Help me understand the correlation of RAM timings to bandwidth

[SPL Tech]

I am trying to understand how to determine which of two different (hypothetical) RAM chips are the fastest. My understanding so far is that the real speed of a RAM is mostly a function of the clock speed (bandwidth) and the CAS latency. From what I have read, a chip with a CAS of 10 and clock of 400 MHz. will perform at the same total speed as a chip with a CAS of 5 and a clock of 200 MHz. Correct?

If so, then why make the change from DDR2 to DDR3 for middle-grade chips like the Sandy Bridge? If the Corei5 can only handle DDR3-1333, then wouldn't it make more sense to go with a DDR2 chip at 1066MHz and CAS 5 then a DDR3 chip at 1333MHz. and CAS 7?

That brings me to my next question. I am considering buying a Corei5 and I am trying to determine what RAM to get. If a Corei5 memory controller cannot control RAM faster than 1,333MHz, then why are there LGA1155 mobos that can run 2,400 MHz RAM?

 

[Knufire]

It can control RAM faster than 1333MHz. However Intel only rates them at 1333MHz. It's just a very conservative rating.

Anandtech did two fairly in depth studies on RAM speed and latency vs real world performance, one in general using Sandy Bridge CPUs and one on iGPU performance using Ivy Bridge chips. I'll try to find them but I'm on mobile.

Sandy Bridge Memory Scaling: Choosing the Best DDR3

www.anandtech.com

Memory Performance: 16GB DDR3-1333 to DDR3-2400 on Ivy Bridge IGP with G.Skill

www.anandtech.com

 

[Woomack]

Memory bandwidth depends from memory controller and memory itself. If you hit max memory controller bandwidth then the only thing you can do is to tighten memory timings for quicker access time but it won't raise bandwidth.

For older Intel series, memory controller was in the northbridge so NB clock could affect maximum bandwidth. In other words FSB. On the other hand, then we had situation when memory wasn't fast enough to reach maximum memory controller bandwidth like we have with newer Intel series and because of longer access times, memory latency was more important ( so tightening memory timings ). It's one of reasons why Core 2 Duo systems with DDR2 were better than the first DDR3 clocked even higher.

Sandy Bridge or newer Intel series have memory controller linked to their internal bus clock which also affects cache and other things. Simply while raising CPU speed you also raise maximum possible memory bandwidth. That's why higher memory clock has some point only when you are overclocking.
In my tests SB without OC was scalling good at least up to DDR3-1866. Ivy has better memory controller but it's not scalling much higher without OC and DDR3-2133 or higher clock will give similar results.
As Knufire said, Intel set default memory speed to DDR3-1333 or DDR3-1600 for all their series ( no matter if it's desktop or laptop ) and depends from CPU/chipset combo. You can't even change it to higher speed unless you have enthusiast series CPU and board. In this way they guarantee problem-free work at this clock and compatibility with memory standards. Most memory manufacturers let you set higher clock but higher memory ratios are marked as OC so not guaranteed even though all will be 100% stable.

Every new memory generation is bringing also lower voltages, lower heat and some other things. With faster and larger CPU cache we also see that tighter memory timings are not so important as couple of years ago.
You can also notice that some well programmed DDR3 kits can perform nearly the same even though one has tighter and the other more relaxed timings ( like 10-10-10 and 11-11-11 ) as sub timings sometimes matter more than the first 5 timings that all are showing as main performance factor. It's just selling much better if you write 2400 CL9 than CL10 or 11

I just want to show that overall memory performance depends from many things and at least in last years it depends more from CPU/IMC speed than memory clock itself.

I would still recommend 1866-2133 for Intel as price isn't much higher than the DDR3-1600 kits.

 

[SPL Tech]

Memory bandwidth depends from memory controller and memory itself. If you hit max memory controller bandwidth then the only thing you can do is to tighten memory timings for quicker access time but it won't raise bandwidth.

For older Intel series, memory controller was in the northbridge so NB clock could affect maximum bandwidth. In other words FSB. On the other hand, then we had situation when memory wasn't fast enough to reach maximum memory controller bandwidth like we have with newer Intel series and because of longer access times, memory latency was more important ( so tightening memory timings ). It's one of reasons why Core 2 Duo systems with DDR2 were better than the first DDR3 clocked even higher.

