which speed is better? 2000/1600/1333? check out the results!

[magnus28]

Finally I can put this one to bed.

Some enthusiasts claim that because they can't run faster ram at the rated speed, that the same ram will actually perform better when downclocked - and at tighter timings. While i'm sure maybe SOME rams might behave this way, most simply will perform best when clocked at a higher frequency.

I've seen some benchmark's pulled off of Tom's hardware claiming that ram downclocked to 1600Mhz out performed the same modules clocked at 1800 and 2000. Let me just say I don't take much stock in benchmarks at tom's hardware- nor would I rely on them as gospel. I decided to run my own benchmarks, my own controlled environment, and use different software to confirm my results.

The Test Conditions:

- All benchmark's were performed right after a warm boot up.
- Each test was performed 3 times, and the best result was recorded.
- Bios was set manually, cpu clock speed remained the same, all volts remained constant.
- Cpu/NB frequency remained as close as possible no more than 50Mhz difference.
- Programs used were SiSandra, Everest, and Passmark.











There you have it, proof that 2000mhz ram is fastest when ran at it's intended speed.
Not only did it win all the tests, but it crushed the other speeds.

 

[xoke]

In a memory only benchmark, I doubt anyone will try to tell you slower is better. Problem is, when you go to translate this in to real world performance, the results are much less significant. Furthermore, given that with todays CPUs and on die memory controllers, it gets harder to overclock RAM and CPU to extreme levels at the same time. Usually in order to get one really high, the other must be more moderate. In this scenario, CPU speed trumps all, thus negating the effects of the faster RAM.

My point is, the above tests are expected as they are merely memory benchmarks. Real world (useful) performance and benchmarks that are not completely memory bound may have different (even opposite) results.

 

[Bobnova]

You didn't change all the timings, 7-7-7-24 and 9-9-9-24.
Try 7-7-7-18 or so.

It also depends on platform, 1156 boards don't seem to like their 8x multi, making for very low memory performance at some speeds that should work well.

 

[magnus28]

You didn't change all the timings, 7-7-7-24 and 9-9-9-24.
Try 7-7-7-18 or so.

It also depends on platform, 1156 boards don't seem to like their 8x multi, making for very low memory performance at some speeds that should work well.

give me a break bob, yes I did.

9-9-9-24

7-7-7-24

6-6-6-20

doesn't matter if I go 6-6-6-18 or 5-5-5-15

I ran some more benchmarks changing the last number, and the difference wasn't worth pasting into mspaint.

doesn't matter what platform you are using, what memory controller, what command rate you are using... I don't care.. bottom line is FASTER frequency translates into FASTER performance, even with loose timings.

Don't agree? then post your proof or Deleted

 

[redduc900]

@ magnus28... While you may think it's OK to slam another member for disagreeing with you by telling him or her to "STFU", it's NOT. You should probably re-read the rules you agreed to when you registered for membership here...

Excerpt from Ground Rules: Absolutely NO personal attacks will be tolerated, no exceptions!

Welcome to OC Forums! Please Read This Before Posting! Guidelines and Rules!
http://www.overclockers.com/forums/announcement.php?f=22&a=66

 

[magnus28]

@ magnus28... While you may think it's OK to slam another member for disagreeing with you by telling him or her to "STFU", it's NOT. You should probably re-read the rules you agreed to when you registered for membership here...

Excerpt from Ground Rules: Absolutely NO personal attacks will be tolerated, no exceptions!

Welcome to OC Forums! Please Read This Before Posting! Guidelines and Rules!
http://www.overclockers.com/forums/announcement.php?f=22&a=66

Agreed.. I did not tell Bob, or any member to STFU. There was no personal attack. If I was unclear in my post, I apologize.

The first part of the post was directed towards Bob, when I disagreed with his comment's that I didn't change the end timings at all.

The last part of the post which was seperated by several spaces in it's own paragraph was a generalized comment aimed at no one in paticular. basicaly if you are going to post your opinion, have a SS and facts to back it up or shut the "heck" up.

Again I apologize for being misconstrued.

 

[Bobnova]

Have a look at the core i3 HWBot benchmark results, especially superpi 32m as it is particularly memory intensive.
On the 540, four of the top five are using the slowest memory multi possible, running in the 700 to 800 range rather then the 2000+ range.

