P4: "error is reproducable..." / Speed

[alexnoe]

OK, I'm using a Pentium 4, 2400, overclocked to 2880 (was 2900 before, but this gave a blue screen once a week, now tring a bit less). Memory is Kingston 512 MB, DDR266, CL2. Memory frequency is 120 MHz, since with an FSB of 160, i could only choose 160 or 120 MHz. CPU temperature is 60°C at 100% CPU load.

According to the readme file, it is very rare to get a "error is reproducable and thus not a hardware problem"-message, but I actually get it several times per exponent. The torture test didn't find anything within 12 hours, while the message comes much more often.
If I really had a hardware problem as a reason for these messages, such as heat or too much overclocking, then the errors would not be reproducable, would they?

Another strange thing is that the benchmark results are 10% slower than comparable systems in the list. I can't imagine that this is caused only by the lower memory frequency

[ET_]

Originally Posted by alexnoe

According to the readme file, it is very rare to get a "error is reproducable and thus not a hardware problem"-message, but I actually get it several times per exponent.
Another strange thing is that the benchmark results are 10% slower than comparable systems in the list. I can't imagine that this is caused only by the lower memory frequency

You are maybe trying to LL-test an exponent near to the FFT soft crossover point, if the errors are "reproducable" it means that the routine is working fine as it reproduces the same values, and the hardware is working as well, as a random failure in the CPU happening in the same exact point of the routine and giving the same result is unlikely to happen, but the error is due to "For reasons too complicated to go into here, the program's error checking is not perfect" :(

As for the benchmarks, try to disable all programs that run in the background, then (if the difference persists) try checking CPU work with a CPU analysis tool like WinTop.

Hope this helps.

Luigi

[Xyzzy]

What happens if you run the CPU at stock speed?

My 2.4B has run at 61C 24x7 since I got it... I've never experienced an error yet... Personally, I don't get all freaked out about CPU temperature...

What is your PCI bus running at? Can you lock it at 33MHz?

Also, in my experience, Prime95 will respond much better to faster memory than to a higher CPU FSB...

In your current setup, you are increasing the work the CPU can do, but you are not bumping the memory up in a equal fashion... DDR memory is already short on bandwidth to begin with...

With dual channel DDR it will be a different story (Supposedly the Granite Bay chipset-based motherboards will be released within the next month!) but for any single channel setup memory performance is critical...

Example: I recently upgraded from a 2GHz P4 with a 100MHz FSB to a 2.4GHz P4 with a 133MHz FSB... My chipset (845G) will run asynchronously, so the memory in both cases is 133MHz... The upgrade, while faster, was not as fast as I expected, because the memory bandwidth in both cases stayed the same... I suspect the bandwidth was close to being saturated at 2GHz and at 2.4GHz it is totally saturated...

Solutions?

I could pop in a DDR333 memory stick since the 845G unofficially supports DDR333...

Dual channel DDR...

RDRAM...

Increase the FSB of the memory as much as possible...

I'm waiting for the dual channel DDR...

[Xyzzy]

Also, I experienced the 10% thing...

Solution:

http://arstechnica.infopop.net/OpenTopic/page?a=tpc&s=50009562&f=122097561&m=7670997135&r=6690916635#6690916635

[alexnoe]

I have an Asus P4B533-E, which supports setting the PCI clock asynchronously to FSB, so it is 33.3 MHz.
I can't set an asynchronous memory clock, so with an FSB of 160, I can only set 120 or 160. I'll look out if I can get a DDR333 module at a good price :)

I hope the dual-channel DDR chipsets and boards will support registered memory (with 4 or 6 memory slots)! Otherwise it would be silly...

[Xyzzy]

Looks like regular old DDR to me... And only 4 slots...

http://users.cybercity.dk/~dsl30381/

Although 2GB will be nice... If you used Crucial DDR266 it would only cost around 500 bux to max the board out...

[alexnoe]

4 pieces of ddr-ram can't work properly if they are unbuffered, unless I've missed some development. If it has 4 slots, then I'd be sure that registered RAM is supported.
If registered RAM is supported, than it would be possible to install 4x1024 MB, which would be about 2000 Euro at the moment.

I've tried installing Intel Application Accelerator.
Prime95 didn't get faster, but Nero DriveSpeed, cdrecord and the such caused BSODs in "aspi32.dll". Deinstalled this IAA crap immediately

[QuintLeo]

4 pieces of DDR can work - non-Registered - in some cases with good design and a lot of luck.

I dunno if I would *trust* such a setup for anything important, though..

[patrik]

The torture test didn't find anything within 12 hours, while the message comes much more often.

I just wanted to comment on my experience about the torture test. I recently (2 weeks ago) overclocked my two computers at home, just taking the standard hardware as high as it would go. (My first overclocking :) )

On the system with only 128 MB memory, the stress test ran for hours without any errors, but then I started Netscape and it failed at once. I had to lower it to 6% overclocking before it became stable.
/Patrik