How is RAM frequency controlled in 2009 Mac Pros with QPI?

[gugucom]

I understand that the Nehalem Gainestown CPUs with W and X TDP rating (X5550, X5560, X5570, W5580, W5590) are designed for 1333 MHz memory speed.

The E rated CPUs (E5520, E5530, E5540) are designed for 1066 MHz DDR3 memory.

Apple runs all E, X and W rated CPUs in the 2009 Mac Pro with 1066 MHz. I have a MP4,1 Nehalem octad with W5590 CPUs which still runs with the original 6 GB memory in 1066 MHz. I'm thinking about more memory and obviously about getting the higher 1333 MHz frequency that the CPUs are designed for.

When I asked an experienced user (Tutor) he told me that this is impossible. Both BTO options of 2,66 and 2,93 GHz CPUs also support the faster memory. So it appears likely that some more tech savvy users have looked into ways to tweak the memory or simply fitted faster memory.

So my question is how could a memory speed upgrade possibly be achieved.

 

[Eidorian]

The RAM speed is based off of the base 133 MHz BCLK value. The problem is accessing the memory multiplier to change it from x8 to x10 for the faster RAM. You don't really have access to the EFI in order to do so.

 

[gugucom]

OK, that is understood. So Apple systematically gives away possible memory bandwidth of those customers who pay for the higher clocked CPU options. Cheaky buggers.

 

[Eidorian]

I take it that dropping in a W or X 55xx Series doesn't change the RAM speed?

Otherwise you can overclock the BLCK to 166 MHz and get ~1333 MHz RAM. Now doing that is another thing entirely but it should be easier to change the BLCK than the memory multiplier.

 

[gugucom]

I take it that dropping in a W or X 55xx Series doesn't change the RAM speed?

No, my RAM speed did not change when I went from E5520 to the W5590 chips. It is still 1066.

Apple are a lazy bunch when it comes to doing EFI work. They rather screw the customers for top end performance than put in the work.

 

[VirtualRain]

No, my RAM speed did not change when I went from E5520 to the W5590 chips. It is still 1066.

Apple are a lazy bunch when it comes to doing EFI work. They rather screw the customers for top end performance than put in the work.

For your RAM speed to increase, three things all need to happen:
1. The memory controller needs to support the desired speed
2. The memory needs to support the desired speed (with programmed SPD to match)
3. The firmware needs to be set to use the SPD and not hard-coded to 1066

When the computer posts, it will fetch the memory speed and timings from the attached DIMM's and settings from firmware to configure the memory controller. It's very possible that using 1333 DIMMS with a CPU that supports it might give you 1333 speeds if Apple is relying on the SPD to set memory controller config.

If they have hard-coded the memory controller to use 1066 (which would be a bad practise IMHO) then you wouldn't get any of the benefits of the added clocks from running faster memory, but you still might get improved memory timings as most enthusiast/performance oriented 1333 memory will have an SPD entry for 1066 as well which may include tighter timings... but again, that assumes Apple has not hard-coded the timings (again a bad practise).

Has anyone tried 1333 memory in a system with a CPU that supports it? Running CPUz in Windows would quickly tell you if it's working or not.

Some light reading on SPD if you're interested... http://en.wikipedia.org/wiki/Serial_presence_detect

 

[gugucom]

For your RAM speed to increase, three things all need to happen:
1. The memory controller needs to support the desired speed
2. The memory needs to support the desired speed (with programmed SPD to match)
3. The firmware needs to be set to use the SPD and not hard-coded to 1066

When the computer posts, it will fetch the memory speed and timings from the attached DIMM's and settings from firmware to configure the memory controller. It's very possible that using 1333 DIMMS with a CPU that supports it might give you 1333 speeds if Apple is relying on the SPD to set memory controller config.

If they have hard-coded the memory controller to use 1066 (which would be a bad practise IMHO) then you wouldn't get any of the benefits of the added clocks from running faster memory, but you still might get improved memory timings as most enthusiast/performance oriented 1333 memory will have an SPD entry for 1066 as well which may include tighter timings... but again, that assumes Apple has not hard-coded the timings (again a bad practise).

