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

Quad Channel vs Dual

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

pcride

New Member
Joined
May 23, 2015
Having been out of the PC hardware industry for many years I started to look into the new technologies and found M.2, Sata Express, better audio on mobos, etc.. One mobo I am hearing about is the quad channel memory which seems to have 2 DIMs on each side of the CPU and there is Memory being sold as quad channel or multi channel.

Any thoughts about this? Is it extreme and better performance?
 
There is a lot of money wasted on high speed RAM solutions by a lot of people these days who assume faster RAM will improve their performance. In the vast majority of systems and applications it just ain't so. Modern DDR3 RAM provides so much bandwidth that most modern processors are not bottle-necked by the RAM. Anything over about 1600 mhz DDR3 yields extremely minor performance gains. Check it out. Do some research on the performance impact of high speed and low latency RAM. People have tested this and there is a lot of documentation out there if you google it.

The exceptions to what I am saying are a few application categories that are very memory intensive (like file compression), servers because they have multiple CPUs and heavy traffic and APUs using the integrated GPU instead of a discrete video card. APUs (Accelerated Processing Units) are common these days with both Intel and AMD and have a GPU component integrated into the chip. Since GPU activity benefits significantly from high memory bandwidth then if you are computing without a discrete video card installed you can benefit significantly from high speed RAM since the GPU wouldl be using system RAM.
 
Mountain said:
Latency wins over frequency (Mhz)
Click to expand...

Only when the bandwidth needs of the CPU are already satisfied. Then the lower latency will give a little performance advantage but not nearly as much as most people imagine and throw down money for. Bottom line is many people have wasted a lot of money buying into marketing hype when they should have looked at what the testing research actually says when applied to real life applications.
 
Actually higher frequency = lower latency so for optimal performance is required some balance. New CPUs have so fast and large cache that latency is less important. We can see it in DDR4 where memory controller/cache speed counts more than memory speed.
 
This^. The memory bandwidth needs of the cores is being met very well by the cache.
 
Higher frequency with low latency comes at a price. I have not seen a noticeable difference which is why I stick with 9-9-9-24 1600Mhz DDR3 ram for all my systems. I've created an excel sheet to help with picking out RAM for your system. however, the upload part of the forum is not working. I included a jpg instead

Latency.pdf.jpg

2400Mhz with a CAS latency of 10 seems to be good.
 
Mountain said:
I included a jpg instead
Click to expand...

Your assumption in the table is very naive. CPU is always fetching 64 bytes from memory not just one word. A single DD4 memory module return 64 bit (8 byte) of data at once. Quad channel setup allows you to fetch 32 bytes at once. A single thread still needs to wait two CL to fetch another 64 bytes even if the memory row was already opened. So the next or/and forth "word" will be read from CPU L1 cache. These words can fell in another 64 byte strip of data and you'll need to wait another 2*CL or even open another row of memory but from four consequent 64 bit word at least two will be read from L1 cache because they share the same cache line.
 
Unclocker,

Actually, its all formula's. Those numbers are not random. Plus, you need to pay attention to detail. Those are DDR3 specifications.

Tracy

PS Besides, I think its all hype anyway. You're not going to "see" the difference from one ram chip to another unless you compare apples to oranges. 1333 to 3100Mhz RAM. You may shave off a few seconds at best. Big deal.
 
Mountain said:
Unclocker,
Actually, its all formula's. Those numbers are not random. Plus, you need to pay attention to detail. Those are DDR3 specifications.
Click to expand...

DDR/DDR2/DDR3 also out 8 bytes at once and CPU cache line is 64 bytes long from very old Pentium processors. I understand that this is a formula but delay for second and third word reads can't be calculated this way because of a cache as I describe it before.

You're not going to "see" the difference from one ram chip to another unless you compare apples to oranges. 1333 to 3100Mhz RAM. You may shave off a few seconds at best. Big deal.
Click to expand...

Agree. For CL limitings bigger than 10 you can possibly see the difference only in 16+ thread application. Only applicable for highend Xeon multicore processors.
 
I feel there should be a comparison ram calculator. we give it 1600Mhz, 9-9-9-24 then gives us the output speed in nanoseconds or what ever it will be.

3100Mhz RAM @ CAS 12 seems slow to me. but the Frequency is almost double of 1600, so it should perform about the same as or slightly faster than 1600Mhz ram. Yes?
 
Mountain said:
I feel there should be a comparison ram calculator. we give it 1600Mhz, 9-9-9-24 then gives us the output speed in nanoseconds or what ever it will be.

3100Mhz RAM @ CAS 12 seems slow to me. but the Frequency is almost double of 1600, so it should perform about the same as or slightly faster than 1600Mhz ram. Yes?
Click to expand...

You should consider all three timings if you want to predict random access time but anyway it will be close to 50+-20ns for both good old DDR and DDR4. It is the time needed to read one line of L1 cache from memory. Chips can't open/close raws much faster than 200Mhz. Dynamic memory random access time doesn't change over time.
Bigger frequency (even with bigger timings) is important in multi threading applications. Memory is formed of independent banks and memory controller do not wait for one bank to open raw. It can send commands to different banks to open different raws needed and then read the data when it will be ready. One thread will wait RAS+RCD+n*CL time but the bandwidth will be filled with other reads and that is why higher frequency is important. If you don't run many process in parallel and your processor has only 4 cores then performance difference between 1600Mhz and 3000Mhz will be less than 10% even in synthetic tests. For 18 core Xeon the difference will be bigger.
 
Last edited:
You seem very knowledgeable. Can you make up an Excel spread sheet that compares two types of RAM? then show the speed improvements? It would put to bed the idea that 3100Mhz is worth $1200, where $120 1600mhz ram is not that far off.
 
I think the big picture of what unclocker is saying is what we already know: Ordinary systems doing ordinary tasks do not benefit much at all from high speed memory because they just don't use the extra bandwidth.
 
You must log in or register to reply here.

Similar threads

Cor3000000323
Mixing RAM, what's better, 32gb kit or 48gb mixed?
Replies
15
Views
6K
don256us
don256us
mackerel
1Rx16 to 2Rx8, filling in 1Rx8 testing?
Replies
2
Views
2K
mackerel
mackerel
Voodoo Rufus
Project Donnager, main rig upgrade
2 3 4 5
Replies
80
Views
4K
Voodoo Rufus
Voodoo Rufus
Brando
Any reason to spend big on a z690 motherboard?
Replies
5
Views
2K
Brando
Brando
B
Ethernet Cables and Best Cat Rating. Some Thoughts
Replies
3
Views
184
Railgun
Railgun
Back