Benchmarks for i7 965

lavalamp · 2008-12-06, 03:33

Hey all, I'm posting from an i7 965 system, if anyone wants any specific benchmarks run on this chip just post in the thread and I'll run them. If you would also like comparisons between hyperthreading enabled/disabled and 1/2/3/4 cores enabled then I can do that too by making changes in the BIOS.

I only have Vista SP1 64 bit installed so far, but I've downloaded a Mandriva ISO so I may have Linux on here soon.

The full system spec is:

Gigabyte GA-EX58-Extreme motherboard
i7 965 @ stock 3.2 GHz (for now), HT enabled
3*2GB 1600 MHz 8-8-8-24 OCZ triple channel
HiS Radeon HD 4870 X2 2GB GDDR5 @ all stock frequencies (for now)

For the very near future, I also have access to an i7 920 with similar RAM (3*2GB @ 1333 MHz 9-9-9-20), EX58-UD5 mobo and same graphics card, also running Vista SP1 64 bit.

Prime95 · 2008-12-06, 03:47

Make sure you use v25.8

Freightyard · 2008-12-06, 05:17

What's the reason not to use 25.7? Even though I will use the 25.8 Windows service version.

Cool! I will be replacing my machine next week with an i7 965:

- i7 965 3.2 GHz (still waiting for anandtech's Nehelem overclocking article)
- 6 GB Corsair 1600 MHz 8-8-8-24
- nVidia 280 GTX
- RAID1 Velociraptors

Hoping to get 4.0 GHz moderately O.C.'d with performance heat sinks and fans.

Prime95 · 2008-12-06, 14:33

Quote:

Originally Posted by Freightyard

What's the reason not to use 25.7?

Fixes to the Affinity code. 25.7 misidentified the hyperthreaded CPUs.

lavalamp · 2008-12-06, 18:51

There was no 64 bit version of 25.8 that I could see, so I guess I used the 32 bit version.

The benchmark was run at 135*24 = 3240 MHz, QuickPath running at 6480 MT/s and the RAM running at 1620 MHz.

I'll run it again once I've got a suitable overclock.

Any other benchmarks people are interested in?

petrw1 · 2008-12-06, 19:32

Quote:

Originally Posted by lavalamp

There was no 64 bit version of 25.8 that I could see, so I guess I used the 32 bit version.

The benchmark was run at 135*24 = 3240 MHz, QuickPath running at 6480 MT/s and the RAM running at 1620 MHz.

I'll run it again once I've got a suitable overclock.

Any other benchmarks people are interested in?

LL about 15 % faster than my Q9550 with 1 or 2 CPUs.
Much better with 3 or 4 CPUs but I don't have 25.8 yet.

Slightly slower on TF but I have the 64 bit version of 25.7.

fivemack · 2008-12-06, 22:59

I'd be very interested in the results (with hyperthreading enabled) from downloading the (650MB) file

http://www.chiark.greenend.org.uk/~twomack/i7test.zip

decompressing it and running

msieve.exe -v -nc -t 4

and

msieve.exe -v -nc -t 8

(if you have time, -t 1 and -t 2 would also be interesting, but I'd expect those to take ~8 hours and ~4 hours respectively). Once you've run the commands, put the msieve.log file up here.

This is a test of GNFS post-processing on a 120-digit number; with -t 4 it takes two hours on a Q6600, if it takes 45 minutes with -t 8 on an i7 then I'll go and buy an i7 next month :)

lavalamp · 2008-12-07, 03:34

I have two lots of -t 8 results from that test now, and I'm currently running -t 2 on a 2.7 GHz i7 920. I'd be running one on this PC but I'm using it and that may throw off the numbers.

I know basically nothing about what this program is and what it does, but what I can say is that at the beginning and at the end of the run, it only runs 1 thread. It only seems to break out the extra threads while running linear algebra. I guess that section is easily broken up into jobs that can be shared amongst threads and the others aren't.

