MP 1,1-5,1 cMP & nMP {MacPro 4,1 & 5,1} Memory Upgrade Compatibility and FAQ

[tpivette89]

They

Just picked up a 4,1 single processor a few days ago and have updated to 5,1.
Would this 4 of these sticks RAM work when I upgrade the processor to a 6 core?
Micron 8GB 2Rx4 PC3L-12800R DDR3 Server Ram Memory MT36KSF1G72PZ-1G6K1LG
thx

They would work, but that is 1600mhz RAM. Would down clock to 1333mhz.

Also, running 4 sticks of RAM would slightly reduce performance over 3 sticks... these CPUs were designed to be run in triple channel mode. If you need the extra RAM, it may be better to find 16GB sticks and run 3x16GB (for 48GB), instead of 32GB (4x8GB).

Just a thought.

 

[Manek43509]

Looking at potentially upgrading the RAM in my Mac Pro, but not sure what I should get. I can't find this processor listed on Everymac for this machine, so it's obviously been upgraded at some point. Ideally I'd like to go up to 1333mHz RAM instead of the 1066mHz. Anyone know which sticks would be compatible with this Mac? Thanks!

 

[tsialex]

Looking at potentially upgrading the RAM in my Mac Pro, but not sure what I should get. I can't find this processor listed on Everymac for this machine, so it's obviously been upgraded at some point. Ideally I'd like to go up to 1333mHz RAM instead of the 1066mHz. Anyone know which sticks would be compatible with this Mac? Thanks!

View attachment 935635

It's a X5690 or W3690. You can know for sure with:

sysctl machdep.cpu.brand_string

With a single CPU tray you already have the optimum memory configuration:

• 3 x 16GB with W3690 (this processor do not support 64GB)
• 3 x 16GB with X5690 for optimum memory throughput
• 4 x 16GB with X5690 for maximum memory with some throughput penalty.

If your DIMMs support 1333MHz, reset the NVRAM 3-times consecutively to force DIMM re-detection.

 

[Manek43509]

Thank you! I'll try that next time I'm back in the studio. How will I know whether the DIMMs support 1333mHz or not?

 

[macsplusmore]

Hi, I am hoping to get some assistance. I am upgrading a Mac Pro 4,1. I have done everything needed to get it to run 1333mhz ram i.e 5,1 firmware update, installed 2 x X5690's, and added 96GB of Ram. However, system report only shows 1066. What am I missing? Any help is greatly appreciated. Ram model is Samsung M393B2K70DMB-YH9 6x16

 

[Manek43509]

It's a X5690 or W3690. You can know for sure with:

sysctl machdep.cpu.brand_string

With a single CPU tray you already have the optimum memory configuration:

• 3 x 16GB with W3690 (this processor do not support 64GB)
• 3 x 16GB with X5690 for optimum memory throughput
• 4 x 16GB with X5690 for maximum memory with some throughput penalty.

If your DIMMs support 1333MHz, reset the NVRAM 3-times consecutively to force DIMM re-detection.

Terminal says:
Intel(R) Xeon(R) CPU X5690 @ 3.47GHz

 

[flyproductions]

How will I know whether the DIMMs support 1333mHz or not?

Most likely the machine has been upgraded with quad-rank-dimms (instead of dual-rank) which "only" run @ 1066 istead of 1333 even with fast upgraded cpus. You can verify this by checking their labels:

2Rx4 = Dual-Rank, 4Rx4 = Quad-Rank

But in real life performance - aside from the cosmetic system-info-thing - this doesn't make much of a difference. Even the plain "speed difference" isn't that much. But the shorter cas latency (CS7 instead of 9), which the quad-ranks run with, reduces the effect to almost zero. In fact under circumstances where latency matters, i.e. random access to small amounts of data, the quad-rank-memory might even be mesurably faster.

 

[wfmwfm]

My experience with my 5680 12core 3.33 cMP this week:

I ordered 8 of the 16gb Samsung server sticks at $18 each off eBay. Pulling my dual processors tray out numerous times (it can be a bit fiddly & stress inducing btw) I could get up to 80gb to register.

The #4, #7 & #8 slots would show the red error LEDs on the motherboard.
Eventually I got 6 of the 16gb to register in slots #1, 2, 3 - 5, 6, 7 along with 4gb sticks in slots #4 and #8.

Off Topic news:
I got my 10.13.6 to firmware 089 finally and installed a $79 MSI RX 560 4gb eBay card. Rocks!

Maybe I'll do Mojave next week:
CCC cloning my current disk to a 1TB SSD and then installing from my original 480gb 10.13.6. Has anyone done this?

 

[flyproductions]

The #4, #7 & #8 slots would show the red error LEDs on the motherboard.

In most cases pulling and carefully reseating the sticks is effective in fixing this.

 

[Noobraino]

Hey guys, quick question about RAM Seller. How reliable is A-Tech Components? They sell very cheap 64gig package for cMP 5.1 which leads me to think there is a catch here because most other sellers have higher price.

Any thoughts about this company?

Thank you.

 

[tsourbier]

Just to report that I successfly upgraded my single proc MacPro 5,1 with a x5690 CPU and 64Gb of Hynix RAM.

