Mac Pro memory performance (Stream)

[hehejames]

Re-writing the post...

In the Table below (attached), Bare Feats and OWC have reported the following results for the Copy, Scale, Add, and Triad functions of STREAM to produce a single speed rating in gigabytes per second.


Are you seeing compatible results when running Geekbench?

I have 16GB installed on a 3.2GHz 8-core. 8 x 2GB using all 8 slots, but only seeing 3.5GB/s

 

[UltraNEO*]

Hello folks -

I'm sure most of us have visited OWC either via recommendation or through advertisement/links and for those currently using a 2008 Mac Pro (8-core); the chart below should be familiar. They provide a separate chart for Mac Pro quad/single cpu....

spam spam spam spam...

Technically speaking, this thread could be considered as spam too!

 

[GoCubsGo]

Technically speaking, this thread could be considered as spam too!

Agreed. This thread could also be considered part of the FAQ which should carry the title "when it is appropriate to quote an entire post?"

 

[hehejames]

spam spam spam spam...

Did you actually read the post? I'm just sharing my personal experience of comparison OWC's test result/claim vs my own test runs using Geekbench. How is this spam? can you not read English?

 

[GoCubsGo]

hehejames in all fairness, you can research the true benchmarks of paired sets and such elsewhere before you buy.

Some of this would also qualify as common knowledge. Depending on which MP you have the way memory is used could vary. I'd imagine.

 

[UltraNEO*]

Did you actually read the post? I'm just sharing my personal experience of comparison OWC's test result/claim vs my own test runs using Geekbench. How is this spam? can you not read English?

Yeah I read it.. Often if you examine some of the other sites, they'll often say their product is superior to the competitors, from my experience it's usually American companies who'll embellish it's products. Personally, I think most shoppers would like to see third-party comparisons tests, it's fairer! Wouldn't you say?

 

[hehejames]

hehejames in all fairness, you can research the true benchmarks of paired sets and such elsewhere before you buy.

You are correct. All system-wide benchmarks using GeekBench submitted by others who own a compatible machine MP 3.2GHz 8-core with 16GB of RAM are seeing similar results. Max for memory stream being 3.5GB/s NOT 7GB/s+++ as stated/mentioned by OWC.

hehejames Some of this would also qualify as common knowledge. Depending on which MP you have the way memory is used could vary. I'd imagine.

It could be... but it wasn't for me and therefore making a post to discuss with the experts/those interested. Also, since you are asking which MP I have/tested; based on your response, you have not read my post, but decided to ask reply... huh?

 

[Genghis Khan]

@ hehejames

dude, you need to chillax

hopefully you are aware of the latency issue (which it seems you are as that is what is in your O.P.). therefore you should be aware that to get the highest bandwidth out of your memory you want to a) have as many slots filled as possible in b) identical pairs

also, check out barefeats article on the harpertown memory issues
http://www.barefeats.com/harper3.html


it seems your issue may be with that specific ram, or it may not be installed properly...recheck and then complain to OWC if you have an issue...

 

[hehejames]

hopefully you are aware of the latency issue (which it seems you are as that is what is in your O.P.). therefore you should be aware that to get the highest bandwidth out of your memory you want to a) have as many slots filled as possible in b) identical pairs

As mentioned on the post. I have installed 8 x 2GB for a total of 16GB. Yes. all 8 slots are properly installed and all 8 modules are identical. All from OWC.

also, check out barefeats article on the harpertown memory issues
http://www.barefeats.com/harper3.html

Yeah... the same benchmark performed from barefeats clearly states fill up all 8 slots and expect to see 7.5GB/s average speed and I'm not seeing it.

And... because I'm not seeing those results... I'm asking the community if they are seeing 7GB/s+++ or not. A VERY SIMPLE question.

it seems your issue may be with that specific ram, or it may not be installed properly...

- If the memory was not installed properly; then I wouldn't be able to see 16GB from system profiler.
- System diag. & profiler clearly reports all 8 RAM's are detected and okay running at 800MHz.

recheck and then complain to OWC if you have an issue...

Yeah. That's what I'm trying to do, but I need additional facts.

As clearly stated on my post... when looking at other results from Geekbench from those who have a compatible machine MP 3.2GHz with 16GB; their results are in line with mine. Memory stream test hitting the wall at 3.5GB/s not 7GB/s+++


Maybe... Geekbench is simply reporting the wrong data. Who knows, but that's why I'm asking others in the forum.

 

[OWCMichael]

We have worked very hard over the past 20+ years to create a name that's synonymous with trust in the Mac community.

The test used was a terminal-based, 64-bit, parallel multi-threaded test that was provided to us from Rob-ART at BareFeats who obtained the same results that we did.

http://www.barefeats.com/harper3.html

Basically the test is not bound by the 3.5GB allocation restrictions of a 32-bit OS like 10.5 - This is why your results using a 32 bit version of Geekbench would vary.

