cores as long as everything else will work fast.
If Intel decide to release 8 cores for 1151 then I bet it will have the same IMC. It's because of chip size and actual architecture. For quad channel will be required additional pins etc.
 
Personally I see no special difference in DDR4 price in last years.

??? This kit, my exact part #, is $124 USD more than I paid when I built this rig. That's over twice my initial cost. DDR4 RAM prices have over doubled here in the last couple years. That's a pretty special difference to me. LOL
 
Your kit, but in general, prices for the same capacity are not so much different. I'm buying/selling RAM pretty often. Difference is in some series. If you compare typical, popular kits then price for the same capacity was +/- 10-15%, not 50%. If you look at more exotic series then all depends on availability of higher binned IC. Good overclocking kits cost much more in some months but it still was not twice as much.
What more, even stores have +/- 30% on some series. I got 16GB 4133 CL18 Team Xtreem kit for ~$300 maybe 3 months ago, the same kit last month cost ~$400. Now price is ~$450 and is not available anymore.
You can't tell me that your 3000 CL14 kit cost ~$125 when I sold mine maybe half year after release for about $250 (used, no warranty). In this year I'm selling 16GB kits for ~$200-250 depends on specs.

EU prices are like this and the main difference is added local tax:
Ripjaws 4, 4x4GB 3000 cost ~$350 in local stores (first X99 series), now 2x8GB 4000 cost about as much
Crucial 4x8GB 2133 cost ~$450, it was still one of the first DDR4, now 32GB 2666 cost about as much (selling right now)
TridentZ 2x8GB 3200 C14 cost me $300, now it cost about the same
TridentZ 2x8GB 3600 C16 cost me about $270, now it's about $300 but hard to get for some reason
Ballistix 2x8GB 2666 cost about $200, now can get them for about $250
Corsair 4x4GB 3000 cost me ~$300, now most 2x8GB 3000 kits cost $250-280
So average 0-20% higher price. This is how I see it in last 4 years.
 
Last edited:
Are you talking solely in your world where this matters or the rest where bandwidth really isn't a big concern?

That said, we know AMD CPUs respond to faster RAM, but, that is due to how the CPU was designed. Intel doesn't have that issue, in particular on the mainstream CPUs and to some extent, HEDT.

I still haven't seen a need for 'typical' enthusiast computer use for more than dual channel.

Yes it is biased towards my personal interests but that doesn't negate its general impact. Gaming reviews elsewhere demonstrate that ram performance does measurably affect frame rates, at times more than low single digit percentages. Maybe not significantly enough as to make or break a deal, but it is not zero impact either. We are not moving in a positive direction as far as the CPU:ram balance is going and it will have to be addressed sooner or later. I'd just prefer sooner. With Ryzen's existence, and Intel's 10nm woes, there is a lot of uncertainty going on at the moment and I think we are at a point where past behaviour is not a good indicator of future direction.
 
Yes it is biased towards my personal interests but that doesn't negate its general impact.
It surely can though. ;)

It depends on the games tested, res, settings, etc. If you run 1080p with a 1080Ti, I can see how it makes a bigger difference (or multi GPU). But when more realisitc setups are used, those values tend to be low single digits and nearly negligible. For example...https://www.techpowerup.com/reviews..._Memory_Performance_Benchmark_Analysis/9.html

Shows a couple of percent at best difference in most cases. ANd the way they test, in particular the 720p results are exxagerrating any difference because who the hell runs 72-p and a 1080??? That is magnifying the difference. Also, who wants to pay an extra 50%+ over middling sticks to buy high end sticks for minimal gains? I'm wondering what that looks like at 2560x1440 or 4K if that changes things...

Last I checked, quad channel RAM just really didn't make a difference in gaming and how many use the PC. Clearly, there are some uses for it and it is necessary to hear from that perspective (though it would be good to mention what perspective that is coming from), but in general it doesn't really bother 95% of people, even here.
 
Games usually base on the balance between bandwidth and access time. Some titles react better to access time, some to bandwidth but it's still not as high FPS increase as we could expect. On Ryzen improvements are mostly in memory bandwidth, on Intels in balance between bandwidth and access time. It's just because of architecture and the fact that on Ryzen all is linked so memory bandwidth affects pretty much everything. On Intels it's like memory read is the most important as it feeds the CPU and fast cache is covering any delays. It's how I see memory scaling on both platforms.
It still looks like memory frequency helps more in current generation than it was on last chipsets.
 
