What makes good performing memory?

[ksv]

I'm about to get some memory for my new home workstation (at least, I guess that's what one calls it) but am curious about what factors differentiate the different types.

I am putting an E3-1240 v2 on a Supermicro X9SCA... I was looking for the DDR3-1600 support as I figured that should max out my use of the machine (ab initio calculations and various simulations, BOINC, etc.). Two things are hitting me though: the relatively few matched modules that seem to be available (may just be a lack of time looking), and especially a comment by mobilenvidia in another recent thread. That person's comment woke me up to the effect of the timings: a latency of 11 vs. 9 for DDR3-1333 modules could hit the performance enough to significantly alter what I thought would be a reasonable jump in speed.

Perhaps then the DDR3-1333 is the best choice until things improve technically?

I'm basically self taught in understanding this area (my specialty is chemistry), and wanted to toss this out: what else goes into making memory work well? I'm sure there are other factors, and was hoping people would be willing to enlighten me...

(BTW I'm glad that I found this forum. I've read through a lot of posts, learned a lot, and expect to learn more! Thanks for any help!)

 

[Patrick]

Just as a thought, I spoke with a few memory makers back in March when the Xeon E5-2600 was launching. The general consensus is that with 1600 CL11 the vast majority of customers would benefit over 1333 CL9 and they were not planning to put out 1600 CL9 in the near future (ECC because you can easily get non-ECC at those speeds.)

 

[ksv]

Just as a thought, I spoke with a few memory makers back in March when the Xeon E5-2600 was launching. The general consensus is that with 1600 CL11 the vast majority of customers would benefit over 1333 CL9 and they were not planning to put out 1600 CL9 in the near future (ECC because you can easily get non-ECC at those speeds.)

Well, I went with the X9SCA MB instead of the Asus P8B because of how transparent the Supermicro board's capabilities were, as well as their very obvious support of Linux. Yet were I to back up to the P8B, I could go with "higher performing" non-ECC memory, with latency timings in the range of 7-8. While I understand the value for ECC, how often do errors occur? There must have been *some* study of this sort of thing.

Do you know whether it's as simple as that comparison, or should I also look at something else when searching for the best memory? That's why I'm trying to get a grip on whatever intricacies exist. I don't know what else to look at, though, and whether this by itself is appropriate.

Part of the reason for my building this rig is to look at what I might try to get for our school in the near future (I teach at a community college, and would like to pull in some computational software, some of which would thrive in a cluster environment, as well as individual workstations).

 

[Mike]

There's nothing wrong with finding the best bang for your buck but memory performance is about the last of your worries with the latest cpus by intel IMO. If what you do requires huge amounts of memory bandwidth you can go up the line and get a Xeon E5 x4xx or x6xx for respectively tripple or quad channel. From what we're seeing is that the RAM with the lower timings often do not live up to the price as the performance differences are rather small or sometimes non-existent in benchmarks. On the E3 platform you're also pretty much limited to ECC Udimms if i'm not mistakin'. That would leave you with not that many vendors and thus options afterall anyway. Also I believe the Ivy Bridge cpus are the first to officially support PC-12800 RAM so it may take a while for those to pop up too.

 

[mobilenvidia]

The 1600Mhz will be faster than 1333Mhz even with the latency being higher, but the gain will be small

I"ve opted for 1333Mhz ECC for my HTPC build as 1600Mhz was not readily available at the time.
But I would have gone 1600Mhz other wise (at same price)

 

[ksv]

I appreciate you folks' thoughts about this. There does not seem to be a lot of dual channel sets of 1600 memory out there at this point. I saw some Kingston ValuRAM at $125-135 for 4x4 Gb, and somewhere else a 2-module set, but it's relatively sparse. Newegg has the 1333 stuff from Kingston (ValuRAM??) and Crucial (ea. 2x4 Gb).

I'd be interested in running the two types against one another in identical calculations. Since I plan to be using an SSD I'm guessing the memory effects will become more apparent (ab initio calcs are storage intensive). If I can RMA the losers back to their store then I'd be happy.


What does having both banks full in a dual channel setup do to performance? I.e. is 4x4 better than 2x8??