Mac Pro Quad FSB Bottleneck limit for proc transplant?

TallGlassOfFail

macrumors member
Original poster
Sep 16, 2009
43
0
Hey guys.

I bought the late 2009 Mac Pro as i really needed it to do my work at uni and be generally more productive as working on my old white polycarb macbook was just not cutting it. Anyrate, so i bought the low end quad core, 2.66Ghz.

Heres my questions:

1) I am wondering if i wanted to upgrade the processor at a later date with a faster quad xeon, what frequency or model could i go up to before the FSB bottlenecked and made anything faster futile?

2) Also would i need to apply thermal paste, or is it just out it comes and in it goes?

3) And one more thing, would the new intel 6 cores being proposed possibly work?

Thanks.
 

TallGlassOfFail

macrumors member
Original poster
Sep 16, 2009
43
0
Thanks for your reply, if its not called FSB then whatever apple calls it the motherboard or some other component, daughter board maybe surely imposes some kind of limit on data transfer. What would that component be and substituted for FSB in my original questions what would the upper limit be?

I read the thread you linked to but i did not find mention of FSB or its equivalent.
 

Billydelp4

macrumors regular
Feb 29, 2008
102
0
Thanks for your reply, if its not called FSB then whatever apple calls it the motherboard or some other component, daughter board maybe surely imposes some kind of limit on data transfer. What would that component be and substituted for FSB in my original questions what would the upper limit be?

I read the thread you linked to but i did not find mention of FSB or its equivalent.
It uses qpi
 

Umbongo

macrumors 601
Sep 14, 2006
4,929
54
England
Hey guys.

I bought the late 2009 Mac Pro as i really needed it to do my work at uni and be generally more productive as working on my old white polycarb macbook was just not cutting it. Anyrate, so i bought the low end quad core, 2.66Ghz.

Heres my questions:

1) I am wondering if i wanted to upgrade the processor at a later date with a faster quad xeon, what frequency or model could i go up to before the FSB bottlenecked and made anything faster futile?

2) Also would i need to apply thermal paste, or is it just out it comes and in it goes?

3) And one more thing, would the new intel 6 cores being proposed possibly work?

Thanks.
1. You can upgrade to any Core i7 or Xeon 3500 series processor. Currently they top out at 3.33GHz, so you would get roughly 25% extra processing power.

2. Yes you should apply it.

3. Regard this as a no, it isn't impossible, just no real chance of it happening on current information.
 

TallGlassOfFail

macrumors member
Original poster
Sep 16, 2009
43
0
1. You can upgrade to any Core i7 or Xeon 3500 series processor. Currently they top out at 3.33GHz, so you would get roughly 25% extra processing power.

2. Yes you should apply it.

3. Regard this as a no, it isn't impossible, just no real chance of it happening on current information.
Thanks for your replies. So the QPI (thanks Billydelp4) would have an upper limit high enough that it would not impose a bottleneck of any of the core i7 or xeon processors? (just wanted to avoid buying a proc that the qpi [now i know of QPI] would prevent the maximum use of).

Strange how we always got to add thermal paste these days, i swear we never used to have to add that stuff when i used to work on hardware during the windows 98 times as i had to for college in the hardware lab. Just used to yank them out and swap them out. Had to do it but hardware is never been my thing, prefer programming.

And if i upgraded to a Core i7 would it be a straight forward swap or is there any messing with firmware? would it use more than 4 cores? and would apple updates brick the machine?

Thanks :)
 

goMac

macrumors 604
Apr 15, 2004
6,759
780
Thanks for your replies. So the QPI (thanks Billydelp4) would have an upper limit high enough that it would not impose a bottleneck of any of the core i7 or xeon processors? (just wanted to avoid buying a proc that the qpi [now i know of QPI] would prevent the maximum use of).
QPI has no limit at all, as far as I am aware. It's all processor dependent.

But likely you're not going to be able to stick a Gulftown in there anyway because of firmware issues. And all the current Xeons and Core i7s have the same memory frequency, iirc.

I only have a Nehalem at work, where I'm not interested in upgrading it, so I'm just working from my understanding, which could be wrong. :)
 

gugucom

macrumors 68020
May 21, 2009
2,136
0
Munich, Germany
Memory speed in the 2009 MP is controlled by firmware to a fixed 1066 MHz level. The primary memory control is done by the CPU itself and can vary between 800 and 1333 MHz depending of the CPU speed level, but as I said Apple has castrated it at 1066. They are not using the low level CPUs and make all their top level CPUs go slower to achieve this. Actually a silly move in my view.
 