The i3 8x memory multiplier is junk, something about it kills bandwidth badly.
That means that above 200bclk you either have to to insanely fast ram (to run the 10x multi), or you run the 6x, which gets better performance then the 8x.

On an AMD platform all the multipliers work well, so you are absolutely right about memory speeds on the platform you tested. You cannot, however, say that the same holds true on a platform you have not tested on.

 

[kayson]

I don't care.. bottom line is FASTER frequency translates into FASTER performance, even with loose timings.

That's not true. Timings are set as a number of clock cycles, which have a constant time period. 2000MHz = 500ps. 1333MHz = 750ps. Ignoring stability, CPU clocking, and memory controllers altogether, there will come a point where the faster speed is negated due to needing more clock cycles between operations. For example: 24 cycles at 2000MHz = 12ns. At the same time, 16 cycles at 1333MHz = 12ns. Beyond this equilibrium point (looser 2000MHz or tighter 1333MHz), it is certainly possible for 1333MHz RAM to outperform its 2000MHz equivalent by having lower latencies.

Yes, practically speaking, you will probably see better performance at higher frequencies with stock timings of each DRAM chip. But you can't discount the timings completely because they can make a difference. Given that most DRAM chips are speed binned anyways, I'd much rather get a 1600MHz stick at a fraction of the cost and OC it then tighten timings. If the memory controller has stability trouble at the higher frequencies (i.e. 1156), tightening timings and lowering frequency can maintain or even exceed performance of a higher frequency operation.

 

[Bobnova]

It comes down to the size of the memory transactions.
If you need to move 2gb of data into or out of ram, then bandwidth (high speed high latency) is king.
If you need to move a billion two byte files into or out of ram, then latency (low bandwidth, low latency) is king.

Memory benchmarks are almost all dealing in huge files, so bandwidth takes the crown. If you use a benchmark that does lots of small memory transactions you'll see latency taking the crown.
It all depends on what you're trying to do.

Personally I would take 1333 5-5-5-15 over 2000-9-9-9-24.

 

[magnus28]

both you and kayson are right to a point.

What you also must concede is that although it takes memory with higher frequency and looser timings a longer period of time to complete a full cycle, it compensates by doing more cycles in a shorter period of time.

I would never take 1333 5-5-5-15 over 2000 Mhz.

And i use my memory for large bandwith hogging games, and moving huge files from hard drive to hardrive. So for me 2000 is prime.

What you also need to realize is that look at the benchmarks again.. the 1333mhz ram with 6-6-6-20 still had a higher latency (41ns) than the 2000 9-9-9-24 (35ns)

so faster read/write, with higher bandwith, and lower latency?

win win for me.


but what the heck Bob i'll try and get my memory to boot at 1333 5-5-5-15 and run the benmarks again - just to prove my point even further.

 

[Evilsizer]

is worth a read if you havent seen/read it already.
How memory frequency affects latency - Tighter timings vs. Higher speeds?

 

[kayson]

What you also must concede is that although it takes memory with higher frequency and looser timings a longer period of time to complete a full cycle, it compensates by doing more cycles in a shorter period of time.

Re-read that. It doesn't really make sense. I think what you meant is that higher frequency and looser timings takes more cycles to complete an operation but compensates by doing each cycle faster. Which is of course true, but the exact numbers and other external factors determine which will actually be faster.

I would never take 1333 5-5-5-15 over 2000 Mhz.

750ps*5 = 3.75ns
500ps*9 = 4.5ns

750ps*15 = 11.25ns
500ps*24 = 12ns

2000MHz will still be faster for a number of reasons aside from operational latency, however, if your memory controller can't handle 2000MHz, you won't see much of a tangible performance decrease due to using 1333MHz. Faster clock speed does not necessarily mean faster operation.

What you also need to realize is that look at the benchmarks again.. the 1333mhz ram with 6-6-6-20 still had a higher latency (41ns) than the 2000 9-9-9-24 (35ns)

I'm not sure how its measuring latency exactly, but its probably measuring it while its measuring the bandwidth. So yes the 2000MHz will have lower latencies because it clocks faster during the operation and thus "responds" faster.

