CFD Online Discussion Forums

CFD Online Discussion Forums (http://www.cfd-online.com/Forums/)
-   Hardware (http://www.cfd-online.com/Forums/hardware/)
-   -   Performance of dual xeon 2643 (http://www.cfd-online.com/Forums/hardware/110411-performance-dual-xeon-2643-a.html)

tally_ho December 10, 2012 18:31

Performance of dual xeon 2643
 
Hello. I am doing FEM simulations, and although they are not CFD but electromagnetics, I hope that you can still help me. I recently got a HP Z820 with dual Xeon 2643 @ 3.3 GHz and 32 GB ram. However, in the FEM simulations I do (using Ansoft software), I notice almost no difference between this machine and my previous machine, which was an Intel core i5-2400 @ 3.1 GHz. I would expect a dual 2643 to be a bit faster!

I have one idea about what could be wrong: The memory is installed as 4x8 GB, but HP recommends that you use 8x4 GB to take advantage of the Xeon's quad channels. Could this be the issue, or is there anything else ?

tally_ho December 11, 2012 04:49

Here is some data I got from PerformanceTest to compare the two. The 2643 is best in the CPU Mark, but the i5 wins in Memory Mark


2643
Memory Mark: 2174
Databse Operations: 65.2
Read Cached: 24874
Read Uncached: 10876
Write: 8452
Latency: 36.1
Threaded: 24057

i5-2400
Memory Mark: 2240
Databse Operations: 76.9
Read Cached: 23899
Read Uncached: 12159
Write: 8482
Latency: 26.2
Threaded: 17825

2643
CPU Mark: 14093
Integer Math: 29677
Floating Point Math: 13772
Prime Numbers: 44.0
Extended Instructions: 83.7
Compression: 22582
Encryption: 3163
Physics: 768
Sorting: 13674
Single Threaded 1869

i5-2400
CPU Mark: 6244
Integer Math: 8308
Floating Point Math: 6593
Prime Numbers: 24.5
Extended Instructions: 22.5
Compression: 7784
Encryption: 1036
Physics: 391.4
Sorting: 4865
Single Threaded 1805

evcelica December 11, 2012 18:07

Are you using the same number of cores comparing the two? If so then I would expect them to be about the same. You will see the dual cpu machine pull away from the i5 as you increase the number of cores. That 14000 score is for 2 XEONS, they are only 7000 each, and if you are not using all the cores, you won't see much increase over the i5.
Also you are only using half the memory bandwidth if you are only populating two of the four channels for each CPU, I would definitely populate all of them. Can you get another identical 4x8GB kit so you can run 8x8GB?

tally_ho December 12, 2012 07:04

Quote:

Originally Posted by evcelica (Post 397031)
Are you using the same number of cores comparing the two? If so then I would expect them to be about the same. You will see the dual cpu machine pull away from the i5 as you increase the number of cores. That 14000 score is for 2 XEONS, they are only 7000 each, and if you are not using all the cores, you won't see much increase over the i5.
Also you are only using half the memory bandwidth if you are only populating two of the four channels for each CPU, I would definitely populate all of them. Can you get another identical 4x8GB kit so you can run 8x8GB?


Actually I was trying to use all the cores on the xeon, but I couldn't. It would not use more than 4 cores even when running multiple instances of my software. Maybe the performance is limited by the memory bandwidth?

I will try to get another 4x8GB kit. The memory currently installed in my computer is the following. Is it any good?
http://www.memory4less.com/m4l_itemd...T31GR7CFR4C-PB

What about just replacing the existing 32GB by 8x4GB non ECC ?

evcelica December 12, 2012 12:57

That is very expensive server memory, CL11 is a little slow, CL9 is pretty much standard. I didn't know you had expensive server RAM, now I see why you want to just get an entirely new kit of non-ECC memory.
You can get non-ECC 8x4GB or an 8x8GB kit for under $300. I don't see why new RAM wouldn't work if you get a 1.5V 1600MHz kit, though RAM can be finicky sometimes so I can't say it definitely will or not.

tally_ho December 12, 2012 13:39

Quote:

Originally Posted by evcelica (Post 397188)
That is very expensive server memory, CL11 is a little slow, CL9 is pretty much standard. I didn't know you had expensive server RAM, now I see why you want to just get an entirely new kit of non-ECC memory.
You can get non-ECC 8x4GB or an 8x8GB kit for under $300. I don't see why new RAM wouldn't work if you get a 1.5V 1600MHz kit, though RAM can be finicky sometimes so I can't say it definitely will or not.

Okay thanks. I think I'll just get a new 8x4 kit and see how that works out.

tally_ho December 17, 2012 10:36

Hello again. I am a bit confused about what memory my system supports. If it is built to use ECC ram, can I then use non-ECC ram ? I mean, doesn't ECC ram have more pins than non-ECC ? What about registered or unbuffered ?

I found this PDF
http://h20331.www2.hp.com/Hpsub/down...igurations.pdf

From this I understand that it can either use unbuffered ECC or registered non ecc. Is that correct?

evcelica December 17, 2012 13:01

If memory is (registered = buffered), it is by default ECC.
Unregistered/unbuffered memory comes in both ECC and Non-ECC, though the latter is very rare.
At first glance I would say you are stuck having to use ECC RAM, but thinking about it more, maybe not, that sheet is just saying it does support, not what it doesn't. You could contact HP or do as that sheet says and "See the Memory Technology White Paper for additional technical information"

Really the smartest thing to do would be to just get 4 more identical modules like you already have, that way you will have 64GB of ECC RAM (8x8), and you know it will be compatible. That was also was my suggestion in my first post. You may be able to get another, but cheaper, 4x8GB kit like this: http://www.newegg.com/Product/Produc...82E16820239235
to add to your existing modules, looks identical to what you already have as far as I can tell.


All times are GMT -4. The time now is 17:41.