TallGlassOfFail

macrumors member
Original poster
Sep 16, 2009
43
0
QPI has no limit at all, as far as I am aware. It's all processor dependent.

But likely you're not going to be able to stick a Gulftown in there anyway because of firmware issues. And all the current Xeons and Core i7s have the same memory frequency, iirc.

I only have a Nehalem at work, where I'm not interested in upgrading it, so I'm just working from my understanding, which could be wrong. :)
So are the Core i7's gulftown? Are you saying the Core i7's would not work? What are the best options?

gugucom said:
Memory speed in the 2009 MP is controlled by firmware to a fixed 1066 MHz level. The primary memory control is done by the CPU itself and can vary between 800 and 1333 MHz depending of the CPU speed level, but as I said Apple has castrated it at 1066. They are not using the low level CPUs and make all their top level CPUs go slower to achieve this. Actually a silly move in my view.
Hmm... So what clock xeon would be wasteful given the memory limit of 1066Mhz?
 

VirtualRain

macrumors 603
Aug 1, 2008
6,304
114
Vancouver, BC
Get the fastest clock you can afford. Given the massive cache sizes used on current Intel processors, you won't be running into any memory bandwidth issues, especially with the tri-channel memory configurations.

For a primer on the Core i7 Nehalem processor family... have a look at this...
http://en.wikipedia.org/wiki/Intel_Core_i7

The current Mac Pro's use the 4-core Bloomfield chips (Xeon part #'s that are identical to the i7 desktop processors except for ECC support). Gulftown is the upcoming 6-core variation that are very similar to Bloomfield except for the added cores and additional L3 cache. They will use the "i7" branding as well.

All Nehalem's have an integrated memory controller (IMC) which eliminates the old FSB so the CPU cores have direct access to the memory controller now. The CPU also connect to the chipset via a much higher speed QPI link now, again removing any constraints formerly imposed by the FSB.

Hence, the only real bottleneck now is the clock speed and the parallelism... the latter of which we are still waiting for software to catch up.
 

nanofrog

macrumors G4
May 6, 2008
11,718
2
QPI has no limit at all, as far as I am aware. It's all processor dependent.
QPI has a throughput limit of 2GB/s (4x 500MB/s PCIe Gen 2.0 lanes). It's faster than FSB, and it doesn't have to carry the memory data either (PCI <used for USB, FW, and NIC chips in the MP; some PCI slots on PC boards>, PCIe, and SATA). DDR3 data is routed directly to the IMC on the die.

But likely you're not going to be able to stick a Gulftown in there anyway because of firmware issues. And all the current Xeons and Core i7s have the same memory frequency, iirc.
Not without new microcode to make it work, which Apple's not going to release to the '09 MP owners. :(

The memory is based on the BCLK frequency (133MHz) and a multiplier, just as the CPU is.

So are the Core i7's gulftown? Are you saying the Core i7's would not work? What are the best options?
Yes and no. They're apparently going to be marketed as i7's (i.e. i7-980 is one of the parts in the range that's been announced), but differ from the other LGA1366 parts in core count. Otherwise, the QPI, IMC, and other features are the same. The X58 (5520 for DP systems) chipsets will be recycled with the Gulftowns. That's why the boards will work, provided there's a firmware update with the proper microcode.

Hmm... So what clock xeon would be wasteful given the memory limit of 1066Mhz?
Of the current Quad core (35xx line), the 3.33GHz model (W3580) is capable of 1333MHz memory.

It's possible Apple fixed the limit, but as it ships with 1066MHz sticks, I actually doubt they did. :(
 

Eidorian

macrumors Penryn
Mar 23, 2005
29,081
287
Indianapolis
QPI has a throughput limit of 2GB/s (4x 500MB/s PCIe Gen 2.0 lanes). It's faster than FSB, and it doesn't have to carry the memory data either (PCI <used for USB, FW, and NIC chips in the MP; some PCI slots on PC boards>, PCIe, and SATA). DDR3 data is routed directly to the IMC on the die.
I think you're mistaking QPI for DMI.
 

nanofrog

macrumors G4
May 6, 2008
11,718
2
I think you're mistaking QPI for DMI.
Not really (LGA1366 only BTW). The QPI lanes attach to the chipset, then the ICH10R to the chipset via DMI. So it does actually reach the CPU via QPI, not DMI. You end up with DMI -> QPI -> CPU for SATA access (it's all bi-directional of course). But ultimately, it has to route through the QPI before it hits the CPU.