Simply put, it is possible to have lower frequency memory outperform higher frequency memory due to a difference in timings. This does not mean it will occur in every case, and it tends to me more true at the initial release of a technology.

Edit: In all this I forgot that DDR is double clocked (yay acronyms) so technically all of the time intervals I listed should be doubled. But it doesn't really make a difference.

 

[ChanceCoats123]

How about some real benchmarks? I +1 what bobnova said. Look at the Hwbot. Sure, you can win Maxmem with some 2400mhz Cas 9 ram, but Super Pi 32m will go minutes faster on tightly timed ram over fast ram. Facts are facts .

Magical thing called "Application Variance."

 

[Archer0915]

I want some real world tests. or a balance of tests. I have been doing this for years (read, longer than some members have been alive) as have many others here. Forget the marketing crap the manufacturers feed you and put this to the real test.

I have run all of the synthetics and they tell me that speed is king, I have run memory intensive applications and timed the results and timings take the majority of the tests.

Look at it this way and understand what memory is.

HDD is used to store data for use later and is sometimes used when too much information is in memory to hold the information for swap and data manipulation.

Memory is where DATA is either moved to and from or it is worked on.
When you work on something in memory it is already in there and transfer is at a minimum and then the timings rule.

In the situation where there is a lot of swapping from cache to memory the fetch is what is important and when you have high latency memory and a lot of these transactions you are slower than having a reduced speed and tighter timings.

I presented the OP a chart (study actually) that I worked on when I was working with DDR2 showing the strengths and weaknesses of both timings and speed.

 

[Archer0915]

How about some real benchmarks? I +1 what bobnova said. Look at the Hwbot. Sure, you can win Maxmem with some 2400mhz Cas 9 ram, but Super Pi 32m will go minutes faster on tightly timed ram over fast ram. Facts are facts .

Magical thing called "Application Variance."

Did we not go through this already? It just seems so familiar

 

[QuietIce]

Finally I can put this one to bed.

The Test Conditions:

- Cpu/NB frequency remained as close as possible no more than 50Mhz difference.

There you have it, proof that 2000mhz ram is fastest when ran at it's intended speed. Not only did it win all the tests, but it crushed the other speeds.

There you have it! By admission up to 2.5% difference in the cpuNB speed with many test results coming in within 2% of each other. Not what I would call definitive in any way. Too bad the CPU-Z Memory tab wasn't included with the tests so we call all see exactly what those cpuNB speeds were at the time.


PS
It should also be noted that there is one test where higher RAM speeds simply failed to perform. If you'll look closely the Passmark 7.0 "Memory - Allocate Small Block" performance consistently goes down as the timing latencies go up. 3625 @ DDR3-1333CL6, followed by 3594, and finally DDR3-2000CL9 getting "crushed" with that poor 3523 score, 3% lower than DDR3-1333CL6. As predicted small memory blocks, which includes 95% of all program RAM access, is better when the RAM is slow and tight ...

 

[Archer0915]

There you have it! By admission up to 2.5% difference in the cpuNB speed with many test results coming in within 2% of each other. Not what I would call definitive in any way. Too bad the CPU-Z Memory tab wasn't included with the tests so we call all see exactly what those cpuNB speeds were at the time ...

I still have not seen a real world battery of tests. And never will I am sure. I hope I am wrong.

Ice do you remember that old 7750 thread where I had my memory down clocked because it was faster that way?

 

[QuietIce]

I still have not seen a real world battery of tests. And never will I am sure. I hope I am wrong.

Ice do you remember that old 7750 thread where I had my memory down clocked because it was faster that way?

I wasn't sure which thread it was but, yes, I remember you having a thread showing that.

To me the biggest problem with the results shown here is the lack of cpuNB speeds. Since all RAM data passes through the cpuNB it's a critical parameter for this kind of benchmark - and the OP knew that from past conversations ...

 

[Archer0915]

I think I may crank up my i rig and do a little testing myself but the issue is timing stuff by hand. I know I could do some compression and encode tests and take some screenshots of the software timed results but nothing caned really does it.

 

[Grnfinger]

doesn't matter what platform you are using, what memory controller, what command rate you are using... I don't care.. bottom line is FASTER frequency translates into FASTER performance, even with loose timings.

Don't agree? then post your proof or Deleted

They say a picture speaks a thousand words, so.....