Dedicated Prime95 rig

[hj47]

Hey peoples,

Since AMD has released its new dual core 7550 and 7750 I've been thinking of creating a dedicated prime rig.

Would these be relatively efficient CPUs for prime, and is 1 gig of memory enough? Or should I get 2?

Additionally, if they are not reasonably priced here in Australia, would an AMD triple core be significantly more efficient than a 7750 or 7550?

& is there a preference to using linux or windows for prime, because for folding I know that linux is terribly faster for SMP folding than windows is.

Cheers :D

[henryzz]

Originally Posted by hj47

Hey peoples,

Since AMD has released its new dual core 7550 and 7750 I've been thinking of creating a dedicated prime rig.

Would these be relatively efficient CPUs for prime, and is 1 gig of memory enough? Or should I get 2?

Additionally, if they are not reasonably priced here in Australia, would an AMD triple core be significantly more efficient than a 7750 or 7550?

& is there a preference to using linux or windows for prime, because for folding I know that linux is terribly faster for SMP folding than windows is.

Cheers :D

if you are planning on doing ll tests dont get an AMD
an intel runs ~twice as fast on ll tests

[hj47]

Why is that?

[henryzz]

Originally Posted by hj47

Why is that?

i dont know why it is but it is a well known fact on this forum
when i upgraded from a 2.4 GHz Athlon 64 to a 2.4 Ghz Core 2 my speed at ll testing doubled per core yet for some other programs my Athlon 64 would have been faster per core

[ET_]

Originally Posted by henryzz

i dont know why it is but it is a well known fact on this forum
when i upgraded from a 2.4 GHz Athlon 64 to a 2.4 Ghz Core 2 my speed at ll testing doubled per core yet for some other programs my Athlon 64 would have been faster per core

Quad-core Phenom doesn't work that bad after all...

Luigi

[fivemack]

Yes, the athlon64 split SSE2 operations into two chunks done in sequence, whilst the core2 and phenom do the whole operation in a single chunk. So a 7550, which is basically half a Phenom, should have reasonable performance.

Though I suspect the price/performance will be significantly better if you use a 'Phenom II' (45nm quad-core) when they come out - the CPU costs about twice as much as the dual-core, but you don't have to have a second case, motherboard and memory.

[S00113]

Originally Posted by henryzz

if you are planning on doing ll tests dont get an AMD
an intel runs ~twice as fast on ll tests

And while we are at it -- trial factoring to limits > 60 bits runs ~ twice as fast on AMD64.

Another wirdness I've found is that AMD64 is faster at 5120K and 6144K FFT than 4096K FFT when multithreading. At 5 threads 6144K FFT is almost as fast as 2048K! This makes me wonder if it is properly optimized for multi core multi CPU AMD CPUs. Here is selected benchmarks from an oldish machine with eight dual core 1 GHz Opterons:

Dual-Core AMD Opteron(tm) Processor 8218
CPU speed: 1000.31 MHz, 16 cores
...
Timing FFTs using 2 threads.
Best time for 3072K FFT length: 122.214 ms.
Best time for 3584K FFT length: 141.534 ms.
Best time for 4096K FFT length: 162.453 ms.
Best time for 5120K FFT length: 148.117 ms.
Best time for 6144K FFT length: 192.784 ms.
Best time for 7168K FFT length: 243.260 ms.
Best time for 8192K FFT length: 296.969 ms.
...
Timing FFTs using 3 threads.
Best time for 3072K FFT length: 113.277 ms.
Best time for 3584K FFT length: 147.447 ms.
Best time for 4096K FFT length: 145.264 ms.
Best time for 5120K FFT length: 105.272 ms.
Best time for 6144K FFT length: 125.552 ms.
Best time for 7168K FFT length: 154.236 ms.
Best time for 8192K FFT length: 184.660 ms.
...
Timing FFTs using 4 threads.
Best time for 3072K FFT length: 140.230 ms.
Best time for 3584K FFT length: 157.839 ms.
Best time for 4096K FFT length: 177.005 ms.
Best time for 5120K FFT length: 139.861 ms.
Best time for 6144K FFT length: 116.143 ms.
Best time for 7168K FFT length: 155.983 ms.
Best time for 8192K FFT length: 209.981 ms.
...
Timing FFTs using 5 threads.
Best time for 2048K FFT length: 87.582 ms.
Best time for 2560K FFT length: 122.746 ms.
Best time for 3072K FFT length: 141.399 ms.
Best time for 3584K FFT length: 160.135 ms.
Best time for 4096K FFT length: 179.035 ms.
Best time for 5120K FFT length: 114.527 ms.
Best time for 6144K FFT length: 90.497 ms.
Best time for 7168K FFT length: 113.099 ms.
Best time for 8192K FFT length: 139.041 ms.
...

5 threads seem to bee the sweet spot. From 6 to 16 threads there is very little improvement, and even degradation for some FFT sizes.

[henryzz]

Originally Posted by fivemack

Yes, the athlon64 split SSE2 operations into two chunks done in sequence, whilst the core2 and phenom do the whole operation in a single chunk. So a 7550, which is basically half a Phenom, should have reasonable performance.

Though I suspect the price/performance will be significantly better if you use a 'Phenom II' (45nm quad-core) when they come out - the CPU costs about twice as much as the dual-core, but you don't have to have a second case, motherboard and memory.

i didnt realise that the problem was solved in phenoms
thanks for a good explanation

[hj47]

I have another question;

Would it be more productive/efficient if I were to use a higher clocked single core with more available cache than a dual core with smaller cache per core?

I'm thinking of either using the Celeron 430 overclocked to 2.66ghz versus, say, a Celeron E1200 clocked at 2.0ghz? They both have 512k cache, the E1200 has it split in half for each core.

+++ Also, is there any advantage in running a 64 bit OS as opposed to a 32 bit, and is there any performance gain/loss to using the linux version of prime to prime95?

[db597]

Originally Posted by hj47

I'm thinking of either using the Celeron 430 overclocked to 2.66ghz versus, say, a Celeron E1200 clocked at 2.0ghz? They both have 512k cache, the E1200 has it split in half for each core.

+++ Also, is there any advantage in running a 64 bit OS as opposed to a 32 bit, and is there any performance gain/loss to using the linux version of prime to prime95?

I'd go for the dual core. Cache does make some difference, but generally the output of a single 2.66GHz will still be less than having 2x2GHz. People often ask the same about dual vs quads too - there the line is more blurred because of memory bandwidth bottlenecks, but I believe the quads are still more efficient (run 3x LL and 1x TF).

The 64bit OS is faster at trial factoring. But for LL and the rest there's no difference. I found no speed advantages of using linux. But if this is a dedicated p95 rig, you might not want to have a monitor etc linked up to it. With linux the advantage is you can make do without the monitor/keyboard etc and simply ssh into the box from another PC.

[lavalamp]

Originally Posted by db597

But if this is a dedicated p95 rig, you might not want to have a monitor etc linked up to it. With linux the advantage is you can make do without the monitor/keyboard etc and simply ssh into the box from another PC.

No need for a monitor with Windows either, simply download UltraVNC (which is free), and you can easily access it visually through another PC on the network, or even over the internet if you set up port forwarding for it.

For overclocking in BIOS and such you would need a monitor though.

If you can stretch to it, an E8200 may perform significantly better than a Celeron due to a much larger cache (6 MB), and the overclocking headroom in the lower models is quite large. With a little patience and luck it seems you can add over another GHz to the stock clock.