overclocking does not increase memory bandwidth??

Phil Weldon wrote:
> 'peter' wrote:
> | I have an intel E4300 on asus P5PE-VM with some DDR 500 memory (PC4000) in
> | single channel mode.
> |
> | When at default bus speed 200Mhz, sisoft sandra memory bandwidth test
> | reports bandwidth of 2918 MB/s
> | Using clockgen to increase bus speed to 250Mhz, sissoft sandra still says
> | the memory bandwidth is 29xx MB/s -- essentially, no change. How can this
> | be?
> _____
>
> Your system is running the memory asynchronously with a memory clock :
> CPU clock ratio of 4:5, so the memory bandwidth hasn't changed (the memory
> clock speed is still at 200 MHz.) Evidently you have the memory clock ratio
> (or whatever the BIOS calls it) set to AUTO so that the memory speed is kept
> to the specification of DDR400 stored in the SPID on the memory modules.
>
> Try changing the memory clock : CPU clock ratio to 1:1. Of course, there is
> some chance that your memory modules will not perform correctly when
> overclocked by 25%. If your BIOS/motherboard allow changing memory
> parameters such as timings and voltage, you may be able to increase the
> maximum stable memory speed. If not, then try a lower overclock speed with
> a 1:1 memory clock : CPU clock ratio. Depending on the applications that
> are important to you, the higher overclock speed is more valuable than
> faster memory; after all, with the 2 MByte L2 few main memory accesses are
> required.
>
> Phil Weldon
>

My take on it, is the results don't make much sense.

The P5PE-VM is an LGA775 board, which uses the previous generation 865G
DDR chipset. It has 1:1, 5:4, 3:2 memory dividers. The 865G officially
supports FSB800, and Asus is cheating when running the chip at FSB1066.
Apparently the internal graphics aren't stable at FSB1066, and this may
be due to however the graphics core is clocked with respect to the
hub clock.

According to posts I've read, if you plug an FSB800 processor into the
board, the Auto setting would select 1:1 for DDR400 RAM. If you plug a
FSB1066 processor, they use the 3:2 divider, and the RAM runs at
DDR355. They did that, because they could not run DDR400 RAM at
DDR426 with the 5:4 divider, with FSB1066.

Now, I thought the divider value stays fixed, once you go from the
BIOS to the OS. Using Clockgen, if an FSB800 processor is present,
and the RAM is running at DDR400, raising the clock from 200 to
250MHz, should be causing the RAM interface to run at DDR500. I
suppose you could check with CPUZ, to verify that, for what that is
worth. (CPUZ and Clockgen are written by the same author, so mistakes
could be consistent from program to program.)

So I don't claim to understand it - if the memory clock is higher, the
bandwidth should have changed as well. Both the processor and the
memory are running faster, and I would have expected different
results in each case.

The only conclusion you can reach, is somehow, the memory clock has
been dropped. And I don't see how that can be achieved in this case,
unless the design of Clockgen is changed. Or the BIOS is somehow
involved in the divider choice, while running live in Windows ?
Does Clockgen offer a slider for the memory clock, and was it
moved to a lower setting ? The choices for memory clock, should
be discrete ratios (as the clock is created inside the Northbridge).

I would have expected an 865G to fall flat on its face at DDR500.
It should have taken some tweaking to get it stable up there.
So for whatever reason, my suspicion is that the memory cannot
really be at DDR500. And I just don't see a mechanism to make that
change, once the BIOS choices have been made.

Paul
 >> Stay informed about: overclocking does not increase memory bandwidth?? 

'Paul' wrote, in part:
| So for whatever reason, my suspicion is that the memory cannot
| really be at DDR500. And I just don't see a mechanism to make that
| change, once the BIOS choices have been made.
_____

I agree with you. The results were so clearly related to a 4:5 memory:CPU
ratio that I did not even realize that the original poster had mentioned the
memory was supposedly DDR500 rather the DDR400. Of course, there have been
problems with SiSoft Sandra reporting more than it knows, and the op did not
mention the version used. More details might make the mystery go away.

