Determining memory latencies and stuff

Matija · Post by **Matija** » Sat Apr 07, 2007 10:00 am

Okay, I'm confused and clueless, so I'm asking for help here.

When overclocking and generally setting memory speed, how the heck do you determine the latencies? It seems like a whole lot of trial & error and memory testing... Plus there's the issue of memory voltage and blah blah blah... And there's the issue of me not knowing much about tweaking modern computers

The reason I'm asking is this:

I have an E4300 @ 9x266. If I set the memory (cheap Kingmax DDR2-800) to DDR2-533 (1:1), I have 4-4-4-12 latencies by default, read from SPD. WinRAR then gives me roughly 890 KB/s compression through its benchmark.

However, when I set the memory to DDR2-800 (2:3), the default latencies are 5-5-5-15, which should be worse, but WinRAR suddenly yells "rock on!" and goes to 1070 KB/s, which is a 20% improvement.

I've then dumped the CPU to 8x300 and DDR2-750 (4:5), latencies again 5-5-5-15, and got around 980 KB/s. This looks strange.

Then I've tested the CPU with 9x300 and DDR2-750. WinRAR barely crossed 1100 KB/s, which is strange, as the CPU was overclocked by 1/8 - I was kind of expecting something close to at least 1200 KB/s.

Edit: Another test at 9x266 with memory at 667 (4:5) - 930 KB/s.

As you can see, I'm clueless. I've read everywhere that it's best to keep CPU and RAM at 1:1, but at least with WinRAR, I'm not seeing any advantages, quite the opposite. I don't have any other benchmarks - I don't know what to look for and I don't know how to see what I'm getting from tweaking memory stuff.

Can anyone enlighten me on this whole issue? How representative is WinRAR, anyway?

derekchinese · Post by **derekchinese** » Sat Apr 07, 2007 11:34 am

This does not seem like anything unusual. When you set the DRAM to ddr-800 at 5-5-5-15, you are getting LOOSE timings, but at HIGHER frequncy. When you set your stuff to 4-4-4-12, you are getting TIGHT timings, but LOWER frequency. It is a trade off. When you slightly overclocked, the benchmark was higher because you have more processing power.

Summing it up, you will get the best benchmarks when you have:
1.) the tightest possible timings (4-4-4-12 is probably the tightest possible on those DRAM's)

2.) the most processing power (you probably should not have a problem getting to around 350mhz x 9)

3.) the highest ram frequency (if your processor is at 350mhz, set your ram to 700mhz so that it is 1:1)

Derek

Matija · Post by **Matija** » Sat Apr 07, 2007 12:48 pm

Okay, a trade-off. Gotcha.

Well, no, I didn't

9x266 at mem 800: 1070 KB/s
9x300 at mem 750: 1110 KB/s

Why is #1 almost as good as #2, and why would #3 (9x300, memory at 600) be uncomparably bad? I'm guessing ~900 KB/s.

Trip · Post by **Trip** » Sat Apr 07, 2007 12:56 pm

Computers have bottlenecks. If you relieve one bottleneck, e.g. cpu power, enough for a certain test you might get another bottleneck, ram frequency.

So one isn't necessarily going to be absolutely more important for a test than another.

I haven't fiddled with any of the new processors, but I would guess processor power is generally most important followed by frequency followed by latency. That's a complete guess though.

sjoukew · Post by **sjoukew** » Sat Apr 07, 2007 2:35 pm

Winrar is one of the few applications that can really push memory and cpu requirements. But are you using rar the entire day?
Check also for the programs you use everyday and are the most demanding, and see what happens.
If you look around somewhat you can find many reviews in which all these timings are tested. (xbitlabs.com, anandtech.com, tomshardware.com)
The general conclusion is that the difference isn't that huge that it justifies expensive and fast ram. Only a couple of applications do benefit from the speed increase, and the difference isn't huge.
The problem is somewhat complicated. The fsb runs at 266*4=1066mhz. The ram runs at 533*2=1066 (double channel) with a divider 1:1. The raw throughput can never be higher than the fsb, whatever the memory frequency is.
When the ram is running faster, the divider isn't 1:1 anymore which is a penalty, and the chipset has to cope with this difference. But then the memory timings are kicking in, if the ram isn't accessed straight forward but in a random order the lookup time becomes important, and then the configuration which can deliver the random data the fastest wins, and sometimes that isn't the 1:1 divider but the highest frequency that wins. It is all about milliseconds, timings, hardware and applications.
When you are using an athlon64 with an integrated memory controller it is an entirely different story.

Matija · Post by **Matija** » Sat Apr 07, 2007 3:38 pm

I have a feeling that the difference between 4-4-4-12 and 5-5-5-15 isn't that big. But memory bandwidth can come in handy... I don't know. To me it looks like more bandwidth = better.