Has anyone tried 1333 memory in a system with a CPU that supports it? Running CPUz in Windows would quickly tell you if it's working or not.

I could obviously ask my supplier to sell me RAM sticks that would work with 1066 and 1333 MHz frequency. I thought that this approach must have been used by other power users before because the BTO CPU options also support 1333. Perhaps I wait some days to see if someone tried. I also thought that the use of QPIs instead of FSB would shift the issue away from the chipset to the CPU used and of course the EFI settings.

 

[wibongo]

I could obviously ask my supplier to sell me RAM sticks that would work with 1066 and 1333 MHz frequency. I thought that this approach must have been used by other power users before because the BTO CPU options also support 1333. Perhaps I wait some days to see if someone tried. I also thought that the use of QPIs instead of FSB would shift the issue away from the chipset to the CPU used and of course the EFI settings.

The Nehalems 'memory controller' is bult into the CPU, however memory frequency and QPI can be independent of each other (much like memory speed and FSB speed in older CPUs).

On a PC mainboard there's usually an option to 'lock' or 'unlock' the QPI and memory speed so you can run the memory at any speed it is rated for or if you're using the 'SPD' detect this is done automatically (in a perfect world anyway). I do have some 1333 DDR3 memory here, and have tried it. It runs at 1066 in my MP however I only have the octo 2.26 model. I'm pretty sure that my model CPU does not support 1333 memory anyway and I really haven't lost any sleep over it.

One of the things to consider, and this may have forced Apple's hand on limiting the speed to 1066, is the availability of higher speed memory WITH ECC. Most of the mainstream desktop DDR3 is NOT ECC, and in my opinion unsuitable for use in a Mac Pro. I wouldn't give up the extra stability of ECC for a marginal increase in memory bandwith. Just my opinion

 

[gugucom]

I do have some 1333 DDR3 memory here, and have tried it. It runs at 1066 in my MP however I only have the octo 2.26 model. I'm pretty sure that my model CPU does not support 1333 memory anyway and I really haven't lost any sleep over it.

You are right, the E5520 is designed for 1066 MHz only. So your trial will not be usefull.

 

[Gonk42]

1333MHz ECC memory is readily available and is roughly the same price (at Crucial at least). It would only work if just three memory slots are used as if any channel has two sticks then it will automatically switch to 1066MHz anyway.

It would be very interesting if anyone does try this with a 2.66GHz or 2.93GHz (or 3.33GHz) Octo machine - I've been curious since the machines came out but no one has tried it yet, as far as I know (with 1333MHz ECC memory -
someone did try non-ECC memory which may not have been properly recognized.)

The other question is, why don't octo Macs recognize Registered RAM. For 4GB chips registered RAM is still much cheaper so it would be very interesting if someone with an octo tried this (being carefull not to mix Registered and unbuffered RAM).

 

[Eidorian]

No, my RAM speed did not change when I went from E5520 to the W5590 chips. It is still 1066.

Apple are a lazy bunch when it comes to doing EFI work. They rather screw the customers for top end performance than put in the work.

As much as I do like the Mac Pro's platform, Apple isn't the best choice if you're going to drop in new processors or want the EFI/BIOS to be accessible and flexible.

VirtualRain, just for the sake of completeness, RAM can be overclocked as well. Though that depends on the cooling and binning of the slower components.

If you have the DDR3-1333 RAM but can't access the memory multipler then the next easiest thing is to overclock the processor's BCLK from 133 MHz to 166 MHz. Then it's 166 MHz x 8 for ~1333 MHz RAM. You'd be overclocking the processor as well. The jury is out on how the logic board and EFI would handle that.

 

[Umbongo]

The other question is, why don't octo Macs recognize Registered RAM. For 4GB chips registered RAM is still much cheaper so it would be very interesting if someone with an octo tried this (being carefull not to mix Registered and unbuffered RAM).

