Quad Channel Memory question

[zombie67]

Lets use this Z10PE-D8 WS (https://www.asus.com/us/Motherboards/Z10PED8_WS/specifications/) mobo for this discussion. It says it is Quad Channel Memory Architecture. It has 8 DIMMs and 2 CPUs (if that matters).

Is there any benefit to running 8x8gb DIMMs vs 4x16gb DIMMs? Or any detriment? I recall hearing that more DIMMs may mean they might run at slower speed. But that just may be my crappy memory.

What if it was the 16 DIMM version (https://www.asus.com/us/Motherboards/Z10PED16_WS/specifications/) of the mobo? Any difference in performance with 16x4gb vs. 8x8gb vs. 4x16gb? They have a footnote that says "2133MT/s@1DPC only". But I don't know what that means.

[John P. Myers]

8 DIMMs are best for 2 socket boards. 16 DIMMs can slow you down. 4 DIMMs cuts your bandwidth in half.



1DPC means 1 DIMM per channel i.e. best to only use 4 DIMMs per Xeon.

[pinhodecarlos]

Here's a post from ATH where he actually compared using dual/quad channel on the same system. Speaks for itself:
http://www.mersenneforum.org/showthread.php?t=20575

[zombie67]

Thanks!

I looked at the linked thread. I have no idea how to read that table. What is a worker?

[pinhodecarlos]

Not sure how to reply to that but let me try to explain.
Lets think of a quad-core machine without HT, so it has 4 cores. Prime95 can use one task to be processed in parallel in all 4 cores or can run 4 single tasks in parallel (4 instances running in parallel). If I am not mistaken you can saythat a worker is the ID who processes the tasks using x number of cores. Worker would be the program itself.

[John P. Myers]

That table shows the iteration times per instance, which means you're looking for the lower times as being better. Running 1 instance(worker) across all 8 cores at 40M gives you about the same times whether you're using quad or dual channel. As you increase to 60M and 80M, you can see quad channel having definite advantages. This is similar to running 1 MT WU which uses all of your cores.

Also, when the number of instances (workers) are increased, you'll see quad channel as a very significant advantage over dual channel. Looking at the column for "8 workers 1 core/worker" you'll see it's about 40% faster on quad channel than dual. This is similar to running 1 WU per core, which is what we're almost always doing.

Like i said in my previous post, 4 DIMMs per Xeon is best (8 DIMMs on a dual socket board - meaning 4 per Xeon). Less than that reduces your bandwidth. More than that also reduces your bandwidth because the RAM speed is dropped (usually from 2133 down to 1866).

This table also applies to non-Xeon CPUs like the 5960X, 5930K, 5820K, 4960X, 4930K, 4820K, 3970X, 3960X, 3930K and 3820K - quad channel makes a big difference in speed.

[John P. Myers]

Also i should mention, a way to possibly save money on ECC RAM is, since you're only putting 1 module per channel, you can use RDIMMs which are usually cheaper than LRDIMMs. LRDIMMs are only usefull when using multiple modules per channel as they help reduce the RAM's speed loss, though it can still occur.

[zombie67]

Okay, similar question for older Xeons. Westmere-EP (https://en.wikipedia.org/wiki/Xeon#3600.2F5600-series_.22Gulftown.22_.26_.22Westmere-EP.22), specifically. Apparently they have channels in sets of 3, if you can believe it.

Normal memory speed is 1066 or 1333. I think that drops down to 800 if you have 2 DIMMs per channel. Please tell me if I have any of this wrong.

So here is the question: For BOINC, is the memory speed difference significant with these older CPUs?

Edit: The difference in cost for a T7500 w/ 48gb (12x4gb) vs. T7500 w/ 24gb + 6x8gb kit = ~$30 dollars.

[John P. Myers]

DDR3 @ 800MHz is gonna hurt noticeably. DDR2 ran faster than that (i still have 2 sticks of 1066MHz DDR2) :p

The speed your RAM will run at is dependent upon the speed of the DDR3 clock of the Xeons installed. If you have 1333 MHz RAM, but E5640 Xeons, your RAM will only run at 1066MHz (disregarding the channel loading). Can probably get a pair Xeons off eBay for $50 that'll support 1333MHz so if you did 2 DPC it would only drop to 1066 instead of 800. Still sucks, but not as bad.