mersenneforum.org

mersenneforum.org (https://www.mersenneforum.org/index.php)
-   Hardware (https://www.mersenneforum.org/forumdisplay.php?f=9)
-   -   LL test doesn't match benchmark (https://www.mersenneforum.org/showthread.php?t=10481)

drew 2008-07-15 06:10

LL test doesn't match benchmark
 
I just got a new Intel Q9450 Core 2 Quad processor. When I began running Prime95 with a test that has FFT length 2560, it shows a per iteration time of about 100 ms.

The benchmark page shows 56.9 ms, and when I run the benchmark, it shows 58.2 ms for the same FFT length, which is probably reasonable. Why is the iteration time so much slower on the actual LL test?

Drew

Cruelty 2008-07-15 07:27

How many LL tests are you running in parallel? What version of Prime95 are you using?

S485122 2008-07-15 11:16

You compare a single running instance (the V4 benchmarks) with your real life results. Another thing to consider is that the benchmark displays BEST iteration times while a real test displays AVERAGE iteration times. (As an aside : if one runs the benchmark several times one can get the BEST of the BEST iteration times : the variation can be huge especially on 64 bit systems.)

As for your specific problem, the questions by Cruelty are very pertinent, other things to consider are your motherboard chipset, memory type..

Jacob

rx7350 2008-07-15 12:18

Just for the sake of comparison, I recently built a Q9450 system to run Prime95. I am running DDR2 1066 memory (5-5-5-15 , 2x1Gb) at stock speed (1066 MHz). I am running the following LL tests, with iteration times.

A double-check exponent, 1280K FFT - 28ms (.028 sec)
Two 2048K FFT exponents - 47 ms (.047 sec)
One 2560K FFT exponent - 61ms (.061 sec)

Use these as real-life benchmarks. I have been pleased with the iteration times, based on the numerous comments in this forum of the limitations of the memory bottlenecks in the current Intel architecture.

I am running version 24.14

What is the speed of your memory? I noticed that the iteration times decreased dramatically as I went from underclocking the memory to raising it to stock speed. I do have a small overclock (6%) on the cpu, but the most significant difference came from raising the memory speed to stock.

Cruelty 2008-07-15 13:08

Also FSB speed is important in case of Intel CPUs + manually setting an affinity for each instance also helps :tu:
BTW: dont' forget that the best performance from C2Q you will get when using motherboards based on Intel chipsets - Intel has much better memory controller than nVidia...

drew 2008-07-15 14:52

Right now I'm running 4 tests, 2 LL tests and 2 P-1 stage 1 tests using version 25.6 of Prime95. The memory is 800 MHz DDR2 RAM. The motherboard uses an Intel P35 chipset.

I assumed the memory would run at stock speed by default. Is this a BIOS configurable parameter?

Drew

S485122 2008-07-15 15:33

Memory speed is configurable via the BIOS ... on certain machines : on Dell professional machines (latitude and optiplex) it is not configurable. O course you should torture test your machine after applying new settings.

Jacob

robo_mojo 2008-07-15 17:15

[quote]Right now I'm running 4 tests, 2 LL tests and 2 P-1 stage 1 tests using version 25.6 of Prime95. The memory is 800 MHz DDR2 RAM. The motherboard uses an Intel P35 chipset.[/quote]
You have probably exhausted your memory throughput.

Cruelty 2008-07-15 17:22

Memory most likely runs @ DDR2-800 however you should set your timings manually - usually SPD contains 5-5-5-15-2T so the question is what are the specs of your RAM.
Also, if you run your CPU @ stock speed then instead of 9x266 set it to 7.5x333 or 6x400 - you will keep your CPU speed @ stock (or close to it) and improve FSB throughput (increasing chipset voltage is advisable then).

drew 2008-07-20 08:40

[QUOTE=Cruelty;137854]Memory most likely runs @ DDR2-800 however you should set your timings manually - usually SPD contains 5-5-5-15-2T so the question is what are the specs of your RAM.
Also, if you run your CPU @ stock speed then instead of 9x266 set it to 7.5x333 or 6x400 - you will keep your CPU speed @ stock (or close to it) and improve FSB throughput (increasing chipset voltage is advisable then).[/QUOTE]
Ok, I just want to give an update:

I changed my FSB to run at 400 MHz instead of 333 to make use of the 800 MHz memory speed, and shortened the latencies from 5-5-5-15 to 4-4-4-12, then reduced the CPU divider from 8 to 7 which yields a slight overclock (2800 MHz instead of the stock 2666). This made a marginal but not dramatic increase in speed.

I think you're right about the memory being the bottleneck, because after this, I put the multiplier back to 8 (3200 MHz), and this didn't speed things up at all. I'm getting iteration times of 90 now (all four tests are in the LL stage now, which reduced the times from 97 to 90).

I also just learned about the benefits of dual-channel memory, so I ordered a second memory stick (I'm sure this was my biggest mistake). I imagine this will bring things back to more typical performance.

Thanks for everyone's help.

Drew

Cruelty 2008-07-21 10:42

Running C2Q in single-channel memory setup doesn't affect your performance... it simply kills it. Single channel memory is not enough to run at full speed on Core2[U]Duo[/U] not to mention quad core configuration.


All times are UTC. The time now is 08:12.

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2020, Jelsoft Enterprises Ltd.