2009 Mac Pros work with Registered RAM.

 

[Gonk42]

2009 Mac Pros work with Registered RAM.

Have you tried this? I don't doubt your veracity, as it definitely should
work, but I've not seen any reports from anyone actually putting it into
an 09 Mac Pro.

It is significant because, even though OWC sells RAM cheaply, in the UK Crucial
(for example) charge £625 for 3x4GB of unbuffered RAM whilst only £388 for 3x4GB of registered RAM - and it is easier to get registered 1333MHz ECC RAM.

 

[Umbongo]

Have you tried this? I don't doubt your veracity, as it definitely should
work, but I've not seen any reports from anyone actually putting it into
an 09 Mac Pro.

It is significant because, even though OWC sells RAM cheaply, in the UK Crucial
(for example) charge £625 for 3x4GB of unbuffered RAM whilst only £388 for 3x4GB of registered RAM - and it is easier to get registered 1333MHz ECC RAM.

Someone posted here about it not long ago.

 

[gugucom]

I have made a special deal with my supplier. He will supply me 6x2GB ECC 1333 at cost. I will do the test and if positive give him a screen shot of my Mac running @ 1333 for his advertising.

If the RAM do not run he takes them back again at cost. So I have no risk of being left sitting on components I cannot use.

I'm paying 240€ for 6x2. I would be paying 600€ for 6x4.

I was thinking that 3x4GB would not work properly in a MP4,1. Or at least it would drop bandwidth considerably. Can anybody confirm that.

 

[Eidorian]

I don't see a problem with 3 x 4 GB. By design the Mac Pro (Nehalem 2009)and its processors are built to work on a triple channel memory connection.

 

[nanofrog]

It's very possible that using 1333 DIMMS with a CPU that supports it might give you 1333 speeds if Apple is relying on the SPD to set memory controller config.

Yep.

But I've not recalled anyone that's tried it, so no confirmation if Apple used SPD or set a fixed value in the EFI. Hopefully, it's via SPD, so those that have the faster clocked units (shipped or swapped), could utilize the faster RAM at thier discretion.

If they have hard-coded the memory controller to use 1066 (which would be a bad practise IMHO)...

I agree it's bad practice, especially as they don't allow firmware access. The only thing that comes to mind for setting a fixed value, is if the ROM capacity was nearing filled and the firmware still incomplete. Resulting in they had to go back and make some compromises to to stay within the capacity limit.

Allowing firmware access would have allowed for the ability to change the memory manually, had capacity been an issue. But Apple hasn't supplied access to firmware during any of the Intel based systems, nor expect an update that will allow it on existing systems.

...I only have the octo 2.26 model. I'm pretty sure that my model CPU does not support 1333 memory anyway and I really haven't lost any sleep over it.

No, as gugucom posted, it's stuck at 1066.

One of the things to consider, and this may have forced Apple's hand on limiting the speed to 1066, is the availability of higher speed memory WITH ECC. Most of the mainstream desktop DDR3 is NOT ECC, and in my opinion unsuitable for use in a Mac Pro. I wouldn't give up the extra stability of ECC for a marginal increase in memory bandwith. Just my opinion

1333MHz UDIMMs were available in time to ship. Cost would make more sense IMO as to why they didn't use them.

 

[nanofrog]

I was thinking that 3x4GB would not work properly in a MP4,1. Or at least it would drop bandwidth considerably. Can anybody confirm that.

Not that I'm aware of. It's not dealing with interleaving at all, as you have 1 UDIMM per channel.

 

[gugucom]

I don't see a problem with 3 x 4 GB. By design the Mac Pro (Nehalem 2009)and its processors are built to work on a triple channel memory connection.

So how would the three DIMMs be distributed over the two RAM banks?

2 sticks in A and 1 stick in B? It sounds strange somehow.

 

[nanofrog]

So how would the three DIMMs be distributed over the two RAM banks?

2 sticks in A and 1 stick in B? It sounds strange somehow.

