8GB->16GB...use 2GB or 4GB DIMMs?

[CaptainJeff]

I have a 2009, octo-core, 2.26GHz Mac Pro (MacPro4,1). I bought it with 8GB of memory, which came with 4x2GB modules, two installed in each "side" of the machine. I'd like to bump this up to 16GB total.

Looking at OWC, I can buy 4 2GB modules of 1066MHz ECC RAM for $145.99. I can also buy 2 4GB modules of 1066MHz ECC RAM for $146.99. So, I can pay one whole dollar more and get the 16GB I want with two extra slots for adding more memory in the future. This seems like the way to go except that I would then be running with 4x2GB and 2x4GB modules. Will that work? Will I take a performance hit? In what way should I install this in order to maximize bandwidth of the triple-channel memory controller?

Thanks!

 

[oculus42]

You need to have only three slots per processor filled to get Triple Channel performance. Filling all eight will drop you to dual-channel mode (3 and 4 share a channel). Details here: http://manuals.info.apple.com/en_US/Mac_Pro_2009_2010_Memory_DIMMs_DIY.pdf

Erm, more to come, accidentally submitted this early.

I would recommend the two sticks of 4GB each to give you the best capacity and performance gain. That gives each processor 8GB (2x2GB + 1X4GB) with which to play. Apparently poor slot selection will trigger a Memory Slot Utility (Buried in /System/Library/Core Services) to tell you what you should do.

 

[Hellhammer]

It will work. As you have unequal amount of modules, just install the 4GB modules to the remaining slots. If you had 3x4GB and 3x2GB, then you should be able to take full advantage of triple-channeling. Triple-channeling only helps in VERY few apps so I would not worry about that. It's very unlikely that you would notice any difference. Besides, having 2x2GB+4GB per CPU might mean that the first 6GB (i.e. 2GB per module) will run in tri-channel and the last 2GB would run in single-channel. Nobody seems to know how current memory controllers really work so to see the difference, you would have to use some benchmarking tools.

Here are some benchmarks