Phil Weldon

"Paul" <nospam.DeleteThis@needed.com> wrote in message news:essfth$v9t$1@aioe.org...
| Phil Weldon wrote:
| > 'peter' wrote:
| > | I have an intel E4300 on asus P5PE-VM with some DDR 500 memory
(PC4000) in
| > | single channel mode.
| > |
| > | When at default bus speed 200Mhz, sisoft sandra memory bandwidth test
| > | reports bandwidth of 2918 MB/s
| > | Using clockgen to increase bus speed to 250Mhz, sissoft sandra still
says
| > | the memory bandwidth is 29xx MB/s -- essentially, no change. How can
this
| > | be?
| > _____
| >
| > Your system is running the memory asynchronously with a memory clock :
| > CPU clock ratio of 4:5, so the memory bandwidth hasn't changed (the
memory
| > clock speed is still at 200 MHz.) Evidently you have the memory clock
ratio
| > (or whatever the BIOS calls it) set to AUTO so that the memory speed is
kept
| > to the specification of DDR400 stored in the SPID on the memory modules.
| >
| > Try changing the memory clock : CPU clock ratio to 1:1. Of course,
there is
| > some chance that your memory modules will not perform correctly when
| > overclocked by 25%. If your BIOS/motherboard allow changing memory
| > parameters such as timings and voltage, you may be able to increase the
| > maximum stable memory speed. If not, then try a lower overclock speed
with
| > a 1:1 memory clock : CPU clock ratio. Depending on the applications
that
| > are important to you, the higher overclock speed is more valuable than
| > faster memory; after all, with the 2 MByte L2 few main memory accesses
are
| > required.
| >
| > Phil Weldon
| >
|
| My take on it, is the results don't make much sense.
|
| The P5PE-VM is an LGA775 board, which uses the previous generation 865G
| DDR chipset. It has 1:1, 5:4, 3:2 memory dividers. The 865G officially
| supports FSB800, and Asus is cheating when running the chip at FSB1066.
| Apparently the internal graphics aren't stable at FSB1066, and this may
| be due to however the graphics core is clocked with respect to the
| hub clock.
|
| According to posts I've read, if you plug an FSB800 processor into the
| board, the Auto setting would select 1:1 for DDR400 RAM. If you plug a
| FSB1066 processor, they use the 3:2 divider, and the RAM runs at
| DDR355. They did that, because they could not run DDR400 RAM at
| DDR426 with the 5:4 divider, with FSB1066.
|
| Now, I thought the divider value stays fixed, once you go from the
| BIOS to the OS. Using Clockgen, if an FSB800 processor is present,
| and the RAM is running at DDR400, raising the clock from 200 to
| 250MHz, should be causing the RAM interface to run at DDR500. I
| suppose you could check with CPUZ, to verify that, for what that is
| worth. (CPUZ and Clockgen are written by the same author, so mistakes
| could be consistent from program to program.)
|
| So I don't claim to understand it - if the memory clock is higher, the
| bandwidth should have changed as well. Both the processor and the
| memory are running faster, and I would have expected different
| results in each case.
|
| The only conclusion you can reach, is somehow, the memory clock has
| been dropped. And I don't see how that can be achieved in this case,
| unless the design of Clockgen is changed. Or the BIOS is somehow
| involved in the divider choice, while running live in Windows ?
| Does Clockgen offer a slider for the memory clock, and was it
| moved to a lower setting ? The choices for memory clock, should
| be discrete ratios (as the clock is created inside the Northbridge).
|
| I would have expected an 865G to fall flat on its face at DDR500.
| It should have taken some tweaking to get it stable up there.
| So for whatever reason, my suspicion is that the memory cannot
| really be at DDR500. And I just don't see a mechanism to make that
| change, once the BIOS choices have been made.
|
| Paul
 >> Stay informed about: overclocking does not increase memory bandwidth?? 

According to cpuz, FSB:memory is 1:1. And just in case you wonder, I was
able to increase the bus/memory from default 200Mhz to 280Mhz (no crashes
running hotcpu) using mushkin DDR500. At 285Mhz, the system freezes after a
few seconds. With DDR400 memory, it crash at 255Mhz. I wonder if I can push
it even higher if I have DDR600 memory. But I'll never find out, because I'm
going to spend the money on a ga965pds3 motherboard and ddr2 memory.

Back to the topic. I'm going to assume this mystery is a bug in sisoft
sandra, because I *decreased* the clock to 100Mhz, and sisoft still reported
memory bandwidth around 3000MB/s.

I use hotcpu's benchmark test, and the memory related figures do
increase/decrease with the bus speed as expected.

Thanks for the responses.
 >> Stay informed about: overclocking does not increase memory bandwidth?? 

'peter' wrote, in part:
| Back to the topic. I'm going to assume this mystery is a bug in sisoft
| sandra, because I *decreased* the clock to 100Mhz, and sisoft still
reported
| memory bandwidth around 3000MB/s.
_____

Thank you for the extra details, and for the heads warning about the SiSoft
Sandra error. SiSoft Sandra has always had a problem in reporting more than
it knows, though the later versions are getting better in this respect.

The newest version of SiSoft Sandra 2007, version 2007.4.11.22 shows the
correct overclock memory bandwidth increase for me on a system using the
Intel 850 chipset and RDRAM.

Try the Cache and Memory Access Benchmark for a better idea of overall
memory access performance; the size of the L2 cache makes a huge difference,
more than memory bandwidth for many applications.

Phil Weldon

"peter" <nospam.TakeThisOut@nospam.com> wrote in message
news:hMsIh.172$0W5.73@trndny05...
| According to cpuz, FSB:memory is 1:1. And just in case you wonder, I was
| able to increase the bus/memory from default 200Mhz to 280Mhz (no crashes
| running hotcpu) using mushkin DDR500. At 285Mhz, the system freezes after
a
| few seconds. With DDR400 memory, it crash at 255Mhz. I wonder if I can
push
| it even higher if I have DDR600 memory. But I'll never find out, because
I'm
| going to spend the money on a ga965pds3 motherboard and ddr2 memory.
|
| Back to the topic. I'm going to assume this mystery is a bug in sisoft
| sandra, because I *decreased* the clock to 100Mhz, and sisoft still
reported
| memory bandwidth around 3000MB/s.
|
| I use hotcpu's benchmark test, and the memory related figures do
| increase/decrease with the bus speed as expected.
|
| Thanks for the responses.
|
|
 >> Stay informed about: overclocking does not increase memory bandwidth??