MP 1,1-5,1 1333 MHz or 1066 MHz ?!?!

[amedias]

Thank You

looking at these as well Quality/Price point of view
https://www.ebay.com/itm/Samsung-M3...156046?hash=item340d2a014e:g:mRkAAOSwRwddC903

Samsung Ram is normally very reliable, and that's a good price. PC3L is the low voltage (1.35) version, it should work fine at 1.5V in a MP but I'm a bit wary about mixing 1.5V and 1.35V Dimms, I probably shouldn't be but I try to stick with one or the other.

I tend to stick with Micron, Samsung, or at a push Kingston or Hynix. They are big brands, who supply a lot of OEM and are normally well tested and reliable.

 

[frilli]

What CPU do you have? Exact model, not just speed.

X5675

 

[macjunkie2013]

I just upgraded my 2010 MP to 6x16GB 96GB and they run at 1333 MHz. My Mac Spec are in my Sig.

96GB (6 x 16GB) Samsung M393B2G70BH0-YH9 PC3L-10600R DDR3 1333MHz DIMM Memory for ~100 on eBay.


The same seller as another set listed here: https://www.ebay.com/itm/96GB-6-x-1...891785?hash=item4b66f52fc9:g:3QEAAOSwo6hc~5JE

 

[Dewdman42]

So on this topic, I have 16gb 2Rx4 PC3L-12800L-11-11-E2-D3 ECC. 96gb in my 5,1 with dual X5690 @ 3.46ghz

About this mac shows it as running at 1333mhz. Fine so far...

however when I use NovaBench to measure the RAM transfer speeds, it is showing ~8300 MB/sec, which I understand to be the transfer speed associated with 1060mhz...

Am i misunderstanding somehing or is this ram clocking down, even though MacOS reports it as being at 1333mhz?

In addition to that info, I have bootcamp also setup on this mac, under windows I can run Novabench and it shows the memory running even slower at ~7500MB/sec....and using CPU-Z, it reports the DRAM frequency at ~660mhz.

 

[h9826790]

So on this topic, I have 16gb 2Rx4 PC3L-12800L-11-11-E2-D3 ECC. 96gb in my 5,1 with dual X5690 @ 3.46ghz

About this mac shows it as running at 1333mhz. Fine so far...

however when I use NovaBench to measure the RAM transfer speeds, it is showing ~8300 MB/sec, which I understand to be the transfer speed associated with 1060mhz...

Am i misunderstanding somehing or is this ram clocking down, even though MacOS reports it as being at 1333mhz?

In addition to that info, I have bootcamp also setup on this mac, under windows I can run Novabench and it shows the memory running even slower at ~7500MB/sec....and using CPU-Z, it reports the DRAM frequency at ~660mhz.

You better forget about Novabench, IMO, that's not a reliable benchmarking tool.

Also, it's not quite practical to convert the transfer speed into DIMM clock speed like that.

Last but not least, you may Google what DDR means, than you should able to understand why it shows ~660MHz.

 

[Dewdman42]

You better forget about Novabench, IMO, that's not a reliable benchmarking tool.

Do you know another better tool I can use to measure what the Mac is actually doing? I am interested to measure this. I use Geekbench for my overall performance measurement according to GeekBench, which is only a relative number anyway compared to others who have ran Geekbench on the same hardware and doesn't give any specifics, while Novabench at least does provide what it measured while reading/writing RAM as a specific metric.

Also, it's not quite practical to convert the transfer speed into DIMM clock speed like that.

Why not?

crucial has a page that explains max peak data transfer rates for each DDR3 frequency:

RAM Memory Speeds & Compatibility

Better understand your computer's RAM speeds & the compatibility of different types of memory to work out how to get the most from your desktop or laptop.

www.crucial.com

According to them, DDR3-1333 should be able to achieve up to 10677 MB/sec, while DDR3-1066 can get up to 8533 MB/sec peak transfer rate. Novabench was measuring ~8300 MB/sec...which sorta seems like even though About This Mac reports my memory as being 1333, but might be clocking it down under the covers anyway under some circumstances...

Under Bootcamp, windows version of Novabench reports even slower around ~7500MB/sec.

If there is another better tool to measure actual RAM performance I would be interested in trying it, but I couldn't find anything that measured speed, only stuff that checks for memory integrity. Do you know of something?

I definitely got a 25% better GeekBench score when I removed the 4 and 8 DIMM modules (128gb down to 96gb ram)....even though in both cases About This Mac reports all the ram as being 1333.

I'm pretty happy with my latest Geekbench score to be honest...

so I'm just kind of asking for understanding I guess...and wanting to understand a little more what the 5,1 is actually doing if and when it might clock down the RAM.

Last but not least, you may Google what DDR means, than you should able to understand why it shows ~660MHz.

alright, copy that. DDR = Double Data Rate..., so 660mhz is a good sign that its attempting 1333mhz actual data throughput... While watching CPU-Z (under bootcamp) I do see that DIMM mhz value flicker and change...which indicates to me that it could theoretically get clocked down?

