Consumer RAM for NGS analysis

[JamesH]

Now that 8GB sticks are available, and you can get socket 2011 motherboards with 8 DIMM slots, here is what I'm thinking:

G Skill 32G(4x8G) DDR3 1333 PC10600 x 2 = $590

Intel Core i7 3930K Processor LGA2011 = $699

Gigabyte GA-X79-UD5 MB, Socket 2011, Intel X79, 8x DDR3, 3x PCI-E3.0 = $348

Total = $1637 to upgrade to a 64GB system

Is there any reason not to do this?

Does the lack of ECC actually affect NGS analysis packages? If so would you notice it or would you get errors in your data that go unnoticed? Does the speed of the RAM have much of an effect?

[Crypticfortune]

I can't think of any compelling reason *not* to. That's a solidly "desktop" oriented motherboard, and most desktops don't use ECC anyway. Yes, without ECC any memory errors are silent unless they happen to occur in some system critical area, and yes memory speed is frequently the I/O bottleneck for NGS applications (that and disk, depending on what you're doing). Memory errors are rare enough that personally I think it's safe enough not to think about them for most computing. For practical purposes, if an error occurs, it'll likely crash your program when some pointer gets screwed up or cause some blatantly impossible result...but if the wild chance that it actually causes some silent error that doesn't crash your system and passes as legit data gives you the heeby-jeebies, then it's time to spring for ECC. This PDF has some nice gritty details about exactly how rare and why, but you're likely looking at about 100 days MTBF.

[Richard Finney]

Sounds good.
Make sure your memory is compatible with your motherboard. Find manufacturers site that says "yes, this memory X works with motherboard Y". Don't guess.

Can you even get ECC memory for this motherboard?
If you could you might spend a little more for it.

You will want to run a serious memory test now and then to make sure the memory is
working if you're flying without the an ECC safety net. Especially when you first set it up.
If a bit breaks, you might not notice bad results. Opinions are mixed on ECC vs. non ECC; servers typically have ECC but desktops don't. I crunch numbers on both types and haven't had any memory problems on the desktop/workstation (that I know of). Since you're not running a bank and your results won't send a space probe to the wrong planet; you'll probably be okay.

Memory speed and cache coherancy is the critical item for fast execution these days.
The CPU is waiting for memory. Tha'ts one reason CPU clocks speeds have flat-lined the last 5 years or so.

[JamesH]

Thanks guys!

I don't think you can get ECC on that MB, main reason to get it was the 8 slots! Sounds like it should be OK to use non ECC, I guess few of the folding overclockers are using ECC http://www.overclock.net/t/486609/gpu-milking-machine!

Would MEMTEST be suitable to test the RAM? Do you think getting 1600Mhz rather than 1333 would be worth it for an extra $100?

[mschatz]

More RAM is great for almost every application, but that still won't be enough RAM to assemble genomes larger than ~100Mbp.

Mike

[chadn737]

More RAM is great for almost every application, but that still won't be enough RAM to assemble genomes larger than ~100Mbp.

Mike

But its more than enough for the majority of NGS applications. For everything from re-sequencing, most RNA-seq, Chip-Seq, exome sequencing, etc.