Quad channel memory sticks for Threadripper

Depends.

Dual-rank modules tend to work a bit faster per given clock speed, but only a bit, and you'll likely only see the benefit in benchmarking, as the CPU can have more open pages on a dual-rank modules (doesn't mean it will.)

Single-rank modules tend to play better with AMD's current memory controller in their Ryzen / Threadripper CPUs, and as such tend to reach higher clock speeds, giving you better bandwidth and Infinity Fabric speeds.

The more RAM channels filled, the more read / write requests can be carried out, but I think you already understand that. Since Threadripper supports quad-channels, you might as well do that, otherwise the extra expenditure for the CPU might be called into question.

As far as balancing the modules in the slots, why wouldn't you? And, why wouldn't you be looking at quad-channel kits? And, you should be looking at the kits listed on the QVL list for your chosen motherboard, so you know what to expect as far as RAM performance is concerned.

4 modules VS 8 at this point, I would personally start with 4. CPUs tend to be pickier with memory modules the higher the module count, as this puts a greater strain on the CPU's memory controller. However, I do believe Threadripper supports ECC Buffered / Registered modules, at which point you should be fine with 4 or 8. Speed should be the same between 4 or 8 modules, provided each of the 4 modules occupies it's own channel.

Were your mac memory experiences tested, or just anecdotal?

16 GB seems a small amount to be equipping such a wide CPU with. I would look at 32 GB at a minimum, 64 GB or more possibly. It's certainly not going to be cheap, but cheap is not a direction a Threadripper build should go anyway.

Start with the CPU and motherboard, obviously. Decide on the motherboard based on the feature set you will use. Then, once you have chosen your motherboard, read through the QVL list for the RAM and pick something out of that. I wouldn't waste 4 slots to get 16 GB RAM. I might not even want to go with 4 x 8 GB to hit 32 GB either, unless it's going to be a long time before you upgrade to the other 32 GB. Filling 4 slots now, gives you only one more possibility for any upgrades without replacing modules.

I really don't have any recommendations for motherboards here. I have no plans to build a Threadripper system, and all of the boards I've looked at are more than adequate. It's not a segment of the market you're typically going to see a lot of corners cut, so just find the board that suits your needs.

You shouldn't be looking at specific memory modules until you've chosen your platform. You can look at it in general, but every motherboard vendor is going to have a QVL list of memory modules you should be choosing from.

You really don't need to look further than the QVL list unless you need something that isn't on the list, or you can't find the modules that are on the list, such as with older memory modules.

As far as specialized modules for AMD VS Intel, no, as long as the modules have voltages and timings that are compatible with the particular CPU's memory controller that you choose, and conform to JEDEC specifications, you can consider them reasonably compatible. If you're concerned with compatibility, again, stick to memory on the motherboard's QVL list.

To answer your second question first, yes, it is recommended to purchase all modules in kit fashion, be it 2 modules, 4 modules, or even 8 modules. That gives the highest assurance of modules that are closest in performance, giving the highest likelihood of working together correctly.

To answer your first question, not necessarily. It depends on what was meant. Memory modules operate at the speed they are set to run at, no faster, no slower, whether you have 1 module or 8 modules.

Where you get into speed differences is going to be both motherboard and memory controller specific. Some memory controllers just won't run a high number of modules, or banks, or won't run them at the highest speeds the memory controller may be rated for. More chips on more modules usually means more strain on the memory controller, unless we're talking about Buffered / Registered RAM (which is why it exists.) This is why reading through your motherboard's QVL is important when choosing your RAM modules. The motherboard manufacturer has gone through and taken the time to test different memory configurations for compatibility. Instead of buying kits in the hopes that they may work at their rated speed, the manufacturer has done this for you and written down the results.

In the end, if your memory runs correctly at it's rated speed, it doesn't matter how many memory slots you use. You're going to find that both ECC and Buffered / Registered RAM isn't as fast as non-ECC, so you're not going to be breaking any speed records here.

The other thing you're going to find is, the high density modules you are looking at to hit 32 GB or 64 GB don't seem to want to run at the highest speeds with Ryzen memory controllers. Can't speak to how they will work with Intel, but I doubt it's going to be significantly better there. If I were planning to have 64 GB or more RAM in a workstation or HEDT level system such as what you're talking about, I wouldn't be looking at high speed gaming RAM modules anyway.

Your upgrade possibility make it harder.

