I hear what each of you are saying: "Geekbench only shows a slight slowdown when DIMM sockets are left empty, so it won't have a large impact in the real world."
This perspective is rational and has merit if Geekbench's top line numbers can be generalized to general system performance.
Unfortunately, they of not generalize.
In fact, it is worse than that. The design of Geekbench and the sole purpose for which it was built in the first place, ensure that it is a very bad proxy for measuring the influence of memory on system performance.
Maybe I'm missing something, but looking at the final geekbench numbers in single and multi-core looks like there is only a 3-5% difference in 2 vs. 4 channels used.
Clearly in memory performance there is a difference.. but it looks like its about a 4% difference conceivable by the user.
That comparison clearly shows that it the additional DIMMs don't matter for most of the tests - just a couple.
Those tests are with a dodeca system, a quad would show even less difference.
Each of you claim that system performance will only be impacted by a few percent by under-populating the memory channels. You base this on Geekbench gross scores. The problem with this is that Geekbench is a CPU level benchmark not a system level benchmark.
What does it measure?
It measures only three classes of thing.
It measures Integer performance, Floating point performance, and Memory performance of CPUs in isolation from the rest of the system.
The Integer and Floating point groups of tests are designed to fit entirely in on-die caches and never touch main memory. These tests together assess the raw number crunching and processing power of the cpu under test. None of these tests read or write data during operation.
The Memory performance section gives perspective on how fast the CPU can perform calculations that rely on external data. Some of these tests also perform significant computation, but each of the algorithms have dependencies which are designed to miss the on-die cache and force data to be read or written during operation.
As you can see, the bulk of the tests performed by Geekbench are specifically designed to avoid touching memory in any way. They are entirely irrelevant to assessing the memory subsystem because they were designed to be irrelevant. Any off-chip access would page fault, stall the thread, and destroy the thing being measured (oops).
Only the memory performance subtests touch memory in any way during operation. Thus, they are the only ones which are relevant to compare when evaluating different memory configurations.
Can you see that?
This system has 4 separate memory channels between DIMMs and the memory controller.
By interleaving data addresses across channels, access to and from memory can occur in parallel on 4 sets of data lines.
The controller can handle 2, 3, and 4 way interleaving limited only by how many channels contain matching DIMMs. DIMMs match by sharing the properties {RDIMM/UDIMM, voltage, speed, size, and maybe internal rank}.
So by choosing to install 1 DIMM you are limiting memory bandwidth to 25% of optimal performance. Install 2, you get 50%. 3 = 75%.
Back to Geekbench.
The memory test subsection posted here, clearly shows that there is a difference. We expect bandwidth to suffer with 2 chips compared to 4 chips. The number of channels has been cut in half so have 50% fewer lanes involved. There is latency, setup, and prefetch queuing in the mix, and the algorithms perform nontrivial work on the data as well so we expect to see less than 50% and do see a 40% difference.
Does this mean that the general system will be 40-50% slower? No. It only means the memory subsystem will be up to 50% slower when actually transferring data between the memory and controller. This does not happen at peak continuously. But it does happen from time to time. Sometime as frequently as 1.866 billion times per second.
Every time that memory is read or written on your system it will happen on a data path that is determined by how the data is interleaved. By choosing to use fewer channels you set a lower bandwidth.
If you do plan on 16GB and still need a number around 4% to compare with something, how about this one? 100 bucks is less than 4% of the base model sticker price. You'll have your 16GB and it will boost your memory bandwidth by 33% in the process.