I've attached the results for 8 threads, one thing that should be taken into account when looking at the times is that in the "square root phase" at the end, the slower 2.7 GHz CPU had to do two more blocks of work at 10 mins each, and those blocks are all single threaded.

Perhaps a more efficient method for using this program would be to run 8 instances, each bound to a single thread and core. Unless, maybe, you want to factor a number really quickly and/or the start and end processes are only a small fraction of the work involved for factoring the numbers of interest.

I've included a couple of screenshots and uploaded a zip here.

Incidentally, I replied to your other thread in the Hardware forum a few days ago.

fivemack · 2008-12-07, 11:05

Thanks very much for these runs. This is a very small test case (because a larger test case would need a much larger zip file); a 'real' run might take twenty hours in the start and end stages together and eight hundred hours on four cores for the chunk in the middle, while using 6GB of memory so it's a bit impractical to run more than one real run on a computer at a time.

On a Q6600, the middle chunk takes 95 minutes; on the i7/3.2 it took fifty and on the i7/2.7 it took 55, both with -t 8. How do the CPU-usage curves for the hyperthreaded run compare with hyperthreaded runs of prime95?

I would be very interested to see the timing and the CPU-usage plots for a -t 4 run on either machine. I see that the cache parameters have been tuned for the L2 rather than the L3 size, so there may be some more performance to eke out there.

When I posted the previous thread I'd forgotten that msieve existed on Windows, and so I could prepare a .zip file and instructions rather than needing a login on someone else's machine. Thanks very much for your help in running this.

lavalamp · 2008-12-07, 18:39

Quote:

Originally Posted by fivemack

(if you have time, -t 1 and -t 2 would also be interesting, but I'd expect those to take ~8 hours and ~4 hours respectively). Once you've run the commands, put the msieve.log file up here.

Well, I ran them, (along with -t 4), and they took nowhere near that long.

-t 2 took about 1.5 hours, and -t 1 took about 2 hours.

It's a bit of a weird situation for -t 4, it was quicker than -t 8 and took only about an hour, compared to ~ 1.25 hours for -t 8.

Of course, you'll have to remove the start and end times to find the real values for the multi-threaded performance of linear algebra.

Anyway, I've uploaded everything here in another zip for you to see.

lavalamp · 2008-12-07, 21:03

Quote:

Originally Posted by fivemack

How do the CPU-usage curves for the hyperthreaded run compare with hyperthreaded runs of prime95?

When running a torture test with 8 threads, the usage is a solid 100% on all 8 graphs.

I don't know what it means that the msieve graphs are not all 100% and the Prime 95 ones are. Obviously I'm not getting a 100% performance increase by running Prime95 twice on one core, so the real utilisation is a lot less than what those graphs indicate for P95, though I don't know what it really is.

Maybe msieve has trouble breaking the work up into more threads? Or maybe the graphs really are representative of the CPU utilisation.

2008-12-06, 03:33	#1
lavalamp Oct 2007 Manchester, UK 1,373 Posts	Benchmarks for i7 965 Hey all, I'm posting from an i7 965 system, if anyone wants any specific benchmarks run on this chip just post in the thread and I'll run them. If you would also like comparisons between hyperthreading enabled/disabled and 1/2/3/4 cores enabled then I can do that too by making changes in the BIOS. I only have Vista SP1 64 bit installed so far, but I've downloaded a Mandriva ISO so I may have Linux on here soon. The full system spec is: Gigabyte GA-EX58-Extreme motherboard i7 965 @ stock 3.2 GHz (for now), HT enabled 32GB 1600 MHz 8-8-8-24 OCZ triple channel HiS Radeon HD 4870 X2 2GB GDDR5 @ all stock frequencies (for now) For the very near future, I also have access to an i7 920 with similar RAM (32GB @ 1333 MHz 9-9-9-20), EX58-UD5 mobo and same graphics card, also running Vista SP1 64 bit. Last fiddled with by lavalamp on 2008-12-06 at 03:34

