fastest RAM configuration

Question

I am planning to buy a workstation for doing heavy image processing parallel computation (for 3-dimensional electron microscopy) under linux. It will have 64 cores (4 AMD opterons of 16 cores each). There will also be 3 GPUs (Nvidia Geforce GTX 780Ti). I want to fit 128 Gb RAM but I am not sure what is the best configuration in term of performance (speed). The planned computation involves quite a bit of data exchange between the CPUs/GPUs and the RAM. The workstation vendor gives three options:

1: 16 x 8GB 240-Pin DDR3 SDRAM 1866 Quad Channel Desktop Memory
2: 4 x 32 GB DDR3 SDRAM ECC Registered DDR3 1866 Quad Channel Server Memory
3: 16 x 8GB DDR3 SDRAM ECC Registered DDR3 1866 Quad Channel Server Memory

Which one would be the most performant? Thank you

Answer 1

I would definitely go with the fully populated channels for performance as already mentioned, though it will make an upgrade later more expensive. In terms of registered versus unregistered, from Wikipedia:

Normally, there is a performance penalty for using registered memory. Each read or write is buffered for one cycle between the memory bus and the DRAM, so the registered RAM can be thought of as running one clock cycle behind the equivalent unregistered DRAM

As you will see though (and as mentioned later in that article), things are not really that simple. There are times that unregistered memory in a multi-channel set up can see reduced memory bandwidth.

You also have to consider the reason why registered memory usually (not always) features ECC, so for accuracy considerations it may be beneficial, especially for lots of calculations. Without the ability to test I would probably default to the 16x registered DDR3 as a good compromise.

Answer 2

You have 4 CPUs, each with a quad-channel memory controller. So having the same amount of memory on each of your 16 channels would be best. Other than that, there's no speed information given for two of your options, so it's hard to compare them.