Sandy Bridge or newer Intel series have memory controller linked to their internal bus clock which also affects cache and other things. Simply while raising CPU speed you also raise maximum possible memory bandwidth. That's why higher memory clock has some point only when you are overclocking.
In my tests SB without OC was scalling good at least up to DDR3-1866. Ivy has better memory controller but it's not scalling much higher without OC and DDR3-2133 or higher clock will give similar results.
As Knufire said, Intel set default memory speed to DDR3-1333 or DDR3-1600 for all their series ( no matter if it's desktop or laptop ) and depends from CPU/chipset combo. You can't even change it to higher speed unless you have enthusiast series CPU and board. In this way they guarantee problem-free work at this clock and compatibility with memory standards. Most memory manufacturers let you set higher clock but higher memory ratios are marked as OC so not guaranteed even though all will be 100% stable.

Every new memory generation is bringing also lower voltages, lower heat and some other things. With faster and larger CPU cache we also see that tighter memory timings are not so important as couple of years ago.
You can also notice that some well programmed DDR3 kits can perform nearly the same even though one has tighter and the other more relaxed timings ( like 10-10-10 and 11-11-11 ) as sub timings sometimes matter more than the first 5 timings that all are showing as main performance factor. It's just selling much better if you write 2400 CL9 than CL10 or 11

I just want to show that overall memory performance depends from many things and at least in last years it depends more from CPU/IMC speed than memory clock itself.

I would still recommend 1866-2133 for Intel as price isn't much higher than the DDR3-1600 kits.

It is interesting you recommend 1866+ considering the links above clearly showed that faster RAM does virtually nothing for real-world applications.

Does overclocking an Ivy Bridge raise the RAM timings as it did with the Core2 Duos? Or is the RAM timing independent of the CPU timing? And for an Ivy Bridge, which I plan to overclock, would I be better off with DDR3 2400 11-13-13-35 or DDR3 1600 9-9-9-24?

 

[Woomack]

I'm recommending 1866 or 2133 for simple reason ... price is nearly as high as 1600 so I think it's better to get higher clocked kit. You just have to check some stores to find better prices. About 2 months ago you could get Team Vulcan 2400 11-11-11 for the price of regular 1600 9-9-9 memory.
Even if you won't see big difference in performance then I think it's better to get faster memory if price is nearly the same. All depends if you find something good in lower price.

Ivy Bridge has separate memory and cpu ratios so while overclocking cpu, you are not overclocking memory itself ... however while overclocking cpu you are overclocking memory controller and cache so memory bandwidth will raise even though you won't touch memory settings.

2400 11-13-13 will be faster than the 1600 9-9-9 especially in multithreading but synthetic benchmarks are not showing daily performance. There can be up to 3-5GB/s difference in bandwidth but in games you will see 1-2 FPS more. So again , all depends from price.
There are tests where you see difference like winrar:
http://www.anandtech.com/show/6372/...-to-ddr32400-on-ivy-bridge-igp-with-gskill/11

 

[SPL Tech]

Crap, I ended up buying the 1600 as it was a tad cheaper. Oh well, I guess I can overclock the RAM slightly if need be. So if overclocking the CPU OCs the memory controller, do I need to overvolt the memory controller separately, or is that handled when I overvolt the core? Also, if the CPU has an internal RAM controller, why do LGA 1155 mobos have a north bridge?

 

[Woomack]

1155 boards have IO hub which is more like southbridge and you can find there additional controllers like SATA, USB and other things. Almost everything what was in the northbridge for older series is now in the CPU.
You don't have to overvolt memory controller up to ~DDR3-2400/2600. Most CPUs will work with memory at least up to DDR3-2400 at stock voltages. Memory controller is in the CPU but it has separate voltage.

Simply if you get DDR3-1600 then you can keep all voltages on stock/auto and if you overclock CPU then depends from desired clock you may need to increase CPU voltage and that's all.