Other World Computing has never and will never intentionally mislead our customers or post incorrect information to dupe anyone into making an uninformed purchase from us.

 

[SolidGun]

We have worked very hard over the past 20+ years to create a name that's synonymous with trust in the Mac community.

The test used was a terminal-based, 64-bit, parallel multi-threaded test that was provided to us from Rob-ART at BareFeats who obtained the same results that we did.

http://www.barefeats.com/harper3.html

Basically the test is not bound by the 3.5GB allocation restrictions of a 32-bit OS like 10.5 - This is why your results using a 32 bit version of Geekbench would vary.

Other World Computing has never and will never intentionally mislead our customers or post incorrect information to dupe anyone into making an uninformed purchase from us.

I really like this response. I have known about OWC since my first Powerbook purchase and I think I will make another purchase this week as I just bought a Mac Pro. It is this type of community feedback that keeps return customers.

 

[hehejames]

We have worked very hard over the past 20+ years to create a name that's synonymous with trust in the Mac community.

Thanks for the response. I trust this community and that was the reason for my posting.

The test used was a terminal-based, 64-bit, parallel multi-threaded test that was provided to us from Rob-ART at BareFeats who obtained the same results that we did.

OWC's web site does not have any reference to the test being performed at 64-bit nor using a custom test app. It would be very nice for consumer to be able to run the same test.

Also, when providing a link to the "STREAM benchmark reference information" or the test tool that may have been used is from the mid/late 90's it doesn't look too promising.

Basically the test is not bound by the 3.5GB allocation restrictions of a 32-bit OS like 10.5 - This is why your results using a 32 bit version of Geekbench would vary.

As clearly stated on the original post; I ran both 32bit and 64bit version of GeekBench. Yes, 64bit reported a higher BW on the Stream test, but it was a mere 10% increase over the 32bit. Based on my test, the system was unable to go over the 4GB/s when tested with GeekBench 64bit.

Other World Computing has never and will never intentionally mislead our customers or post incorrect information to dupe anyone into making an uninformed purchase from us.

Very good. Can you now please update your web site? Would love to see an actual instruction for customers to reproduce the performance test or have results from GeekBench or other standardized testing tool.

 

[barefeats]

Re-writing the post...

In the Table below (attached), Bare Feats and OWC have reported the following results for the Copy, Scale, Add, and Triad functions of STREAM to produce a single speed rating in gigabytes per second.

Are you seeing compatible results when running Geekbench?

I have 16GB installed on a 3.2GHz 8-core. 8 x 2GB using all 8 slots, but only seeing 3.5GB/s

As stated in the third paragraph of the original article, we at Bare Feats Lab were testing with a "special 64-bit parallel multi-threaded version of STREAM." Geekbench does not a use a multi-threaded version of STREAM. Our custom STREAM was created for us by the author of Geekbench to squeeze out the maximum throughput of the memory bus.

Here's the link to the original article:
http://barefeats.com/harper3.html

 

[hehejames]

As stated in the third paragraph of the original article, we at Bare Feats Lab were testing with a "special 64-bit parallel multi-threaded version of STREAM." Geekbench does not a use a multi-threaded version of STREAM. Our custom STREAM was created for us by the author of Geekbench to squeeze out the maximum throughput of the memory bus.

Here's the link to the original article:
http://barefeats.com/harper3.html

Yes... Your report/summary does clearly state such information, but not on the OWC's site!

 

[OWCMichael]

Fair enough. Despite having links to the test for more information on the test itself, we have added the same wording to our test site to help eliminate any possibility of misunderstanding.

 

[Andrew Henry]

Fair enough. Despite having links to the test for more information on the test itself, we have added the same wording to our test site to help eliminate any possibility of misunderstanding.

Glad to see you can admit you're wrong. Thanks for clearing that up for everyone who might be misled.

 

[Tesselator]

Yeah... the same benchmark performed from barefeats clearly states fill up all 8 slots and expect to see 7.5GB/s average speed and I'm not seeing it.

And... because I'm not seeing those results... I'm asking the community if they are seeing 7GB/s+++ or not. A VERY SIMPLE question.

