Is Quad Channel Memory Slow Or Is It OSX ???

Question

Level 1

70 points

Is Quad Channel Memory Slow Or Is It OSX ???

Booting up Windows XP and running HotCPU gives me these
memory benchmark results. That's a heck of a lot different
than the Mac OSX benchmarking utils give me. I'm using a
Quad-channel memory configuration which is supposed to yield
about 21Gigs/sec and these results look pretty close to that.

000042| 0010 000000c0 Nov, 17 - 02:09:24 | Memory memcpy Method : 13913 MB/s
000043| 0010 000000c0 Nov, 17 - 02:09:25 | Memory C++ Method : 17842 MB/s
000044| 0010 000000c0 Nov, 17 - 02:09:25 | Memory repmovsd Method : 13034 MB/s
000045| 0010 000000c0 Nov, 17 - 02:09:26 | Memory FPU 64-bit Method : 12307 MB/s
000046| 0010 000000c0 Nov, 17 - 02:09:27 | Memory Assembly 32-bit Method : 20857 MB/s
000047| 0010 000000c0 Nov, 17 - 02:09:28 | Memory Assembly 64-bit Method : 13913 MB/s
000048| 0010 000000c0 Nov, 17 - 02:09:29 | Memory MMX 32-bit Method : 11428 MB/s
000049| 0010 000000c0 Nov, 17 - 02:09:29 | Memory MMX 128-bit Method : 13333 MB/s
000050| 0010 000000c0 Nov, 17 - 02:09:30 | Memory SSE 128-bit Method : 13333 MB/s
000051| 0010 000000c0 Nov, 17 - 02:09:31 | Memory SSE movntq 64-bit Method : 16000 MB/s
000052| 0010 000000c0 Nov, 17 - 02:09:32 | Memory SSE movntq 128-bit Method : 20000 MB/s
000053| 0010 000000c0 Nov, 17 - 02:09:34 | Memory SSE prefetch movntq 128-bit Method : 20000 MB/s
000054| 0010 000000c0 Nov, 17 - 02:09:35 | Memory SSE movntps 128-bit Method : 20000 MB/s
000055| 0010 000000c0 Nov, 17 - 02:09:36 | Memory SSE prefetch movntps 256-bit Method : 16842 MB/s

So my question is. Is it just OS X that can't take advantage
of the hardware configuration or is something else going on
here?

This test was performed in Parallels WinXP Corperate Eddition.
With this utility: http://www.softpedia.com/progDownload/Hot-CPU-Tester-Pro-Download-5086.html#down load_locations

MacPro, Mac OS X (10.4.8), 2x2x2.66 | 2gig | LW3D

Posted on Nov 16, 2006 10:12 AM

Reply

Answer 1

Nov 16, 2006 1:46 PM in response to Tesselator

Well my hotcpu results are much worse than yours, this is a 4x1GB machine:

000045| 0010 000001ec Nov, 16 - 21:39:43 | Memory memcpy Method : 3018 MB/s

000046| 0010 000001ec Nov, 16 - 21:39:44 | Memory C++ Method : 2807 MB/s

000047| 0010 000001ec Nov, 16 - 21:39:44 | Memory repmovsd Method : 2990 MB/s

000048| 0010 000001ec Nov, 16 - 21:39:45 | Memory FPU 64-bit Method : 2807 MB/s

000049| 0010 000001ec Nov, 16 - 21:39:45 | Memory Assembly 32-bit Method : 2735 MB/s

000050| 0010 000001ec Nov, 16 - 21:39:46 | Memory Assembly 64-bit Method : 2758 MB/s

000051| 0010 000001ec Nov, 16 - 21:39:47 | Memory MMX 32-bit Method : 2758 MB/s

000052| 0010 000001ec Nov, 16 - 21:39:47 | Memory MMX 128-bit Method : 2831 MB/s

000053| 0010 000001ec Nov, 16 - 21:39:48 | Memory SSE 128-bit Method : 2882 MB/s

000054| 0010 000001ec Nov, 16 - 21:39:49 | Memory SSE movntq 64-bit Method : 3902 MB/s

000055| 0010 000001ec Nov, 16 - 21:39:49 | Memory SSE movntq 128-bit Method : 3902 MB/s

000056| 0010 000001ec Nov, 16 - 21:39:49 | Memory SSE prefetch movntq 128-bit Method : 3950 MB/s

000057| 0010 000001ec Nov, 16 - 21:39:50 | Memory SSE movntps 128-bit Method : 3855 MB/s

000058| 0010 000001ec Nov, 16 - 21:39:50 | Memory SSE prefetch movntps 256-bit Method : 3902 MB/s

000059| 0010 000001ec Nov, 16 - 21:39:51 | Memory SSE2 movapd 128-bit Method : 0 MB/s

Running SiSoftware sandra I get 5800MB/sec for memory bandwidth.

For combined cache and memory the memory sems to fly:

32KB blocks = 362300 MB/s
1MB blocks = 100234 MB/s
16MB Blocks = 10108 MB/s

This is on x32.

Cheers

Andy

Reply

Answer 2

Nov 16, 2006 2:07 PM in response to Tesselator

Well I have just had a look around on the net and your figures are either a windup, you have a magic MacPro or the figures are inacurate as you are running parallels.

Reply

Answer 3

Tesselator Author

Level 1

70 points

Nov 16, 2006 4:29 PM in response to AndrewCapon

Well I have just had a look around on the net and
your figures are either a windup, you have a magic
MacPro or the figures are inacurate as you are
running parallels.

What's a windup?

I guess they could be inacurate but I dunno why they would be.