Unless you're talking about the LGA1156 parts that don't have QPI, and attach to the chipset via DMI.
 

Eidorian

macrumors Penryn
Mar 23, 2005
29,081
287
Indianapolis
Not really (LGA1366 only BTW). The QPI lanes attach to the chipset, then the ICH10R to the chipset via DMI. So it does actually reach the CPU via QPI, not DMI. You end up with DMI -> QPI -> CPU for SATA access (it's all bi-directional of course).
Then you're mistaken for the bandwidth available to QPI. It's either 19.2 GB/s or 25.6 GB/s depending on the processor.
 

nanofrog

macrumors G4
May 6, 2008
11,718
2
Then you're mistaken for the bandwidth available to QPI. It's either 19.2 GB/s or 25.6 GB/s depending on the processor.
We were both wrong here. Oops. :eek:

The X58/5520 chipsets work out to 12.8GB/s max (based on the 3.2GHz model).

Each X58 QuickPath Interconnect uses 21 unidirectional differential pairs in each direction, for a total of 82 pins per QPI. At the highest bandwidth, each QPI can transfer up to 12.8 GB/s usable in each direction simultaneously using the QPI protocol. The protocol transfers information in units of 80 bits (called "flits") which contain 8 bits of error correction, 8 bits of QPI routing information, and 64 bits of data.
Source.

With the newer part (3.33GHz and others used in MP's), the math works out to:

2.26GHz CPU = 9.04GB/s
2.66Ghz CPU = 10.64GB/s
2.93GHz CPU = 11.72GB/s
3.33GHz CPU = 13.32GB/s

What I don't get, is with even close to that bandwidth, why the ICH10R had to be limited to ~660MB/s. :confused:

I'm under the impression they fixed the band of the DMI sub-section in these chipsets (based on 6x mechanical drives @ 110MB/s each = 660MB/s).
 

PaulD-UK

macrumors member
Oct 23, 2009
83
1
What I don't get, is with even close to that bandwidth, why the ICH10R had to be limited to ~660MB/s. :confused:
Hi
I read a description about a year ago that the limit was deliberate because it stopped something further up the data path to the CPU falling over - because of a clocking problem or something...

I can't find the link, and after a year with my memory ;) I might not have the details of the problem correct.

Also I seem to remember a discussion about how the Nehalem QPI was currently implemented at 1066MHz, with only the external RAM access going up to 1333MHz.
 

nanofrog

macrumors G4
May 6, 2008
11,718
2
Hi
I read a description about a year ago that the limit was deliberate because it stopped something further up the data path to the CPU falling over - because of a clocking problem or something...
Clock differentials can easily be a problem (timing hell), but I'd have expected more. The fact the QPI's bandwidth is dependent on the CPU's rated frequency, it varies from part to part.

I am curious though, and will try to find the time to research this again (didn't find anything before, but its been awhile now).

Also I seem to remember a discussion about how the Nehalem QPI was currently implemented at 1066MHz, with only the external RAM access going up to 1333MHz.
:confused: I think you're getting crossed with Apple's memory clocks fixed to 1066MHz for the IMC in their implementation of EFI, even for CPU's that can use 1333MHz.

QPI is based on the CPU frequency (BCLK * multiplier = CPU operational frequency = QPI frequency). As a result, the max throughput varies per part.

Unless the information published by Intel is false. :eek: :p Or what I've seen is incomplete in terms of differences with the Xeon SP and DP parts vs. MP parts (i.e. QPI capable 75xx series).

nanofrog, should i go for a xeon or a core i7? if core i7 it can only be quad core right? If i am understanding correctly your saying it would need to be LGA1366 which on the intel Core i7 webpage says all i7-900 series.

http://www.intel.com/products/processor/corei7/specifications.htm
LGA1366 = i7-900 series (until the Gulftowns hit)
LGA1156 = i5-750 & i7-800 series

But I'm not sure what you're asking. So what exactly are you trying to do?
DIY a system, swap CPU's in a Quad '09 MP for a faster part,....?
:confused:
 

TallGlassOfFail

macrumors member
Original poster
Sep 16, 2009
43
0
Unless the information published by Intel is false. :eek: :p Or what I've seen is incomplete in terms of differences with the Xeon SP and DP parts vs. MP parts (i.e. QPI capable 75xx series).


