Mamory bandwidth comparison
 

(I have just realized, that it is my first post on cfd-online since about 7 years...:)

My question concerns comparison of two ancient workstations in terms of memory access performances on "theoretical basis". Which configuration would be faster assuming all else being equal:

1.
2x Xeon X5650 (4 memory channels)
3 memory channels per proc. board
2x 8Gb dimm per processor

2.
2x Xeon X5675 (4 memory channels)
2 memory channels per proc. board
4x 4Gb dimm per processor

Lets assume that model consumes 20Gb of RAM.
I cannot give a guess: as far as I understand the theory, optimal configuration would be to saturate all memory channels of motherboard with dimms with 1 dimm/1 memory channel. From the second point of view, I am not sure what could be influence of lower number of motherboard memory channels per single proc. in configuration 2.
A bit of explanation: the justification is needed in purpose of deciding which one of these 2 configurations to buy for cheap, so I have to take a proc. speed into account;
I realize that for more telling comparison, assuming the same processors woule be better.

Thanks,
Luk

I actually had to look this up, that's how old these CPUs are ;)
Both have triple-channel memory controllers:
https://ark.intel.com/content/www/de...intel-qpi.html
https://ark.intel.com/content/www/de...intel-qpi.html

Assuming you run memory at the same transfer rate on both systems, I think the difference in theoretical maximum memory bandwidth is negligible.
Option 1 gets you two memory channels due to only using 2 DIMMs per CPU
Option 2 gets you two memory channels due to only having two channels per CPU available on the board.

At least with option 1, you could just drop in two more DIMMs. They are like 10$ each.

Thanks for rapid answer,
(You are right, my mistake - was looking at exactly the same intel page for proc. specs, but somehow I have stuck with convcition about 4 memory channels per proc.).


Ok just to clarify: lets assume now, that the model is bigger and consume 32-48Gb of RAM. We are free to add/reconfigure RAM, money aside.
In such case:
Option 1 gets some advantage unchoked by running 3 memory channels after adding 1 dimm per proc.
Option 2 does not speed up no matter what because motherboard 2 memory channels per proc. is still a bottleneck.

Right?
Thanks in advance.

Luk

Exactly. If you still want to spend money on these systems, do it for option 1.
Maybe swap CPUs around, the X5675 is slightly faster.
You can write me a PM if you are in Europe, there are loads of 16GB DDR3-1333 reg ECC modules in my basement.

What a provocation, sir!:)

(Will remember I am right beyond your western border).
Thanks a lot.


Luk