I tried it first with the default 16mb and then I upped it
to 64MB and then to 128MB. I have run the test about 5 or
6 times and each time tho it varies, I get very similar
results.

What/where did you look on the internet to make you think
that the MacPro being used is smoking crack? Educate me
bro!

The Hot-CPU program only sees a 16MB L1 cache and it thinks
that there is a single core 5150 at 2.66. So I guess there
could be something funky going on. Still, show me what you
found. Please. 😀

Reply

Answer 4

Nov 16, 2006 11:13 PM in response to Tesselator

Hi,

Windup is an English term for "pulling our legs", which looking at your reply you obviously were not.

The peak actual bandwidth mentioned in lots of places is arround 5.5GB/s as it is on my machine and the max theoretical bandwidth is 8 GB/s.

A good explanation can be found at http://www.anandtech.com/mac/showdoc.aspx?i=2811&p=4 make sure to look at the next page of the article as well.

Hope this helps

Andy

P.S. Incase the article depresses you, there are a couple of good advantages of this memory system.

1. Memory is bi-directional and can read and write to memory at the same time.

2. Hass ECC and CRC so is very reliable

Reply

Answer 5

Tesselator Author

Level 1

70 points

Nov 16, 2006 11:34 PM in response to AndrewCapon

Hi,

Windup is an English term for "pulling our legs",
which looking at your reply you obviously were not.

K, I kinda thought it meant rigged or something but wasn't
sure.

The peak actual bandwidth mentioned in lots of places
is arround 5.5GB/s as it is on my machine and the max
theoretical bandwidth is 8 GB/s.

A good explanation can be found at
http://www.anandtech.com/mac/showdoc.aspx?i=2811&p=4
make sure to look at the next page of the article as
well.

Hope this helps

Yeah this is the second time I read that article. 😉 I got
a little more out of it this time. Still there are other
articles that repeat the 21gig/sec theoretical spec and
then ponder the reasons why they can't get that. So, that's
why i was wondering as well. When I saw the results of the
Hot-CPU tests I went ah-ha! It's something is OS X (maybe). :P

I wonder what's going on that I get those results? Very
strange. Yeah, even "magic" if they're correct. 😉

Andy

P.S. Incase the article depresses you, there are a
couple of good advantages of this memory system.

1. Memory is bi-directional and can read and write to
memory at the same time.

2. Hass ECC and CRC so is very reliable

Yeah, those are good things. I want the 21gig/sec. tho 😀
Or even 12 like the Opron procs get. 4gig to 5gig is kinda
on the slow side.

Thanks for the explaination man!
Much appreciated!

Reply

Answer 6

Baryo

Level 1

55 points

Nov 16, 2006 11:54 PM in response to Tesselator

Let me get this straight - you got those numbers under Parallels???
Not BootCamp.

Reply

Answer 7

Tesselator Author

Level 1

70 points

Nov 17, 2006 11:14 PM in response to Baryo

Yes sir.

Parallels Build: 1970
3x300gig Drive MaxLine III RAID 0
10.8.4 Xeon 2.66 5150
4x512MB ram
Allocated 1024MB to Parallels WinXP - Corperate eddition.

Reply

Answer 8

Tesselator Author

Level 1

70 points

Nov 18, 2006 2:03 AM in response to Tesselator

http://download.visualware.com/pub/cpuinfo/cpuinfo.exe
Says I have a 11603.5 MHz processor but memory is only
"8582 MB/sec sustained transfer rate"
It also says: L1: 32K (8-way set associative, 64byte line size)
whatever that means. :P

http://www.benchmarkhq.ru/fclick/fclick.php?fid=290
This says I get between 15 and 16 GB/sec too. If you try
this one be sure to give it a large enough amount of ram to
deal with - the default is a little small. I tried it up to
256MB...

Hmmm, these can't ALL be wrong can they? OK, what's going
on? Who passed the crack pipe to my Mac? >:-[

Now I'm starting to think that OS X does have something to
do with it after all.

Reply

Answer 9

Nov 19, 2006 8:34 AM in response to Tesselator

I woyuld say it is a parallels problem, probably the high performance counter is wrong, this would cause inacurate benchmarks.

If you run these benchmarks when booting from XP they give much more realistic figures.

Cheers

Andy

Reply

Answer 10

Baryo

Level 1

55 points

Nov 19, 2006 10:40 AM in response to Tesselator

Hi,
I do some Physics with my Mac Pro basically terminal based wicked fast Fortran numeric calculations.
You could say my life's work would look much shorter if any configuration of the Mac Pro would have gone up to 16GB/Sec.
on me Mac Pro (16GB RAM) I never got anything over an effective 6.8GB/Sec and this is over the best Optimized Algorithm I (and the really talented guy that wrote the ****** thing) can think of.

It's a utility Glitch!!
And if it isn't call me I have a job for you!

Reply

Answer 11

Tesselator Author

Level 1

70 points

Nov 19, 2006 3:23 PM in response to Baryo

Hi,
I do some Physics with my Mac Pro basically terminal
based wicked fast Fortran numeric calculations.
You could say my life's work would look much shorter
if any configuration of the Mac Pro would have gone
up to 16GB/Sec.
on me Mac Pro (16GB RAM) I never got anything over an
effective 6.8GB/Sec and this is over the best
Optimized Algorithm I (and the really talented guy
that wrote the ****** thing) can think of.

It's a utility Glitch!!
And if it isn't call me I have a job for you!

Can I have the job even if it isn't true? I know Fortran! 😀

OK... So it's probably a util glitch. Dang. K, I'm going
to try some stopwatch tests and see. There's enough of a
difference between 7GB/sec and 20GB/sec that it oughtta
show up in some simple stop-watch tests.

Be back with that another day.

Reply