2008-12-06, 05:17	#3
Freightyard Nov 2008 San Luis Obispo CA 2⁷ Posts	What's the reason not to use 25.7? Even though I will use the 25.8 Windows service version. Cool! I will be replacing my machine next week with an i7 965: - i7 965 3.2 GHz (still waiting for anandtech's Nehelem overclocking article) - 6 GB Corsair 1600 MHz 8-8-8-24 - nVidia 280 GTX - RAID1 Velociraptors Hoping to get 4.0 GHz moderately O.C.'d with performance heat sinks and fans. Last fiddled with by Freightyard on 2008-12-06 at 05:18

2008-12-07, 03:34	#8
lavalamp Oct 2007 Manchester, UK 1,373 Posts	I have two lots of -t 8 results from that test now, and I'm currently running -t 2 on a 2.7 GHz i7 920. I'd be running one on this PC but I'm using it and that may throw off the numbers. I know basically nothing about what this program is and what it does, but what I can say is that at the beginning and at the end of the run, it only runs 1 thread. It only seems to break out the extra threads while running linear algebra. I guess that section is easily broken up into jobs that can be shared amongst threads and the others aren't. I've attached the results for 8 threads, one thing that should be taken into account when looking at the times is that in the "square root phase" at the end, the slower 2.7 GHz CPU had to do two more blocks of work at 10 mins each, and those blocks are all single threaded. Perhaps a more efficient method for using this program would be to run 8 instances, each bound to a single thread and core. Unless, maybe, you want to factor a number really quickly and/or the start and end processes are only a small fraction of the work involved for factoring the numbers of interest. I've included a couple of screenshots and uploaded a zip here. Incidentally, I replied to your other thread in the Hardware forum a few days ago. Last fiddled with by lavalamp on 2008-12-07 at 03:35

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Benchmarks	MurrayInfoSys	Information & Answers	3	2011-04-14 17:10
LLR benchmarks	Oddball	No Prime Left Behind	11	2010-08-06 21:39
benchmarks	Unregistered	Information & Answers	15	2009-08-18 16:44
Benchmarks for 24.12	Prime95	Software	60	2005-06-11 07:35
Benchmarks	Vandy	Hardware	6	2002-10-28 13:45

2008-12-06, 03:47	#2
Prime95 P90 years forever! Aug 2002 Yeehaw, FL 20076₈ Posts	Make sure you use v25.8

2008-12-06, 22:59	#7
fivemack (loop (#_fork)) Feb 2006 Cambridge, England 2·7·461 Posts	I'd be very interested in the results (with hyperthreading enabled) from downloading the (650MB) file http://www.chiark.greenend.org.uk/~twomack/i7test.zip decompressing it and running msieve.exe -v -nc -t 4 and msieve.exe -v -nc -t 8 (if you have time, -t 1 and -t 2 would also be interesting, but I'd expect those to take ~8 hours and ~4 hours respectively). Once you've run the commands, put the msieve.log file up here. This is a test of GNFS post-processing on a 120-digit number; with -t 4 it takes two hours on a Q6600, if it takes 45 minutes with -t 8 on an i7 then I'll go and buy an i7 next month :)

2008-12-07, 11:05	#9
fivemack (loop (#_fork)) Feb 2006 Cambridge, England 6454₁₀ Posts	Thanks very much for these runs. This is a very small test case (because a larger test case would need a much larger zip file); a 'real' run might take twenty hours in the start and end stages together and eight hundred hours on four cores for the chunk in the middle, while using 6GB of memory so it's a bit impractical to run more than one real run on a computer at a time. On a Q6600, the middle chunk takes 95 minutes; on the i7/3.2 it took fifty and on the i7/2.7 it took 55, both with -t 8. How do the CPU-usage curves for the hyperthreaded run compare with hyperthreaded runs of prime95? I would be very interested to see the timing and the CPU-usage plots for a -t 4 run on either machine. I see that the cache parameters have been tuned for the L2 rather than the L3 size, so there may be some more performance to eke out there. When I posted the previous thread I'd forgotten that msieve existed on Windows, and so I could prepare a .zip file and instructions rather than needing a login on someone else's machine. Thanks very much for your help in running this.