 

[tsialex]

Why not?

crucial has a page that explains max peak data transfer rates for each DDR3 frequency:

RAM Memory Speeds & Compatibility

Better understand your computer's RAM speeds & the compatibility of different types of memory to work out how to get the most from your desktop or laptop.

www.crucial.com

According to them, DDR3-1333 should be able to achieve up to 10677 MB/sec, while DDR3-1066 can get up to 8533 MB/sec peak transfer rate. Novabench was measuring ~8300 MB/sec...which sorta seems like even though About This Mac reports my memory as being 1333, but might be clocking it down under the covers anyway under some circumstances...

Why on earth are you expecting unregistered non-ECC RAM bandwidth performance when using registered ECC RAM?

 

[Dewdman42]

why on earth? I'm not "expecting" anything Alex, just asking questions. Any insight you have will be very welcome!

 

[Dewdman42]

by the way, my understanding is that ECC memory is only ~2% slower then non-ECC. The disparity I have mentioned is more like 20% so I don't think ECC is the reason.

 

[tsialex]

by the way, my understanding is that ECC memory is only ~2% slower then non-ECC. The disparity I have mentioned is more like 20% so I don't think ECC is the reason.

It's not just 2%, 72 bits have to be transferred instead of the 64 bits of the non-ECC RAM, but the lion share of the performance discrepancy is the additional latency of the register controller.

 

[Dewdman42]

I have read the opposite of what you are saying now. So you feel ECC is 20% slower then non-ECC? without more info I cannot agree with that.

here is crucial's point of view on that: https://www.crucial.com/support/art...Gb3u4tWmEYf5k0uX5En3e_COhBMZd-OkaArsoEALw_wcB

 

[Dewdman42]

Can you please explain about the Register Controller you mentioned?

 

[tsialex]

Can you please explain about the Register Controller you mentioned?

Your Crucial link is relevant only for desktop memory, non-ECC UDIMM. You are using ECC registered DIMM, a type of DIMM that have considerable performance tradeoff when compared to the standard UDIMM, the performance penalty is necessary to allow considerable more memory to be installed - the register controller acts as a buffer and adds latency, usually 1T with 2 ranks RDIMMs, reducing the electrical load with the memory controller.

non-ECC UDIMM performance is not directly comparable with ECC RDIMM.

Btw, Intel has a very good paper about the RDIMM issues with Nehalem and Westmere Xeons, it's on the doc site.

 

[Dewdman42]

This is another interesting article that gets into the "registered" part of it too:

ECC and REG ECC Memory Performance

Recently, a workstation PC was reviewed on a well-known review site that included a Quadro video card, but no ECC memory. In the comment section of that review, a heated discussion occurred with some readers stating that ECC was bad for workstations since it is slower than standard RAM. In this...

www.pugetsystems.com

In any case I'm still wondering more or less if the 5,1 is ramping down the Ram clock speed under any circumstances (regardless of what is reported in AboutThisMac, I'm gonna try some of these other benchmarking tools mentioned under bootcamp also. I do not think ECC nor Registered buffer can explain ~8300MB/sec measured performance according to Novabench, when DDR3-1333 can theoretically get up to 10677MB/sec. But still, ~8500MB/sec very well could be the max memory transfer speed achievable on the 5,1 for other reasons of some kind.... Registered ECC might be dropping mine down to 8300MB/sec in Novabench.

As I said before, this is really just a curiosity for me to understand what is going on, I am pretty darn happy with the performance of my 5,1 right now.

 

[Dewdman42]

more data please Alex. I do not believe that registered ECC will drop performance 20%. show more facts please to support that theory.

At that time I bought my ram it was very difficult to even source ram for this machine at all and many sources would not even provide exact specs...otherwise I would have preferred non-ECC, totally get it. Every little bit helps.

Still I just want to understand what the 5,1 is capable of...and specifically if and when it might be down-clocking the RAM. As I said earlier...

and I see no reason to believe that non-ECC memory would achieve 10677MB/sec hypothetical memory speeds....nor really an real difference in practical performance using real world apps...less then 1% difference if I am understanding everything correctly.

I looked around on Novabench's website and could not find any other 5,1 getting faster then 8500MB/sec memory transfer speeds either...

 

[tsialex]

more data please Alex. I do not believe that registered ECC will drop performance 20%. show more facts please to support that theory.

At that time I bought my ram it was very difficult to even source ram for this machine at all and many sources would not even provide exact specs...otherwise I would have preferred non-ECC, totally get it. Every little bit helps.

Still I just want to understand what the 5,1 is capable of...and specifically if and when it might be down-clocking the RAM. As I said earlier...

and I see no reason to believe that non-ECC memory would achieve 10677MB/sec hypothetical memory speeds....nor really an real difference in practical performance using real world apps...less then 1% difference if I am understanding everything correctly.

I looked around on Novabench's website and could not find any other 5,1 getting faster then 8500MB/sec memory transfer speeds either...