The references for the 4 RAM modules were:

Hynix 16GB (1x16GB) PC3L-10600R 2Rx4 RDIMM 1333MHz DDR3 ECC REG LV

These sticks may be sold under several references related to HP servers(?) :
664692-001
A0R59A
647901-B21
628974-081
632204-001
647653-081

Note that other makers also make these references (I noted one users here had a good experience with Micron memory stick...).

I went for Hynix because it was the cheapest compatible memory I could find on Techbuyer (present at least in the US, Europe and Australia). FYI price was 24.15€ per stick (+ 20% VAT if applicable). Free next day delivery in my case (I got a used X5690 from them as well).

Note this seller, which whom I have no affiliation, specializes in the new and refurbished server/server parts market. At least in France they were not directly aware of this Mac Pro niche market, they were at first a bit surprised of my request and checked if I knew what I was doing (very thoughful of them).

T.

 

[unita_logica]

Hello everyone
I was thinking to buy 3x 16 GB PC3-10600 DIMMs from DataMemorySystems for my cMP5,1 (W3565) but I couldn’t find if these are dual rank or quad rank. Does anybody know?

16GB DMS Certified Memory DDR3-1333 (PC3-10600) 2048x72 CL9 1.5v 240 Pin ECC Registered DIMM - DMS

Buy 16GB DMS Certified Memory DDR3-1333 (PC3-10600) 2048x72 CL9 1.5v 240 Pin ECC Registered DIMM with a Lifetime gaurantee from Data Memory System

www.datamemorysystems.com

 

[basslik]

WoW, so it appears I went overboard with buying extra ram. I figured it was a good choose to max out my 4,1> 5,1 Mac Pro 2010 with 128megs of Micron MT36JSF2G72PZ-1G9N1KG 16GB DDR3 Registered ECC PC3-14900 1866Mhz.
I figured since pro Tools went all 64bit Pro Tools would use every ounce of the ram ?

Now what about Open Core having the ability to run 256 ?

 

[basslik]

It's a X5690 or W3690. You can know for sure with:

sysctl machdep.cpu.brand_string

With a single CPU tray you already have the optimum memory configuration:

• 3 x 16GB with W3690 (this processor do not support 64GB)
• 3 x 16GB with X5690 for optimum memory throughput
• 4 x 16GB with X5690 for maximum memory with some throughput penalty.

If your DIMMs support 1333MHz, reset the NVRAM 3-times consecutively to force DIMM re-detection.

Alex how much of a throughput hit if I'm maxed out on 128gigs (ALL 8 SLOTS) on my Mac pro 5,1 ?
Also Open Core can address 256 (supposedly)?
THANKS

 

[tsialex]

Alex how much of a throughput hit if I'm maxed out on 128gigs (ALL 8 SLOTS) on my Mac pro 5,1 ?

It's extremely variable, depends if the app is heavily memory bound, like some scientific applications. For most apps is a ~5% hit, but it's not something that you know without real world tests.

Also Open Core can address 256 (supposedly)?

THANKS

While OpenCore can provide the infrastructure to support some types of 32GB DIMMs, it's you that have to manually config for each DIMM. Look at what is necessary before spending any money, it's not something easy or that can be easily automated and you can't boot without the OpenCore config.

 

[basslik]

Thanks Alex. too old of machine to spend on ram just to see if it yield any meaningful results I would imagine.
I read on here about two years ago. more ram is more ram regardless of triple channel config. seems plausible.

 

[GoJohnGo]

There's a good thread here on the performance differences between 3 and 4 sticks: https://forums.macrumors.com/threads/macpro5-1-ram-configuration-question.2096233/

 

[pywm]

First post here, please be gentle. I have read this whole thread plus all the related threads on MacRumours.

I just upgraded my cMP (4,1->5,1, Dual X5690) to 256GB RAM (32G x8, Hynix 4RX4 PC3L-10600R), using Martin Lo's OpenCore 0.9.2, running Monterey.

Question: why is the memory throughput so much lower in MacOS than in Windows 10/Linux with 256GB RAM?

In MacOS, using AmorphousMemoryMark, with 192GB (32Gx6), SEQ1M T8 read is about 22GB/s, but with 256GB (32Gx8), it is only about 10GB/s, while SEQ1M T1 degraded slightly but still high single digit.

When comparing to Windows 10/Linux, I use Geekbench 5 multi-core score as a proxy. Single-core scores are almost the same for all the following different cases.

Baseline (192GB):
Geekbench 5 Multi-score: MacOS 6913, Windows 10[OpenCore] 6644, Linux[no OpenCore] 6228

More RAM (256GB):
Geekbench 5 Multi-score: MacOS 4228 (-39%), Windows 10[OpenCore] 5770 (-13%), Windows 10[no OpenCore] 6225, Linux[no OpenCore] 6510

I understand the clock speed runs at 800Mhz with 256GB, but in Windows 10 the CL drops to 6 to compensate so an effective speed drop of 13% is acceptable. But in MacOS the drop is 39% and still underperforms Windows by another 27% under the same condition is perplexing.

