Memtest bandwidth changes with multiplier

Anyone else notice that bandwidth reading in Memtest changes with the CPU multiplier? For example, with the exact same memory timings, if I run FSB x 10 or FSB x 11, I get different bandwidth scores...why would this affect memory at all?

Guzuta said:

Anyone else notice that bandwidth reading in Memtest changes with the CPU multiplier? For example, with the exact same memory timings, if I run FSB x 10 or FSB x 11, I get different bandwidth scores...why would this affect memory at all?

Click to expand...

Because the memory controller on A64 systems is right on the CPU. This means that memory bandwidth will increase/decrease with overall CPU speed, regardless of how fast the actual memory modules are running. For example running the CPU at 240x10 with memory at 200 MHz (5/6 divider) will yield a higher bandwidth than CPU at 230x10 with memory at 230 MHz (1:1). Don't believe it? Try it.

what about differences in memory bandwith with the same end cpu speed just diff multipliers and HTT speeds.

for example: 270x10=2700 VS 300x9=2700 ???

eric003 said:

what about differences in memory bandwith with the same end cpu speed just diff multipliers and HTT speeds.

for example: 270x10=2700 VS 300x9=2700 ???

Click to expand...

It would depend on the divider you had your RAM set to (so what speed the RAM was running)

If you had a GOOD 1GB set of RAM that would run 300HTT then it would certainly be a little better than 270....but if your RAM won't do 300, then you have to set a divider.
Basically, when you overclock, you find the highest your RAM will operate (By lowering the multiplier on your CPU to remove that from the equation) Then you set a low divider on your RAM and max out your CPU. Now you have both maximums.....so you find a good combo of CPU multipliers, HTT speeds, and MEM dividers to get both to their max speeds, or as close as possible.

Some times it worth it to go 50 Mhz or so slower on your CPU, if you will gain a decent amount with your RAM speed.

This was discovered a while ago when A64s just came out. Think of the whole CPU as the memory controller, if the CPU is faster the bandwidth goes up. HTT speed isn't all the important as low latency. The A64 design is a lot like the old AXP where Raw bandwidth makes a huge difference. It's all about low latency and high CPU speed.

The memory controller runs at the speed of the CPU, the divider is always in use taking (actual CPU speed)/divider = 200mhz ->Or what ever you want the RAM at.

sorry I wasn't specific enough in my question... I MEANT taking the ram speed to be the same in both examples.. so If the rams runs at, lets say, 200 or 250, (doesn't really matter as long as its the same) I was just wondering when the cpu speed is the SAME, which is better for memory bandwith, high HTT or high multiplier..

Your RAM runs at the same speed of the HTT (if it's running on a 1:1 ratio divider) A higher HTT will make your RAM run faster. So if the HTT is set at a speed that your RAM can run with then that's good as you'll have more memory bandwidth. But if your HTT is higher than the rams max speed, then you'll have to use a divider to lower the RAM speed to something it can handle.

A multiplier will not affect the RAM.


Did that answer your question?

The multiplier will affect how much bandwidth you get. As I stated CPU speed affects bandwidth because the actual memory controller is the CPU and it runs at the same speed at the CPU. There's no such thing as 1:1 on the A64.

Do a simple experiment. Try same HTT speed, Try the max mult, then a slower mult. You'll see the lowered bandwidth with the slower mult.

jivetrky said:

Your RAM runs at the same speed of the HTT (if it's running on a 1:1 ratio divider)

Click to expand...

This I understand without any confusion... as this is a fundamental concept of overclocking..

jivetrky said:

A higher HTT will make your RAM run faster. So if the HTT is set at a speed that your RAM can run with then that's good as you'll have more memory bandwidth. But if your HTT is higher than the rams max speed, then you'll have to use a divider to lower the RAM speed to something it can handle.

Click to expand...

This I also understand. cause increasing the HTT will increase the END cpu frequency (when multiplier kept same).

tebore said:

The multiplier will affect how much bandwidth you get. As I stated CPU speed affects bandwidth because the actual memory controller is the CPU and it runs at the same speed at the CPU. There's no such thing as 1:1 on the A64.