Athlons are a different beast, I know. But I'm puzzled about the C2Ds. It almost looks like it's best to get as much memory bandwidth as possible for a given CPU FSB.

JimX · Post by **JimX** » Sun Apr 08, 2007 12:30 am

When the ram is running faster, the divider isn't 1:1 anymore which is a penalty, and the chipset has to cope with this difference. But then the memory timings are kicking in, if the ram isn't accessed straight forward but in a random order the lookup time becomes important, and then the configuration which can deliver the random data the fastest wins, and sometimes that isn't the 1:1 divider but the highest frequency that wins.

I have a feeling that the difference between 4-4-4-12 and 5-5-5-15 isn't that big. But memory bandwidth can come in handy... I don't know. To me it looks like more bandwidth = better.

I have an E6400 on a P5B-E with 2x1GB OCZ DDR2-800.

CPU is set at 320x8=2.56GHz on 1,2 Volts, memory at 800MHz on 2.0 Volts, 4-5-5-13 set by SPD (according to CPU-Z), a 4:5 ratio. Would it be better to lower the memory frequency to get a 1:1 ratio, or tighten timings a bit?

I do not want to go higher on the CPU, temps and power consumption are nice and low. Around 115W idle, 145W load with 4 HDDs, a 6600 GPU and peripherals, according to the (probably inaccurate) APC Powerchute thingy. What to do?

Matija · Post by **Matija** » Sun Apr 08, 2007 9:29 am

Do you have any benchmarks, and which ones?

With that kind of memory, you might be able to hit DDR2-960 with a bit more relaxed latencies, so you'd have a 2:3 ratio.

I would be really, really interested in seeing what would happen then.

You should be having ~1150 KB/s in WinRAR right now, and probably around ~1400 KB/s with memory that high.

But okay, that's pure memory bandwidth. What I'm interested in is this: where exactly (as in, what kind of benchmarks and real-world usage) do you benefit more from tighter latencies than from more bandwidth?

JimX · Post by **JimX** » Sun Apr 08, 2007 10:09 am

I haven't done any benchmarks.

This is my HTPC and file server, its only heavy work is ripping DVDs. Ripping using some old DDR2-533 was not much faster than my old 3,2 Northwood, so I got the Platinum memory (cheap) and did a bit of overclocking.

I am willing to push it a bit more, but the CPU frequency has to stay low(ish), for the reasons stated in my previous post.

But okay, that's pure memory bandwidth. What I'm interested in is this: where exactly (as in, what kind of benchmarks and real-world usage) do you benefit more from tighter latencies than from more bandwidth?

I don't know, that's why I'm asking.

Which is better? What should I do? What?

Matija · Post by **Matija** » Sun Apr 08, 2007 10:20 am

Don't ask me, I started this thread

JimX · Post by **JimX** » Sun Apr 08, 2007 10:42 am

So, I just tried DDR2-960, without adjusting any voltages, and it hanged... But I was saved by this Asus C.P.R. thing.

I think the memory needs more voltage in order to work higher than 800.

I'm useless...

sjoukew · Post by **sjoukew** » Sun Apr 08, 2007 10:53 am

I Finally found the article again in which this is all tested fully from xbit labs. It took a while before I got it, but here it is

I even found part 2 which is testing ddr 2 1000.

JimX · Post by **JimX** » Sun Apr 08, 2007 11:24 am

Thanks! I've got some reading to do...

Matija · Post by **Matija** » Sun Apr 08, 2007 1:14 pm

Very interesting. It seems that I was correct in my assumptions...

DDR2-800 at 5-5-5-15 is always better than DDR2-533 at 3-3-3-10 for a CPU at default FSB, and almost always better than DDR2-667 at 3-3-3-10.

If you do a massive FSB overclock (to 400 MHz), then DDR2-1000 (4:5) at 5-5-5-15 is equal to DDR2-800 at 4-4-4-12.

In the second article, no RAM could do 3-3-3-10 at DDR2-800, however.

So, to conclude, unless you can get really really extra tight latencies (3-3-3-10) at DDR2-667 for your low-FSB CPU (266-333), you are better off using a higher memory frequency with more relaxed latencies. DDR2-800 at 5-5-5-15 tops DDR2-667 at 4-4-4-12.

So that's what I'm going to stick with, even if I go to 8x300 instead of 9x266: DDR2-800. The only question now is how to determine if I can knock off a latency or two.

Edit: And it looks like, in fact, WinRAR *is* representative and you can use it to determine the best memory settings.

JimX · Post by **JimX** » Sun Apr 08, 2007 1:32 pm

That's what I thought too. I'll stick to where I am, a 4:5 ratio, and maybe tighten my timings to 4-4-4-12 to see if it works.

Thanks everyone!

Edit: I raised the FSB to 325 for 2.6GHz, lowered the timings to 4-4-4-12. Everything works!

Still on 1.2V VCore, 2.0V memory.