For ram bandwidth scaling, we have a similar problem as with cores. You can increase either speed or number of channels. However in this case, more channels don't need software to support it and it just works. The simplistic way I look at it, is that for a given total amount of core performance, it should be balanced with a certain amount of ram bandwidth. The amount of ram bandwidth as a proportion to core has been falling as it simply hasn't kept up. Things like bigger caches that come with more cores help mitigate around that, but it'll still be there in the background. I'd rate dual channel as barely adequate for fast Intel quad cores. Now we're at 6 fast cores, with 8 fast cores on the horizon, I think bringing back support for 3 or 4 channel, at least at the higher consumer end, would help a lot. Lower models can remain on dual. I don't know if the two can be combined onto the same mobo implementation/socket though. If there might be some termination differences between those configurations. Previously to get this core count, you'd need HEDT platforms with quad channel so that wasn't so much a problem. Also at much higher core counts, they came with lower clocks in general, helping to take the edge off the worst.
For multi core, multi threading the dual or quad channel memory is not for bandwidth it is for latency. Dual channel or quad channel memory can be accessed by each channel separately. Multi channel memory does not help much with running a single program and the cores accessing the same data. Also the data can be prefetched in cache already like it would be on single channel memory.
Ganged versus unganged
Dual-channel was originally conceived as a way to maximize memory throughput by combining two 64-bit buses into a single 128-bit bus.[disputed – discuss][citation needed] This is retrospectively called the "ganged" mode. However, due to lackluster performance gains in consumer applications,[10] more modern implementations of dual-channel use the "unganged" mode by default, which maintains two 64-bit memory buses but allows independent access to each channel, in support of multithreading with multi-core processors.[11][12]

"Ganged" versus "unganged" difference could also be envisioned as an analogy with the way RAID 0 works, when compared to JBOD.[13] With RAID 0 (which is analogous to "ganged" mode), it is up to the additional logic layer to provide better (ideally even) usage of all available hardware units (storage devices, or memory modules) and increased overall performance. On the other hand, with JBOD (which is analogous to "unganged" mode) it is relied on the statistical usage patterns to ensure increased overall performance through even usage of all available hardware units. https://en.wikipedia.org/wiki/Multi-channel_memory_architecture

There is not much performance to be gained going from dual channel to quad channel. Bench marks linksHyperX Predator RGB 2933 CL15 - DDR4 Quad, dual, single Channel memory test: https://www.purepc.pl/pamieci_ram/h..._cl15_test_pamieci_ddr4_quad_channel?page=0,3

Edit: I was not on this page when I posted this.
 
Last edited:
I was reading an article where a small motherboard didn't have quad channel due to size constraints, so manufacturer made it only dual channel. The author ran benchmarks and there was barely any difference between quad and dual, except for specialized ways of usage, such as zipping and unzipping and some certain games, but most of the time it was so negligent that it was not important. In some cases, actually, dual channel even outpaced quad channel ram.

Someone was posting this link somewhere. Taco will look for herm

- - - Updated - - -

Aye mate, here she is! :cheers:

https://www.google.com/amp/s/www.pc...ocking-truth-about-their-performance.amp.html

View attachment 200085

Genius!
 
You can't tell me that your 3000 CL14 kit cost ~$125 when I sold mine maybe half year after release for about $250 (used, no warranty). In this year I'm selling 16GB kits for ~$200-250 depends on specs.

I can and will. It was $99 two years ago. My receipt:
RAM.JPG

It's $205.99 USD today. (Time of posting this)
RAM2.JPG
 
For multi core, multi threading the dual or quad channel memory is not for bandwidth it is for latency. Dual channel or quad channel memory can be accessed by each channel separately. Multi channel memory does not help much with running a single program and the cores accessing the same data. Also the data can be prefetched in cache already like it would be on single channel memory.

The main point of dual channel, quad channel, is bandwidth. It is raid 0 for ram. That can indirectly help latency, for example, by finishing a transfer operation faster it can start the next one sooner.

About the only case I can think of where multi-channel can help with latency is where NUMA applies. You want ram local to each NUMA node and software or OS has to be smart enough to use that local ram to avoid having to travel across NUMA nodes. Traditionally this was only a "problem" for multi-socket systems but since Threadripper that can happen in a single socket also... also it doesn't actually improve latency, it just prevents it from getting worse.


I think I'll have to do some benching of my own, although I'm not sure how to go about it yet. I do want to run a realistic scenario, so no 720p for me. What are some "twitch" games where competitive players strive for high fps? CS:GO? I don't have that and don't want to buy it just for a bench, so am open to other suggestions. Probably will aim to run 1080p at settings optimised for fps. Note I have another side project on the go, where I'm investigating the difference between AMD SMT and Intel HT, so it will be after that.
 
Twice as expensive! Man I wish I could get something like that..

For the speed I have mine running, G. Skill wants $256 for the 2x8 GB set, at looser timings. That's why some of us paid a little more for the C14 3000 RAM, to get the Samsung B dies. The price difference now makes the $20 difference betweem 3000 MHz and 3600 MHz at the time of purchase a real bargain. I mentioned way back then that we were in the Golden Age of RAM and graphics cards. The "premium" I paid for my RAM was well worth it, as was the $259 I paid for a RX 480 8 GB.

While all these RAM dependent cases may be great stuff, we have people going back to 8 GB on builds because of prices. It won't matter how great dual, quad, or octo channel RAM is if people can't buy it.
 
Back