Memory bandwidth question

[balamw]

I have a large memory intensive application that I need to run. (Unfortunately the app runs on 64 bit Windows. In case that matters).

I currently have a 2008 model Mac Pro 2x2.8GHz and a Dell Precision T5500 Nehalem populated 2xE5504s, but the Mac trounces the Dell in all the benchmarks I have run on it, mainly because of memory throughput.

From the Apple site it looks like the Nehalem machines can be a lot faster at least with benchmarks http://www.apple.com/macpro/performance.html a factor of 2 or even 1.5 would be very helpful.

So the question is:
How sensitive is the memory bandwidth to the CPU clock speed in the current MP family? Would the dual 2.26 perform as well as the 2.93 in the comparison when shoving tons of data in and out of memory. Has anyone run the STREAM benchmark on a stock 2.26 and a 2.93 GHz.

I think one bottleneck I have in the Dell is the processors, which only support running the RAM at 800 MHz instead of 1066 or 1333 MHz. The 2.0 GHz processors seem about as fast as a 2.4 GHz Conroe when all is said and done so I would expect roughly similar CPU performance to the 2.8 GHz box with the 2.26 GHz (maybe a bit less).

So I have a couple of options on the table.

1) Reproduce the Mac Pro from the current line and max it out with OWC RAM
2) Upgrade the CPUs in the Dell to take better advantage of the RAM.

Any opinions/input/comments/thoughts etc... are welcome.

B

 

[netkas]

memory has its own freq, and its not high for macpro

if u want some real fast memory, better use nehalem pc with 3x ddr3-1600 ram, anyway u need it for windows, not macos.

 

[balamw]

memory has its own freq, and its not high for macpro

if u want some real fast memory, better use nehalem pc with 3x ddr3-1600 ram, anyway u need it for windows, not macos.

Even if I wanted to build my own. Can you point me to a motherboard that can support 32GB of 3xDDR3-1600 and benchmarks to show that the memory is utilized to its full potential?

I have been looking and even finding one that supports lots of memory at 1333 is not so easy, plus is no guarantee that the throughput isn't bottlenecked elsewhere.

B

 

[Umbongo]

Do you have your memory optimally placed on the Dell? Even with 800MHz DDR3 should offer more bandwidth than the 2008 Mac Pro.

You can't run faster DIMMs for Xeons on any system/board I know of, so if you want a dual 2.26GHz you are looking at 1066MHz tops. A 2.66GHz+ setup would allow you to use 1333MHz DIMMS.

edit: There are a bunch of rules to keep in mind with Nehalem memory:

• The processor has to support the memory speed.
  
• Assuming dual ranked DIMMs (pretty standard):
  1 DIMM per channel = 1333MHz speeds
  2 DIMM per channel = 1066MHz speeds (Tyan have overclocking on at least some of their boards to run 6 per socket at 1333MHz)
  3 DIMM per channel = 800MHz speeds
  
• Unbuffered memory is quicker if you only use 1 DIMM per channel. Otherwise use Registered, which you may want anyway with large capacities. Non-ECC, again quicker, but you won't find Registered non-ECC so the same rules apply.
  
• Fill all channels with the same amount of DIMMs!

 

[gugucom]

Memory bandwidth should maximise with the W-spec Nehalems with six QPIs and 1333 MHz frequency. The Mac Pro currently doesn't run those processors and if you put them in the RAM will not run so fast.

http://www.asus.com/product.aspx?P_ID=sqbdCm0nmFxn3sS4&templete=3

This board with two W5590s on the other side will run up to 144 GB 1333 MHz RAM with six QPIs. It should actually be unbeatable by any Mac Pro.

The only standard feature it misses vs a Mac Pro is firewire. But with so many PCIe and PCI slots that shouldn't really be a big problem.

 

[balamw]

Do you have your memory optimally placed on the Dell? Even with 800MHz DDR3 should offer more bandwidth than the 2008 Mac Pro.

Yup. We tried every permutation in the book and then some, called Dell, called Crucial. They both say it's working as designed.

It just gets trounced by the Mac no matter what we try. It is quite sad, but for our task at hand the Dell T5500 (24 GB dual ranked DDR3 1333 from Crucial) just barely keeps pace with a PWS390 with a Conroe 6600 (8GB unbuffered DDR2 800 also from Crucial). Of course since the T5500 has more RAM it is more useful, but it ain't particularly fast for my jobs due the memory bottleneck and my 2 thread limit.

Memory bandwidth should maximise with the W-spec Nehalems with six QPIs and 1333 MHz frequency. The Mac Pro currently doesn't run those processors and if you put them in the RAM will not run so fast.

http://www.asus.com/product.aspx?P_ID=sqbdCm0nmFxn3sS4&templete=3

This board with two W5590s on the other side will run up to 144 GB 1333 MHz RAM with six QPIs. It should actually be unbeatable by any Mac Pro.

Would the six QPIs be effective if I am only running 1-2 threads? I believe the memory fill task in the code I am running may even be limited to a single thread which may be a bottleneck in and of itself. This is the kind of surprise I am trying to avoid.

Thanks for the pointer. I had not run across that one in a recent search. Do you have any personal experience with it? Any thoughts on the Z8PE-D12? It should be sufficient for my needs with 4GB sticks.

The main reason I am seriously considering the Mac for this again is that every time we build a box or even get one from Dell or another vendor there are surprises that take days to weeks to resolve, whereas the Mac Pro was useful 20 minutes out of the box. Time=$$$ and if it gives us a measurable performance boost so much the better.

B

 

[Umbongo]

Yup. We tried every permutation in the book and then some, called Dell, called Crucial. They both say it's working as designed.

It just gets trounced by the Mac no matter what we try. It is quite sad, but for our task at hand the Dell T5500 (24 GB dual ranked DDR3 1333 from Crucial) just barely keeps pace with a PWS390 with a Conroe 6600 (8GB unbuffered DDR2 800 also from Crucial). Of course since the T5500 has more RAM it is more useful, but it ain't particularly fast for my jobs due the memory bottleneck and my 2 thread limit.

I'm suprised. The T5500 bandwidth should be around 2.5 times that of the Mac Pro and should have less latency. Maybe it is the processors limiting it in some way, or some software issues, or something weird when it is downclocking the memory to 800MHz or what ever it does. There is obviously an issue present that by technical specs shouldn't be there. I can understand wanting to go to a Mac for the simplicity and headache free experience and might be worth trying. The option I'd probably first look at would be getting 2.26GHz processors and 1066MHz memory so I had everything as it was supposed to be (Mac, Dell or custom). Shouldn't make a difference, but the less variables the better.

 

[balamw]

I'm suprised. The T5500 bandwidth should be around 2.5 times that of the Mac Pro and should have less latency.

Yeah, that's how I sold it to my management and now having to go through the same thing over again I want to get it right.

B

 

[gugucom]

Thanks for the pointer. I had not run across that one in a recent search. Do you have any personal experience with it? Any thoughts on the Z8PE-D12? It should be sufficient for my needs with 4GB sticks.

I'm only going by specs here. I know that Asus makes good boards and if I were trying to max bandwidth I would surely be looking at that mobo.