Quad Core

[rx7350]

http://www.xbitlabs.com/news/cpu/dis...610234912.html


Sounds like this architecture change plus DDR3 1600Mhz memory would go a long way toward eliminating the memory bottleneck that PRIME95 experiences on current Intel hardware.

[db597]

Originally Posted by rx7350

Sounds like this architecture change plus DDR3 1600Mhz memory would go a long way toward eliminating the memory bottleneck that PRIME95 experiences on current Intel hardware.

Thanks for the info rx7350. I suspect that the DDR3 1600 itself would go far towards solving the problem. More wondering if "native" quad core would have any impact if the memory speed stayed the same.

With "native" quad, the cores can talk to each other without having to go through the FSB. But in this case, there's little or no talking to each other going on in the first place.

[Timmy]

i dont know if anyone has seen this new processor yet which is a MIT startup company. http://www.tilera.com/products/processors.php

[dsouza123]

A quad core notebook is now available.

http://www.xtremenotebooks.com/index...&model_id=1266

Xtreme 917V ACCELERATOR

17" WSXGA+ (1680 x 1050) Super Wide Angle Glass View LCD
NEW! Intel® Q6600 (2.40 GHz x 4) QUAD Core™ CPU - 8MB L2 Cache 1066MHz FSB (Q6700 available)
512MB DDR2 667MHz Micron RAM 1 DIMM (up to 4 GIG)
60GB SATA 150 HDD 5400 RPM (160GB 7200 RPM available)
8X DVD±R/RW Burner with 4X Dual Layer Write Speed Multi Drive - Included

Price: $3,359.00

Also available with the Q6700 (2.66 Ghz x 4)

[sonjohan]

Ok, I've just ordered myself a Quad Q6600 with
2 Gb RAM, 800MHz.
I'll follow your guidelines for configuring the definitive Prime95 threads.

If I want to Torture Test the darn thing, should I torture-test it with 4 different instances, or just 1? And what configuration for the torturetest?

Originally Posted by S485122

Based on my experience with two quadcores, you should go for the fastest memory you can afford. Prime95 speed seems to be proportionnal with memory speed (tested with 675 MHz 5400, 800 MHz 6400 and 1066 MHz 8500 memory.) I think that Cruelty and I have posted enough benchmarks about this in this thread and others... Memory above 8500 must be usable (sufficient voltage amongst other factors) and is very expensive, I am not shure if even good motherboards can use it at their advertised speed for our kind of work. At the moment this would mean that the expensive but good choice is 1066 MHz (8500) memory.

The combination of core 0 LL, core 1 TF, core 2 LL and core 3 TF gives very good results : you can allocate most of the 4MB of shared second level cache per core to LL since TF runs mostly in L1 cache.

[rx7350]

If you don't mind, post your detailed parts list. After you get the thing built and start testing exponents on it, I would be very interested in knowing what the iteration times are, especially if you LL test four exponents.

I'm wanting to build a quad-core system, but am concerned about the memory bottleneck when LL testing four exponents.

Whenever I've tested a new build using PRIME95, it seems that the most thorough test is to test exponents in all cores. It seems to flush out errors better than a torture test, plus it will put the maximum thermal load on the system to test your cooling solution.

[jasong]

Originally Posted by sonjohan

Ok, I've just ordered myself a Quad Q6600 with
2 Gb RAM, 800MHz.
I'll follow your guidelines for configuring the definitive Prime95 threads.

If I want to Torture Test the darn thing, should I torture-test it with 4 different instances, or just 1? And what configuration for the torturetest?

I'm pretty sure I have your same processor. There are only two quad-cores being sold by Intel, and Q6600 is non-'Extreme', correct?

My recommendation is to run two instances of Mersenne number testing, one instance on core 0 or 1 and the second instance on core 2 or 3. On the cores you have left, run anything you want, as long as it doesn't require a lot of bandwidth. And when I say a lot, I mean A LOT. Most of the other prime projects, if run alongside the two instances of Prime95, won't slow it down by a significant amount. The only project I can think of that could possibly be a problem, and this is a SWAG based on what I've read online, would be NFSNet. I don't have any experience there, though, except for reading the forum.

[patrik]

I have just upgraded one of my computers with a Core 2 Quad processor. Unfortunately the fast memory I ordered didn't arrive yet, but I've moved two DIMM's of slower memory from another computer and have run some benchmarks.

Timings are for running four, three, two (in two ways) and one instance of Prime95 (well, mprime actually) simultaneously, with affinity set to the respective core. At each benchmark all instances are running exponents with the same FFT size, but I did this for both 1024K and 2048K FFT's. Also, nothing is overclocked.

2 x 512 MB 667 MHz DDR2 memory (Kingston KVR667D2N5/512 at default speed)

Current double-check (1024 kB FFT)

(milliseconds/iteration)
Core 0 Core 1 Core 2 Core 3
 40.9   40.9   40.7   40.8
 37.5   37.5   34.9    -
 35.5   35.4    -      -
 34.0    -     34.0    -
 33.5    -      -      -

Current first-time test (2048 kB FFT)

(milliseconds/iteration)
Core 0 Core 1 Core 2 Core 3
 83.5   83.1   83.2   83.4
 77.1   77.0   71.9    -
 72.3   72.2    -      -
 70.1    -     69.9    -
 68.7    -      -      -

And one DIMM (single-channel memory), if anyone is interested:
1 x 512 MB 667 MHz DDR2 memory (Kingston KVR667D2N5/512 at default speed)

Current double-check (1024 kB FFT), single channel memory

(milliseconds/iteration)
Core 0 Core 1 Core 2 Core 3
 57.1   57.0   56.9   56.8
 46.9   46.8   39.3    -
 39.3   39.2    -      -
 36.4    -     36.4    -
 34.2    -      -      -

Current first-time test (2048 kB FFT), single channel memory

(milliseconds/iteration)
Core 0 Core 1 Core 2 Core 3
117.2  116.4  116.0  116.0
 98.6   98.0   83.1    -
 80.6   80.5    -      -
 75.9    -     75.8    -
 70.8    -      -      -

Note that is is faster to run two exponents on cores 0 and 2 than on 0 and 1.

[rx7350]

Thanks for the benchmarks PATRIK. Obviuously, running four LL tests shows just how severe the memory bottleneck is. For comparison, an Intel dual-core at 2.4Ghz testing two 2048K FFT exponents has iteration times of about 47 milliseconds.

If you get a chance, post some additional benckmarks when you get your 'fast' memory.

[Dresdenboy]

Well done, patrik!

Originally Posted by patrik

Note that is is faster to run two exponents on cores 0 and 2 than on 0 and 1.

This is because the core pairs 0/1 and 2/3 share one L2 each.

[Kevin]

Doing a little bit of experimenting right now. I have a Quad Core Q6600 running on stock settings, and (2x1GB) PC8500 Corsair XMS memory@1066mhz. I started off running 40M first time tests on cores 0 and 2, and then did factoring on cores 1 and 3. My timings for the 40M tests were about .001 seconds below what's listed on the benchmark page (it maxes out at 64 ms per iteration). I decided to see what the hit was if I switched the two factoring tests to double-checks. The iteration time for the 40M exponents has only dropped to 72 ms per iteration when I'm doing nothing else (double-checks are still on self-test, so I don't know what the timing hit is for them yet).

Does anybody know what the difference is in credit for the same amount of CPU time spent on LL testing as opposed to factoring? It'd be nice to know what kind of hit it would take to make running factoring on the second core more advantageous than two LL tests with the memory hit included.