I've already wrote that it's not the additional 8 bits of the ECC RAM the major responsible for the decreased performance when comparing desktop to server memory, but the additional latency of the register controller circuit and ranking - it's a trade-off that allow the usage of more memory (more than the double of the total memory when using RDIMM instead of UDIMM with Nehalem/Westmere). Btw, Nehalem/Westmere Xeons are not directly comparable with newer Xeons benchmarks where Intel implemented dual memory controllers and completely different memory access/interleaving.

Buffered RAM adds at least 1T with the clock/control/address signals for dual rank DIMMs and the penalty increases considerably with bigger rankings - this is easy to check since 4Rx4 DIMMs when work with Nehalem/Westmere usually downgrade the memory controller reference frequency to 533MHz with most DIMMs downgrading to 400MHz. Read the Intel Nehalem and Westmere documentation to learn that are also several other penalty factors like ranking, interleaving and relaxed timings when using different memory types.

 

[Dewdman42]

That’s all very interesting. How do you arrive at 20% from that?

 

[tsialex]

That’s all very interesting. How do you arrive at 20% from that?

The only penalty number I've used at all is that registered RAM adds at least 1T to the timings and this alone is a considerable penalty on itself, plus Nehalem/Westmere memory controller adds more penalties that newer Xeons don't have. Please read the Intel Nehalem/Westmere documentation for the memory controller and learn why things are this way.

 

[h9826790]

Why not?

1) That's the max theoretical speed you can never reach in real world

2) You don't know how much memory bandwidth was taken up by other background process

3) You don't know how NovaBench measure that transfer rate (e.g. Write data to a RAM drive? Or ZeroMemory? Or CopyMemory? etc)

4) That max theoretical speed (from your link) isn't for triple channel setup

5) You don't know the system latency / overhead
......

Anyway, for your info, you aren't only getting 80% of the performance you should have. Your number is waaaaaaay below that.

For triple channel 1333MHz DDR3 memory config, the max theoretical memory bandwidth is ~32GB/s, not 10667MB/s. The number you measured is just ~25% of the theoretical max.

However, it means nothing. Your system is working correctly and performing as expected. IMO, that's the only thing we want to know by running benchmarks.

 

[Dewdman42]

Anyway, for your info, you aren't only getting 80% of the performance you should have. Your number is waaaaaaay below that.

Not sure I understand your English here...some double negatives in there...

For triple channel 1333MHz DDR3 memory config, the max theoretical memory bandwidth is ~32GB/s, not 1066MB/s. The number you measured is just ~25% of the theoretical max.

The number I quoted from Crucial's website was 10,667 MB/sec, not 1066

why did the crucial website cite the numbers they did, which I quoted above and that you don't agree with now?

is it just a mere coincidence that this number they cited, which is more in line with that NovaBench reports? what is the relevance of that number from crucial and how does it relate to what was reported by NovaBench?

However, it means nothing. Your system is working correctly and performing as expected. IMO, that's the only thing we want to know by running benchmarks.

well recently I dropped my memory from 128gb to 96gb and got a 25% boost in benched performance. The AboutThisMac info was showing all 8 DIMM's as running at 1333mhz, but apparently...pulling out the DIMM's from slots 4 and 8, which are shared dual channel....increase benchmark score by 25%. Were they causing things to clock down, even though not displayed as clocked down in AboutThisMac? I don't know, but that is why I'm asking these questions just to understand and make sure I'm not missing any other significant opportunities for improvement.

 

[Dewdman42]

plus Nehalem/Westmere memory controller adds more penalties that newer Xeons don't have. Please read the Intel Nehalem/Westmere documentation for the memory controller and learn why things are this way.

This part ^^^^^ may explain things, thanks for pointing that out.. Between that and the Registered ECC, might be enough to deliver 1066mhz effective memory transfer speeds, despite 1333mhz clock being used. That effective transfer measurement is ~20% less, according to what I read from crucial. according to h9826790, it might be 75% under the theoretical max possible data transfer rate...so I'm not sure about the disparity between his info and Crucial's info...but anyway...This isolated thing is not al that matters and overall I am quite happy with my 5,1 multicore GeekBench score (~6900). Just wanting to understand as I said...

Anyway, I personally need more memory for what I do, so its basically not an option fro me to use non-ECC...

 

[h9826790]

why did the crucial website cite the numbers they did, which I quoted above and that you don't agree with now?

That's for single channel.

For your memory config, you should use this number as reference.

 

[h9826790]

Not sure I understand your English here...some double negatives in there...

You believe that you "only having 80% of performance", but the reality is "way below that".

 

[h9826790]

The number I quoted from Crucial's website was 10,667 MB/sec, not 1066

My fault, post fixed. But that 32GB/s is correct.

 

[h9826790]

is it just a mere coincidence that this number they cited, which is more in line with that NovaBench reports? what is the relevance of that number from crucial and how does it relate to what was reported by NovaBench?

Crucial shows your the calculated theoretical max. Your NovaBench result is nowhere near that.