LGA1366 = i7-900 series (until the Gulftowns hit)
LGA1156 = i5-750 & i7-800 series

But I'm not sure what you're asking. So what exactly are you trying to do?
DIY a system, swap CPU's in a Quad '09 MP for a faster part,....?
:confused:
As i said in the first post of the thread i want to swap out my existing proc that came with the machine for a higher clock one. I have the 09 mac pro. So yeah the latter "a faster part". But obviously i don't want it to be problematic, difficult, incompatible, no messing with firmware this and flashing that and i just want it to work with [no/minimal] fuss. Also don't want it to be wasteful given various bottlenecks and pound (£) per (x)Mhz that won't be fully utilised.

At current i have the 2.66Ghz Quad Xeon as evidenced below in my attachments:
 

Attachments

TallGlassOfFail

macrumors member
Original poster
Sep 16, 2009
43
0
One more thing, apart from clock freq, do the 3.33Ghz have anything different architecturally or otherwise from what i have (2.66Ghz) ? If not, could i not just alter the clock freq? and if so, how would i do this?
 

nanofrog

macrumors G4
May 6, 2008
11,718
2
As i said in the first post of the thread i want to swap out my existing proc that came with the machine for a higher clock one. I have the 09 mac pro. So yeah the latter "a faster part". But obviously i don't want it to be problematic, difficult, incompatible, no messing with firmware this and flashing that and i just want it to work with [no/minimal] fuss. Also don't want it to be wasteful given various bottlenecks and pound (£) per (x)Mhz that won't be fully utilised.

At current i have the 2.66Ghz Quad Xeon as evidenced below in my attachments:
The attachements weren't necessary. ;)

I'd keep to the Xeon 35xx family myself, though others have swapped them for faster i7-9xx parts. I'd recommend a search, as there's info in MR on this.
But it is possible.

One more thing, apart from clock freq, do the 3.33Ghz have anything different architecturally or otherwise from what i have (2.66Ghz) ? If not, could i not just alter the clock freq? and if so, how would i do this?
Within the Xeon family, the memory can run at 1333MHz on certain parts, but won't matter with '09 MP's, as the memory speed is fixed at 1066MHz in the EFI firmware for that system (1333 was tried, and failed to operate at full speed). :(

The i7-900 series vs. Xeon 35xx series, is the Xeons have ECC support enabled. That's it. Otherwise, they're the same part clock per clock.
 

TallGlassOfFail

macrumors member
Original poster
Sep 16, 2009
43
0
The attachements weren't necessary. ;)
Just wanted to show off my setup :p hahaha.

I'd keep to the Xeon 35xx family myself, though others have swapped them for faster i7-9xx parts. I'd recommend a search, as there's info in MR on this.
But it is possible.
Google product searched and found that there is not much in the price between the Core i7s and Xeons, sometimes even the xeons were more expensive.

The Xeons:
http://www.google.co.uk/products?q=W3580&show=li&scoring=p

The i7's: (first page featuring i7's after mobo results once in price order, lowest first)
http://www.google.co.uk/products?q=i7+975&scoring=p&show=li&sa=N&lnk=next&start=54&gnum=18


Within the Xeon family, the memory can run at 1333MHz on certain parts, but won't matter with '09 MP's, as the memory speed is fixed at 1066MHz in the EFI firmware for that system (1333 was tried, and failed to operate at full speed). :(

The i7-900 series vs. Xeon 35xx series, is the Xeons have ECC support enabled. That's it. Otherwise, they're the same part clock per clock.
Given that Core i7 is faster and apparently less expensive it would be the superior choice?
 

Umbongo

macrumors 601
Sep 14, 2006
4,929
54
England
Given that Core i7 is faster and apparently less expensive it would be the superior choice?
No, they are exactly the same speed. Every model of Core i7 has a matching Xeon 3500 part. The prices might differ slightly due to more consumer component demand for the Core i7 brand.