Place one in DIMM slots 1, 2, and 3 (source).
Slot 3 shares with slot4, slot 7 shares with slot 8.
If you double the capacity later (4GB UDIMM's again), you add to 5,6, and 7 to preserve triple channel operation.

Hope this helps.

 

[bbadalucco]

So how would the three DIMMs be distributed over the two RAM banks?

2 sticks in A and 1 stick in B? It sounds strange somehow.

Would this potentially work with a W3580???

 

[gugucom]

Place one in DIMM slots 1, 2, and 3 (source).
Slot 3 shares with slot4, slot 7 shares with slot 8.
If you double the capacity later (4GB UDIMM's again), you add to 5,6, and 7 to preserve triple channel operation.

Hope this helps.

THX nano, a case of RTFM.

Actually I would be using only three out of six memory channels that way. So I would certainly be loosing memory bandwidth. I will buy 6 x 2GB and use the slots (1,2,3,5,6,7).

If the 1333 works out and I decide to go to 24 GB, I will buy only 4 x 4GB and place them in slots (1,2,5,6). Slots (3,4,7,8) I will fill with 2GB modules and sell off the remaining two. That way I will be having optimum memory bandwidth and avoid too much selling of excess modules.

Sometimes I hate the penny pinchers at Apple. Those MP4,1 should really all come with 1333 MHz memory and six RAM slots per CPU. So each memory channel should share two RAM slots properly like other boards do. Some have even three slots share one memory channel for a total of 18 slots.

 

[gugucom]

Would this potentially work with a W3580???

You have only three memory channels in four slots. If you want to mix you should have 4 GB in slots (1,2) and 2 GB in slots (3,4).

As your CPU is W designated I would expect it to be designed for 1333 MHz as well. Obviously you cannot mix the original 1066 MHz modules with the 1333 as it would pull down the whole lot.

 

[nanofrog]

THX nano, a case of RTFM.

Actually I would be using only three out of six memory channels that way. So I would certainly be loosing memory bandwidth. I will buy 6 x 2GB and use the slots (1,2,3,5,6,7).

If the 1333 works out and I decide to go to 24 GB, I will buy only 4 x 4GB and place them in slots (1,2,5,6). Slots (3,4,7,8) I will fill with 2GB modules and sell off the remaining two. That way I will be having optimum memory bandwidth and avoid too much selling of excess modules.

Sometimes I hate the penny pinchers at Apple. Those MP4,1 should really all come with 1333 MHz memory and six RAM slots per CPU. So each memory channel should share two RAM slots properly like other boards do. Some have even three slots share one memory channel for a total of 18 slots.

Technically, using 1, 2, and 3 is still triple channel operation, but to get anything to the second CPU, the memory data has to be routed over the QPI channels, which are slower.

Your idea to balance the pair of IMC's with UDIMM's is the best way to do it. Less to be transferred over the QPI's, as each IMC has 3 channels of it's own to play with. If one IMC isn't loaded, and the software can utilize the additional memory, it can be loaded on the other IMC, and transferred over QPI's.

You have only three memory channels in four slots. If you want to mix you should have 4 GB in slots (1,2) and 2 GB in slots (3,4).

As your CPU is W designated I would expect it to be designed for 1333 MHz as well. Obviously you cannot mix the original 1066 MHz modules with the 1333 as it would pull down the whole lot.

The CPU is capable, but we don't yet know if the firmware is using SPD or a fixed value to set the memory speed. If it's SPD, replacing the existing memory with 1333 will work with full throughput. Otherwise it will be stuck to 1066. (Apple would really have screwed their users. Again).

I'm looking forward to your experiment here, as we can finally get a definitive answer.

 

[KG2002]

So, if I am using the base Quad 2.66 model and plan to upgrade it later on to the 6 Core chip (if it is going to work) or to a faster Quad core (if 6 core is not going to be an option), I should be getting 1333 RAM? I am reading it is backward compatible, am I right?

I just got the box and scratching my head which RAM to get so I am not spending $$$ again 6 months from now. I need 12G or 16G and they do not come cheap.