Yes, I get 8 gigabit or 1 gigabyte speeds!

http://www.haxial.com/download/Benchmark1000-Mac.sit

Benchmark:         CPU & RAM
Version:           1.000
Date/Time:         24 June 2009, 01:34:29 AM
Operating System:  MacOS 10.57

Total Time:        18,820 millisecs (0:18)
Average Speed:     831.6M bytes/second

                     Time   Speed
Memory Copy:        1,005   1.0G/sec
Memory Fill:        2,766   1.6G/sec
Memory Clear:       2,697   1.7G/sec
Memory Equal:       1,094   1023.7M/sec
CRC:                3,208   199.5M/sec
Quick Hash:         1,139   561.8M/sec
Secure Hash:        1,535   208.4M/sec
Encryption:         5,376   29.7M/sec

Memory stream test hitting the wall at 3.5GB/s not 7GB/s+++

Maybe... Geekbench is simply reporting the wrong data. Who knows, but that's why I'm asking others in the forum.

Very good. Can you now please update your web site? Would love to see an actual instruction for customers to reproduce the performance test or have results from GeekBench or other standardized testing tool.

Geekbench? I wouldn't take anything it said seriously.

 

[VirtualRain]

Yikes! Isn't this memory performance a key justifier to upgrade to a Nehalem-based Mac Pro?

Nehalem's tri-channel 1066 DDR3 is three times faster and I'm assuming FB-DIMM latency is an added killer.

Here's a chart that I compiled some time ago comparing the bandwidth of various memory configurations and what would be FSB limited (orange/red) for pre-nehalem architectures... I think Harpertown has a 333MHz FSB which would be the limiting factor for that platform.

Link to more info here.

EDIT: Fixed link to chart

 

[Tesselator]

Those are the theoretical limits (specs) for the memory. It's theoretical in a system environment. Your own post in another thread shows it better IMO:



There's about a 40% increase with each added channel. If it were three times (300%), it would for sure be a "key justifier" tho still wouldn't be any form of justification for the price hike!

 

[VirtualRain]

Those are the theoretical limits (specs) for the memory. It's theoretical in a system environment. Your own post in another thread shows it better IMO:



There's about a 40% increase with each added channel. If it were three times (300%), it would for sure be a "key justifier" tho still wouldn't be any form of justification for the price hike!

That chart shows Nehalem scoring 22GB/s in tri-channel mode! Isn't that a lot more than 3x what some people are seeing here with Harpertown?

At any rate, in real-world benchmarks memory bandwidth seems to play a relatively small role.

 

[Tesselator]

That chart shows Nehalem scoring 22GB/s in tri-channel mode! Isn't that a lot more than 3x what some people are seeing here with Harpertown?

No.

I get 13.6Gb/s on my 2006 2.66.

At any rate, in real-world benchmarks memory bandwidth seems to play a relatively small role.

It's substantial for me - I know that.

 

[nanofrog]

At any rate, in real-world benchmarks memory bandwidth seems to play a relatively small role.

This would be dependant on usage. A lot of software can't utilize it, but some can. Those intended for server use would have a decent chance, as does some scientific applications and rendering packages.

 

[VirtualRain]

No.

I get 13.6Gb/s on my 2006 2.66.

GB or Gb? The table for Nehalem is in GB... 22GB/s.

The max data bandwidth of a 64 bit wide 1333MHz (333MHz quad-pumped) Harpertown bus is 10.6GB/s. So Nehalem's max theoretical memory bandwidth is over twice that of Harpertown. I don't know what impact fully buffered memory has on performance, but I know it ain't good!

It's substantial for me - I know that.

I can't argue with you there, but with such large cache (especially on Harpertown) you are fairly unique. In fact, if memory speed is that critical to you, you really would benefit from Nehalem more than the average guy.

 

[Tesselator]

GB or Gb? The table for Nehalem is in GB... 22GB/s.

The max data bandwidth of a 64 bit wide 1333MHz (333MHz quad-pumped) Harpertown bus is 10.6GB/s. So Nehalem's max theoretical memory bandwidth is over twice that of Harpertown.

Surely they are mixing the two up? You even say here that it's only twice the theoretical limit. For them NOT to have made an MB/Mb error would mean they're results indicated nearly 20X performance. And I'm almost sure the memory bandwidth is not twenty times greater. I guess I could be wrong tho. Oops, let's hope that never happens.

I don't know what impact fully buffered memory has on performance, but I know it ain't good!

It ain't bad. There is a performance penalty for using buffered memory tho. It introduces latency to the memory request and requires more power consumption due to the buffer chips. If I remember right each read or write is held between the memory bus and the DRAM for one cycle. So it's one clock cycle slower than unbufered DRAM - and only on the 1st cycle of a burst for SDRAM.

I can't argue with you there, but with such large cache (especially on Harpertown) you are fairly unique. In fact, if memory speed is that critical to you, you really would benefit from Nehalem more than the average guy.

I hope not twenty times. LOL

 

[nanofrog]

It ain't bad... If I remember right each read or write is held between the memory bus and the DRAM for one cycle. So it's one clock cycle slower than unbufered DRAM - and only on the 1st cycle of a burst for SDRAM.

Yup.