My PI and I have moved to a new institution and I've been assigned with acquiring a lab PC/workstation. I've only been given about $1900 to work with, but this can likely be flexible.
I'm wondering what the current bioinformatics/comp biology opinions are about the specs for each component (i.e., amount/speed of RAM, number of CPU cores, CPU clock speed, etc). I have experience with PC components and building PCs but haven't done this for 5+ years now. I would also like to run Ubuntu on this machine.
The PC will mostly be used for processing things like RNA-seq and single-cell RNA-seq data, training ML models for this data, and image manipulation. Our old rig was slow and frequently ran into problems due to insufficient memory (32gb of RAM).
There was just recently an update to an old thread in this forum, but the discussed options are clearly above what your PI can/wants to spend and also possibly what you need. I think moving your workloads to the cloud (e.g. with Nextflow/nf-core) might be a better alternative, in particular when you are not running heavy analyses too frequently?
As you can probably surmise from the responses so far, that's not a lot of money and large RAM is the most critical component.
I will offer a different take, which is to buy a used/refurbished computer with better specs than can't be obtained when buying new computers. About four years ago I had $5K to spend on two computers, and bought two workstations, each with 12 CPUs, 64 Gb RAM and NVIDIA GPUs. They are still in use, but there are many things that can't be done with that kind of memory.
Two years ago I had $4K to spend and was tempted to buy a single computer with similar configuration as those two above but with more memory. Instead, I bought two used HP Z820 Workstations (Xeon CPU E5-2680 v2 @ 2.80GHz with 20 cores and 40 threads). The kicker is that each has 256 Gb of RAM. These are 7-8 years old by now and have relatively slow CPUs and RAM, but 40 CPUs will still get the job done faster than 12 CPUs in the above configuration. I have had only one problem in the past two years where the 256 Gb RAM was a bottleneck. Both of them still run great after two years despite being on and doing something all the time.
I don't want to leave the links here, but these computers are easy to find in the refurbished section on Amazon. In fact, there are 128 Gb configurations that are not much above $1K.
That is a creative/fresh take and one that can be considered if the institution allows for purchase of refurbished/off-lease equipment. Some places may not. If computer equipment does not malfunction within the first 2-3 months it is likely to keep working for a while. So as long as some limited warranty is included it may be enough to cover that critical period.
Most of them have reasonable warranty both from the original seller and by Amazon. What I like is they are true heavy-duty workstations, and I mean heavy in figurative and literal sense (each weighs about 45 lb / 20 kg). Everything inside is modular so installing a hard disk does not require a screwdriver -- just plug in and play. The fan is quiet given the beast it has to cool down. It will require an adapter for SSDs (about $10) and for newer GPUs (about $20), yet it has no problem powering up two GPUs and 2+2 HDs and SSDs.
If that is what you need to work with then work backwards from that amount. Get the most memory (ideally you will get ECC RAM which is pricier, a google search currently shows 128G DDR4 ECC (2 x 64 GB) ~$600) you think you are going to need. Follow that up by required storage.
Once those two things are decided you will need to consider a CPU/MB combo that fits within the remaining amount (AMD is currently champion of desktop CPU's) that will need to include a power supply/case/UPS (if power is unreliable).
If you don't want to deal with individual component makers on warranty issues then getting a per-configured workstation (which would be more expensive) would be the way to go.
There was just recently an update to an old thread in this forum, but the discussed options are clearly above what your PI can/wants to spend and also possibly what you need. I think moving your workloads to the cloud (e.g. with Nextflow/nf-core) might be a better alternative, in particular when you are not running heavy analyses too frequently?