[Mech_E]

Hey all,

I was wondering the answer to this question:

All things equal what is the overall performance gain of using quad-channel DDR3-1866 MHz vs quad-channel DDR3-1600 MHz memory for CFD or any other simulation application?

Thanks for your thoughts!

MT

[evcelica]

Here is a CFX Benchmark run on 4 processor cores:
http://www.cfd-online.com/Forums/att....benchmark.jpg
looks like about 10% performance gain

[7sunnyshikhar]

can any one explain the best possible way to use ram.....

how many modules per channel...???

should one go for less number of bigger modules or more number of smaller modules.....for ex - 4*8 vs 8*4....

how does memory channel bandwidth gets affected on using more then one module per channel....???

does it any way differs how we will install ram modules in a 4 core 4 memory channel processor then in a 8 core 4 memory channel processor......i have this notion that since memory channels hold key, so increase in number of cores per memory channel wont give considerable boost....

[evcelica]

I don't think it matters very much whether you use 4GBx8 or 8GBx4 in a single socket system. But I've done some benchmarking and found 2 DIMMs per channel showed higher memory bandwidth. (everything else was not equal though, and these were dual socket systems)

Machine #1: Dual Xeon E5-2643: Memory: 16GB DIMM for each Memory channel.
ECC - 1600MHz
Aida64 Read: 74.538 GB/s
Aida64 Read: 54.918 GB/sec
Aida64 Copy: 63.733 GB/s
Latency: 79.0 ns

There could be other factors influencing these benchmarks though. I would just get whatever memory kit meets your needs, and not worry too much about 8x4 or 4x8.

Machine #2: Dual Xeon E5-2680:
ECC - 1600MHz
Aida64 Read: 90.019 GB/s
Aida64 Read: 81.766 GB/sec
Aida64 Copy: 83.443 GB/s
Latency: 82.7 ns