Click to expand...

I'm cool with this too, The memory controller runs at the speed of the cpu (end Frequency)

tebore said:

Do a simple experiment. Try same HTT speed, Try the max mult, then a slower mult. You'll see the lowered bandwidth with the slower mult.

Click to expand...

umm, obviously because the max multiplier test run has a much greater end cpu speed then when you do a slower mutliplier.. for example 200x10=2000 > 200x5=1000

MY SOLE QUESTION is this:

Given 2 different test cases: (any info or settings that I dont mention, assume to be held constant for both test cases, i.e. they are the same for both test cases)

Test case 1: 337 HTT x 8 (multi) =2700mhz (cpu speed)
Ram runs on divider, ram is running at 250

Test case 2: 270 HTT x 10 (multi) =2700mhz (cpu speed)
Ram runs on divider, ram set to run at 250

SOOO, the speed of the ram in both cases is running at the SAME speed. the end cpu speed is ALSO the SAME, therefore because the memory controller runs at the same speed of the cpu, 1:1, I am assuming that the memory controller in each of these test cases runs at the SAME speed.

THE QUESTION IS: which test case has the greater memory bandwith... or do they have the same memory bandwith????????

eric003 said:

Given 2 different test cases: (any info or settings that I dont mention, assume to be held constant for both test cases, i.e. they are the same for both test cases)

Test case 1: 337 HTT x 8 (multi) =2700mhz (cpu speed)
Ram runs on divider, ram is running at 250

Test case 2: 270 HTT x 10 (multi) =2700mhz (cpu speed)
Ram runs on divider, ram set to run at 250

THE QUESTION IS: which test case has the greater memory bandwith... or do they have the same memory bandwith????????

Click to expand...

The bandwidth should be very close, and I believe it will depend on one other factor which you haven't mentioned: total Hypertransport speed, i.e. HTT x HTT multi. My thinking is that whichever case has the higher Hypertransport speed will have the higher bandwidth (they can't be the same for both cases based on the numbers you've chosen).

eric003 said:

MY SOLE QUESTION is this:

Given 2 different test cases: (any info or settings that I dont mention, assume to be held constant for both test cases, i.e. they are the same for both test cases)

Test case 1: 337 HTT x 8 (multi) =2700mhz (cpu speed)
Ram runs on divider, ram is running at 250

Test case 2: 270 HTT x 10 (multi) =2700mhz (cpu speed)
Ram runs on divider, ram set to run at 250

SOOO, the speed of the ram in both cases is running at the SAME speed. the end cpu speed is ALSO the SAME, therefore because the memory controller runs at the same speed of the cpu, 1:1, I am assuming that the memory controller in each of these test cases runs at the SAME speed.

THE QUESTION IS: which test case has the greater memory bandwith... or do they have the same memory bandwith????????

Click to expand...

Actually this scenario is not exactly possible. Here's why:


Test case 1: 337 HTT x 8 (multi) = 2696MHz (cpu speed)
w/ divider 150 -> resulting memory freq: 245.1MHz

Test case 2: 270 HTT x 10 (multi) = 2700MHz (cpu speed)
w/ divider 180 -> resulting memory freq: 245.45MHz

Thus the latter config will have slightly higher b/w. But this is only due to higher CPU frequency.

HTT / divider themselves have no effect on the amount of b/w for K8 systems because the IMC has no contact with the HTT-base (reference) frequency.

Tebore said:

The multiplier will affect how much bandwidth you get. As I stated CPU speed affects bandwidth because the actual memory controller is the CPU and it runs at the same speed at the CPU. There's no such thing as 1:1 on the A64.

Do a simple experiment. Try same HTT speed, Try the max mult, then a slower mult. You'll see the lowered bandwidth with the slower mult.

Click to expand...