OpenCore seems to introduce some performance penalty. I've used Martin Lo's default custom memory settings, I have also tested with my own settings based on dmidecode from Linux.

Does anyone have a theory about this? And better still, a way to fix it? Could change the platform code to MacPro7,1 help at all?

 

[CIA]

Not sure how applicable it is, but the cMP works best in triple channel mode for memory, not quad.

 

[flyproductions]

In MacOS, using AmorphousMemoryMark, with 192GB (32Gx6), SEQ1M T8 read is about 22GB/s, but with 256GB (32Gx8), it is only about 10GB/s, while SEQ1M T1 degraded slightly but still high single digit.

You are loosing Triple Channel when "upgrading" from 192 to 256. So the fourth socket should always be left open, if performance is your goal. 256 is just for an "impressive" "About This Mac".

As easy as that!

 

[pywm]

You are loosing Triple Channel when "upgrading" from 192 to 256. So the fourth socket should always be left open, if performance is your goal. 256 is just for an "impressive" "About This Mac".

As easy as that!

Thanks, I understand losing triple channel, and in theory, it would become a dual channel. I've tested using all 16GB sticks and all 8GB sticks before, the difference is only 5% between 3-slots vs 4-slots.

My question is why MacOS perform so much worse than Windows 10 under the same hardware setting. The amount of underperformance is about 26% which is too large.

 

[flyproductions]

My question is why MacOS perform so much worse than Windows 10 under the same hardware setting. The amount of underperformance is about 26% which is too large.

Looks like the Geekbench-Score is just more memory-dependent for MacOS. What troughput do you get with all slots populated in Windows?

 

[pywm]

Looks like the Geekbench-Score is just more memory-dependent for MacOS. What troughput do you get with all slots populated in Windows?

Yes, Geekbench is more memory dependent, that's why I am using it as a proxy when testing memory. I don't have the throughput data on Windows yet. I will try it tonight using Aida64.

Under Windows 192GB runs at 1066Mhz at CL7, while 256GB is 800Mhz at CL6. If you do the maths, the difference is about 15% which is about right from the geekbench differential.

My current hypothesis is that MacOS under 256GB is running at 800Mhz but CL9. However I don't know how to get the CAS Latency number in MacOS. IF this theory is correct, maybe we could patch OpenCore to force it to run at CL6 which should improve the speed under 256GB.

 

[pywm]

Yes, Geekbench is more memory dependent, that's why I am using it as a proxy when testing memory. I don't have the throughput data on Windows yet. I will try it tonight using Aida64.

Under Windows 192GB runs at 1066Mhz at CL7, while 256GB is 800Mhz at CL6. If you do the maths, the difference is about 15% which is about right from the geekbench differential.

My current hypothesis is that MacOS under 256GB is running at 800Mhz but CL9. However I don't know how to get the CAS Latency number in MacOS. IF this theory is correct, maybe we could patch OpenCore to force it to run at CL6 which should improve the speed under 256GB.

My hypothesis is wrong. CAS Latency doesn't impact bandwidth.

I have tested the bandwidth in Windows 10 using Aida64 Extreme. In MacOS, using AmphoursMemoryMark running three times, picking the best score (SEQ1M T8 Read).

Baseline (192GB - 32GBx6):
Windows 10 27.98GB/s latency 72.7ns, MacOS 25.46GB/s
Windows memory timing: 1066Mhz, 7-7-7-20 CR1, DRAM:FSB Ratio 4:1

Max RAMS: (256GB - 32GBx8):
Windows 10 24.58GB/s latency 81.9ns, MacOS 10.56GB/s <= MacOS is doing something wrong here.
Windows memory timing: 800Mhz, 6-6-6-16 CR1, DRAM:FSB Ratio 3:1

Mixed RAMs: (192GB - 32GBx4 + 16GBx4):
Windows 10 27.97GB/s latency 73.5ns, MacOS 24.77GB/s
Windows memory timing: 1066Mhz, 7-7-7-20 CR1, DRAM:FSB Ratio 4:1

It is interesting that when I am testing different combinations using 16GB sticks (2Rx4 12600R) in MacOS, there is only two tier of bandwidth:
16GBx6 (3-slots) and 16GBx4 (2-slots) both give around 25.7GB/s
16GBx2 (1-slot) and 16GBx8 (4-slots) both give around 17.1GB/s

 

[flyproductions]

Baseline (192GB - 32GBx6):
Windows 10 27.98GB/s latency 72.7ns, MacOS 25.46GB/s
Windows memory timing: 1066Mhz, 7-7-7-20 CR1, DRAM:FSB Ratio 4:1

Where do you get this timings and the "true" speed from. My memory,x sticks of 32, 192 in total, "claims" to run at 1.333. But i guess, this is just what i entered into the config.plist. Is there any way to check what the speed really is in MacOS?

Max RAMS: (256GB - 32GBx8):
Windows 10 24.58GB/s latency 81.9ns, MacOS 10.56GB/s <= MacOS is doing something wrong here.

I'm finally more surprised, how Windows can keep the trougput that high, while triple channel is definitely lost. Or there beeing no real difference in the troughput with triple vs. dual channel in Windows.