I am wondering what is more appropriate for the workstations I am building mainly for WGA and RNAseq of Eukaryotic genomes.
32-64 GB DDR4 RAM (2800 or 3000 MHz) vs 16 GB DDR4 RAM (3600 MHz)
Motherboard is supporting 128 GB max RAM. I will be adding more RAM at one point currently limited by the budget . Above is for the initial configuration.
Processor : AMD Threadripper 1900x
Storage : 2X 250 GB NVME and 2 X 1TB SATA III
GPU: NVIDIA GTX 1050i
Any other suggestions also greatly appreciated. (I wont be able to use AWS since bandwidth limitations at my University).
RAM speeds make incredibly negligible differences to overall performance. As others have said: get as much as you can. You'll be bottlenecked by I/O or CPU long before RAM anyway.
There is one thing you can't find a substitute for and that is enough RAM. If you don't have enough you simply won't be able to analyses you want, no matter what speed it is.
What's enough RAM though? This week I noticed that 256GB isn't enough RAM for e.g. indexing a large reference with default Bowtie2 settings. Sure I can increase offrate but it goes directly against my aims..
If you are going to be serious about assembly and RNA-seq, you must get 256GB at the very least. And those RAM clock rates are not important, that is standard 2400 MHz DDR4 modules will do just fine (AMD Zen and Zen+ CPUs are far more sensitive to RAM clocks than Intel CPUs because of the way their Infinity fabric works, though): you will be far more limited by single-core CPU performance, CPU-caches and IO. 2TB of long-term disk storage is not enough, either. You don't need a GPU, unless this machine will be used for machine-learning or nanopore base-calling.
P.S.
Your choice of CPU, RAM and GPU indicates that you want a desktop that will be primarily used for personal stuff (e.g. gaming), which is not particularly compatible with a research-oriented HPC. A research HPC can only generate value, if it is loaded 24/7 with research tasks. These machines use server CPUs, server RAM (ECC) and, as far as deep-learning is concerned, GPUs with large amounts of VRAM. You don't buy these machines for personal use. I believe you should rethink this purchase and find a way to solve your bandwidth issues instead.
WGA is whole genome assembly? Unless you are talking small genomes, even 128GB may be too few memory.
Describe in more detail what you will be doing, expected sizes of datasets, organisms you will work.
Part of the skill in using cloud services is finding how to reduce the bandwidth required.
RamRS may be knowing better.
RAM speeds make incredibly negligible differences to overall performance. As others have said: get as much as you can. You'll be bottlenecked by I/O or CPU long before RAM anyway.