From all the information that I can find (and I'll admit that I can't seem to full grasp it) if you have your RAM ratio set to 1:1 (or set to ddr400 or 200mhz) in the BIOS, then changing the CPU multi will not change the mem bandwidth. But if you have your RAM ratio set to anything other than 1:1 in the BIOS, then it will change the bandwidth to some degree. Somehow it plays out that way with the math of things. My brain doesn't quite calculate like it used to (brain damage from hemorage) so i can't seem to full understand the way this all plays out in overclocking. Someone who does, please give a word or two and clear this up (if that's possible)

Here's a couple good links I found:

http://wiki.extremeoverclocking.com/images/a/a6/OscarChart.JPG

http://forums.amd.com/index.php?showtopic=55881

I don't remember seeing this information before, so I'm wondering if any of the overclocking stickies even talk about this. It seems like it has relevance in the whole scheme of things, as far as overclocking.

thanks killrbuckeye, that clears it up.

thanks to everyone else for their help as well...
as always, ocforums is an excellent source of knowledge

as to largon's comment... yes I know that dividers dont exist to give exact equal numbers...and you cant pick 337.5 as an HTT speed... I merely was trying to create a hypothetical example...

jivetrky said:

From all the information that I can find (and I'll admit that I can't seem to full grasp it) if you have your RAM ratio set to 1:1 (or set to ddr400 or 200mhz) in the BIOS, then changing the CPU multi will not change the mem bandwidth.

Click to expand...

jivetrky, I think you're getting memory "bandwidth" confused with memory "frequency". The memory bandwidth is the rate of data throughput to the memory (MB/sec), and on A64 systems this depends primarily on overall CPU speed, regardless of how the memory clock frequency is derived. For example, running 240x10 with memory at 200 MHz will yield a much higher memory bandwidth than running say 220x10 with memory at 220 MHz.

well actually the memory bandwidth will be relative to the speed of the memory bus. And that is the only thing that will affect the memory bandwidth. so if your memory is running at 200Mhz then you have 3200MB/s of memory bandwidth (where PC3200 comes from). IF your memory is running at 220Mhz, then you have 3520MB/s of memory bandwidth.
This has NO relation to the CPU speed. But, because the mem controller is in the CPU, you will have lower latency between the controller and the cpu core, which will yeild better performance with memory access.

I don't have the ability to test your what you say about 240x10 with memory at 200 MHz vs. 220x10 with memory at 220 MHz, But I would be interested in seeing someones benchmark results in such a comparison. I'd like to see a couple of different memory benchmarks if possible.

While I'm not saying you are incorrect that the higher CPU frequency will give better memory performance, the bandwidth does not change.

Like this?

Hehe, well those reports support my theory that given equal overall CPU speed and equal memory speed, but different HTT speed, the Hypertransport Link speed will determine which scenario gives better bandwidth (1000 MHz case just edges out the 800 MHz case).

Now could you post results showing memory bandwidth scaling with CPU speed while memory speed is held constant?

well actually what would prove or disprove this is to run your HTT at 200 with both tests and just increase your multiplier and nothing else. And run a couple memory benchmarks (like sandra or STREAM) not superpi, which is affected much more by CPU speeds than the others.

EDIT: by changing your HTT you are changing the internal dividers between the CPU and ram. (according to the chart I posted earlier) So, since you have an FX chip, you can just change the multipier and keep everythign else the same. (which is what this thread is supposed to be about)

jivetrky said:

well actually what would prove or disprove this is to run your HTT at 200 with both tests and just increase your multiplier and nothing else. And run a couple memory benchmarks (like sandra or STREAM) not superpi, which is affected much more by CPU speeds than the others.

EDIT: by changing your HTT you are changing the internal dividers between the CPU and ram. (according to the chart I posted earlier) So, since you have an FX chip, you can just change the multipier and keep everythign else the same. (which is what this thread is supposed to be about)

Click to expand...

The pickey I posted was to show that the hypothesis that "At the same CPU speed and the same memory speed, multipliers and dividers have little or no effect, at least on Superpi".

As for your spggestions, how about the attached.

You still changed your HTT (and there's no comparison, there's only 1 benchmark)

keep your HTT at 200 and change your multiplier to like 15 (to give you 3ghz)
and then do it at 14 or 13


But you can't really use Superpi to prove or disprove because superpi is dependant on the CPU speed.

Changing your HTT (even though you have a divider that makes your RAM goto stock) will still offset the RAM speed slightly