Depending on whatever the RAM is non-ECC or ECC each rank is 64 or 72 bit, since the CPU have four channels and likely 8 memory slots if you just use four memory modules of single rank ram you just have one rank / channel, if you instead use dual rank modules you'd have two / channel, if you used 8 single-rank modules you'd also have two / channel, if you used 8 dual-rank modules you'd have four, if it supported quad-rank modules and that was something you'd get four of you'd still have four / channel and if you got eight of those you'd have eight / channel.

You can read about memory ranks here:
forums.crucial.com/t5/Crucial-memory-for-PC-systems/What-is-a-memory-Rank/ta-p/126346

AMD have their own memory guide for gaming:
https://community.amd.com/community/gaming/blog/2017/07/14/memory-oc-showdown-frequency-vs-memory-timings
AMD specific:
1) If the RAM can run at command rate 1T then GDM / Gear Down is likely better off. Otherwise on.
2) BGS happened to be better on with game, off with benchmark.
Memory:
3) Dual rank AT SAME CLOCK SPEED AND LATENCY is faster, because the second rank can perform work while the first one refreshes, however running more ranks is harder for the memory controller and as I've seen on AMD pages before what I really wanted to show is that with more memory modules or dual rank you're less likely to hit the higher frequencies. And hitting the higher frequency is likely better than having dual rank.
4) Optimized timings is better than auto.. Duh..
5) In their case the best stable timings (low cas latency * memory period) won over the best stable frequency (with loser timings.)

Here you can find the post about number of memory modules and ranks and what speeds are supported:
https://www.anandtech.com/show/11244/the-amd-ryzen-5-1600x-vs-core-i5-review-twelve-threads-vs-four
2 DIMM single-rank: 2666 MHz.
2 DIMM dual-rank: 2400 MHz.
4 DIMM single-rank: 2133 MHz.
4 DIMM dual-rank: 1866 MHz.

So as you can see for memory frequency using as few modules with only single-rank is best. And if you got the choice of using fewer memory modules but dual-rank memory to reach the capacity you want vs using more memory modules of single-rank then using fewer memory modules of dual-rank is better. The more stuff you put on the memory controller = the worse.

So the optimal configuration with 16 GB of ram is to use four 4 GB SR modules to put one rank / controller, preferably as fast as is supported too (where according to the test AMD did for Ryzen timings matter more than clock frequency but you shouldn't then just look at the CL value since that's _CLOCK PERIODS_ not nano seconds, to get the actual time you have to multiply that with 1/frequency (period), I would just do 18/3466 = 0.0051933064, 16/3200 = 0.005, so the 3200 MHz CL16 would have just a tiny bit better timing than the 3466 MHz CL18, but they are of course much closer than what CL16 and CL18 alone suggest.

I'd suggest to not use 8 memory modules and even less so 8 dual-rank memory modules if you don't have too.

Now the problem with your question become that since you want 16 GB now and maybe 32 GB later if you use four modules now you get the best performance now because you get quad-channel support and just one rank per channel, however if you then add four more modules to get 32 GB rather than get a new kit of four sticks you suddenly load each channel with two modules likely making it harder for it to run at the same performance.
On the other hand if you just get two 8 GB modules now you'll just get dual-channel which will give you less bandwidth than what quad-channel would give you, though when you later add two more 8 GB modules you'll still just have four modules giving you the best performance then!
Here's someone testing quad vs dual channel:
https://forum-en.msi.com/index.php?topic=291116.0
He say blue is dual and red is quad but if you look at the charts the number for the runs are switched over in the memory chart plus he say his dual-system performed worse so the colors doesn't match up with what he claimed there and the outcome is that the memory benchmark ran better with quad-channel memory and the CPU benchmark did too!

So you should try to get quad-channel.

As such in the end I guess my recommendation (for performance really..) would be to well, preferably get 4 x 8 GB SR sticks right now if that was an option though RAM are expensive now or get 4 x 4 GB SR sticks now and later when you want to upgrade sell that kit and get 4 x 8 GB SR then, if you don't want to sell any and buy something new I assume getting 4 x 4 GB SR now and upgrade to 8 x 4 GB SR later may be the better alternative, that way you always use quad-channel but may lose a bit of frequency after the upgrade. Going 2 x 4 GB SR now and upgrade to 4 x 8 GB SR later would mean you'd have to use dual-channel now which would lower your performance.

He had a 16 core chip, with 8 core maybe it will matter less because less cores may need less memory bandwidth.

2 = single-channel = very slow.
1 = dual-channel = still slower than necessary.
4 = quad-channel with 2 modules / slot = you may drop some frequency here (also in case of 8 x 2 GB you have no room to upgrade!)
3 = quad-channel with 1 module / slot = the best configuration.

For upgradability 3 will become 4 and 1 will become 3 when you upgrade if you add more modules =P