SPCR

Thanks for the analysis.

Thanks for the in-depth analysis. I've been reading on a few sources for the same comparison and they seem to conclude that DDR4-2800 is the sweet spot in terms of performance VS value. Anything over that is too expensive for the marginal gain.

Memory pricing is pretty darn flat, unless you go for some of the more outrageously marketed parts. For grins, here's what pcpartpicker had for lowest pricing from Kingston, G.Skill, and Crucial for 2x4GB:

2133/15 $48
2400/15 $50
2666/15 $59
2800/16 $62
2800/15 $63
3000/15 $65
3200/16 $65

Biggest jump is from 2400 to 2666.

Thank you! Great article, and strongly reinforces what I had read elsewhere (while adding something I hadn't read before -- that if using the CPU's integrated GPU, speed matters).

One thing that wasn't mentioned (probably because it was outside the scope of the article) is what difference does it make, if any, going with a 2x configuration rather than a 4x. For example, for 32gb you can get (and pay a hefty premium for) 2x16gb modules, or you can use 4x8gb modules. I've read some older posts which say that a 2x configuration is "better" (perhaps performs "better", perhaps just the memory controller stays a bit cooler), but can't find anything that applies to Z170. Does anyone have any thoughts?

For most applications, I don't think it matters. Here are some of the subtleties:

- dual channel memory controller with four RAM slots means each channel controls two of the slots.

- RAM timing is affected by trace and component capacitance. More load capacitance may mean the RAM has to run at lower speed/looser timings. Go back several generations and you may see where the mobo manual spells out max RAM speed based on the slots filled. These days, the memory controllers are pretty rugged and it may only affect you if you are trying to OC/run very fast/very tight timings...and while higher capacity RAM modules may have higher load capacitance, one 8GB module is still going to be less than two 4GB modules.

- Most people only need 8GB of RAM. So, the cheapest way to go about it is 2 x 4GB. If you need an additional 2x4GB, it won't affect you (again, unless you are a RAM OC'er). Same philosophy applies if you think you need 16 or 32GB.

Thanks, Steve.

In my case, I do need 32GB, however at present 2x16gb modules are very scare and much more expensive than 4x8gb modules. I definitely have no interest in OC'ing the memory, only the CPU, and thanks to your article, I'll get whatever speed is on sale. I just wanted to make sure that there's really no advantage to spending extra to get those expensive 2x16gb modules. Sounds like the only benefit might be if I ever wanted > 32GB, however I've read many times that trying to match an additional pair of modules to an existing one is asking for trouble. (Though, again, I have to wonder if that advice isn't for older chipsets with quad channel controllers, rather than the dual channel of the Z170.)

Anyway, sorry for getting off-topic, and thanks again for your article, Lawrence.

Ken5d wrote:Thanks, Steve.

In my case, I do need 32GB, however at present 2x16gb modules are very scare and much more expensive than 4x8gb modules. I definitely have no interest in OC'ing the memory, only the CPU, and thanks to your article, I'll get whatever speed is on sale. I just wanted to make sure that there's really no advantage to spending extra to get those expensive 2x16gb modules. Sounds like the only benefit might be if I ever wanted > 32GB, however I've read many times that trying to match an additional pair of modules to an existing one is asking for trouble. (Though, again, I have to wonder if that advice isn't for older chipsets with quad channel controllers, rather than the dual channel of the Z170.)

Anyway, sorry for getting off-topic, and thanks again for your article, Lawrence.

Actually, I've heard the opposite, i.e. the newer RAM is harder to match, I think they said because of the higher frequencies and/or lower voltage.

Thanks for the laborious testing, Lawrence! Yours is the first Photoshop scaling test that I have seen using the new z170 chipset. Excellent! Did you by chance do any testing of Photoshop with the IGP?

It's hard to know how the Hardware Heaven Photoshop test applies to the average Photoshop user, but judging from all the other tests, it's unlikely to be different. You may want to consider substituting stitching photos for panoramas +/- assembling HDR's instead of the blur filters and such used in the Hardware Heaven test for future testing, though of course it may not make one bit of difference in the results. More testing always leads to more questions! Thanks again for all the hard work!

Excellent article. Given how important system RAM is for iGPU performance, it'll be interesting to see what happens when developers really start to sink their teeth into DX12. If they can make good on Microsoft's promises of efficient, concurrent use of disparate GPUs, system RAM might become far more important for 3D application performance, even in systems with dGPUs.

Thanks for the tests. It's interesting that unlike DDR3 scaling a couple generations back, not even memory-intensive apps really scale with the more money you pour into the memory sticks.

DRAM timing specs are in cycles - not nanoseconds. You need to do the math to figure out if 15 on DDR4 is more latency than 9 on DDR3.

At 3000 MHz, 15 cycles is 5ns. At 1600 MHz, 9 cycles is 5.6ns. The DDR4 has less latency.

Any reason why a speed such DDR4-2666 CL15 wasn't benchmarked as well?

As CA_Steve's pricing a few posts up shows, while pricing is pretty flat, there is a bit of a jump, and I know if you're seeking optimal value then the numbers might be useful.

OTOH, if pricing for memory faster than DDR4-2666 really is flattening out that much (even in the past few weeks, availability's gotten a lot better) then I can see how it might not be all that crucial to test in some eyes.

Ken5d wrote:One thing that wasn't mentioned... is what difference does it make, if any, going with a 2x configuration rather than a 4x ... Does anyone have any thoughts?

You do have the comparison of 2x8 vs 4x8, and the difference is as good as negligible.
(Whether the little difference present is due to more memory or more channels isn't clear though.)

continuum wrote:Any reason why a speed such DDR4-2666 CL15 wasn't benchmarked as well? ...

There's obviously no reason to do it, given the tiny difference in performance between 2133 and 3000.
2666 can be expected to fall in the middle between them.

I'm more curious about how much is lost when using even slower DDR3L memory. At what amount of memory and speed does performance take a dive? How about using 2x4GB instead of 2x8GB? What about the laptops sold with only 1x4GB or 2+4GB of RAM?