Workstation Build for Bioinformatical Analysis [High Threads and High Ram]

[r3dc3t]

Budget (including currency): 10K-15K Euros

Country: EU

Games, programs or workloads that it will be used for: Ubuntu, R, Python, Software requires high-end GPU for calculations

Other details (existing parts lists, whether any peripherals are needed, what you're upgrading from, when you're going to buy, what resolution and refresh rate you want to play at, etc):  Fresh build.

 

Hi,

 

I am a Ph.D. student in live sciences, and I need to analyze large datasets produced by NGS (next-generation sequencing). 

I use Ubuntu for convenience and community support. R and Python are the main languages that I use, and objects that we are working on are reaching 70-100 GB (in the future, I probably need more). 

And some packages require raw data to work with, in these cases it requires to use 32-64 threads and >256GB ram.

And, some other packages require high-end GPU (with ram and compute power).

 

Since lots of packages and tools are in constant development, there is a constant flow of errors, and some of them require administrator privileges to fix them.  For this reason, I cannot use shared server space (most of the time, response time and solutions are too slow for this kind of work). 

Some of my colleagues solved their problem by getting a high-end workstation. Our IT department requested for me to come up with a list. Since my tech-savvy is limited to upgrading the SSD on my laptop, I want to ask for your recommendations and opinions.

 

I think these are what I need;

1. high thread CPU (>=64)
2. high/reasonable fast RAM (ideally 512 gb)
3. High-capacity storage for passive files (HDD min 20 TB)
4. Fast storage for active projects (nvme SSD type min 4 TB)
5. support for multiple Nvidia GPU
6. support for 10 gigabits ethernet speed

 

Thanks in advance for your time, input, and opinions.

 

 

 

[venomtail]

AMD just released their new Threadripper line. 7970X and 7975WX could be good fits for you. Rest of parts should just fall into place, just do you know if the GPU you need is an NVIDIA one that uses CUDA or is AMD still a valid option, allowing you to save money. Do you even need a workstation GPU or will just a 4090 or 7900XTX be just enough?

[r3dc3t]

17 minutes ago, venomtail said:

AMD just released their new Threadripper line. 7970X and 7975WX could be good fits for you. Rest of parts should just fall into place, just do you know if the GPU you need is an NVIDIA one that uses CUDA or is AMD still a valid option, allowing you to save money. Do you even need a workstation GPU or will just a 4090 or 7900XTX be just enough?

I will need CUDA. (exp, https://github.com/pinellolab/dictys?tab=readme-ov-file FAQ)

[brob]

1 hour ago, r3dc3t said:

 

I think these are what I need;

1. high thread CPU (>=64)
2. high/reasonable fast RAM (ideally 512 gb)
3. High-capacity storage for passive files (HDD min 20 TB)
4. Fast storage for active projects (nvme SSD type min 4 TB)
5. GPU
6. support for 10 gigabits ethernet speed

 

You may want to adjust 5 to "support for multiple Nvidia GPU".

 

In 3, is HDD storage in the system really necessary or can access to network storage be used?

 

 

[r3dc3t]

11 minutes ago, brob said:

 

You may want to adjust 5 to "support for multiple Nvidia GPU".

 

In 3, is HDD storage in the system really necessary or can access to network storage be used?

 

 

Thanks for the recommendation (5).

Network storage and options would be slower. Our institutional network storage works well if you want to store.  Anything other than that causes problems.

My plan was, to copy the data from long-term network storage to a build-computer, then perform a md5sum check.  Use it for the analysis, after that remove from the HDD.

[Somerandomtechyboi]

Epyc 7713 64c 128t zen3

 

Gigabyte mz32-ar0 lots of pcie x16 slots for multi gpu and also 16 ddr4 slots

 

Theres 2 choices for ram

M386a4g40dm0 32gb 4drx4 2133 lrdimm which are 27.34€

Hma84gr7cjr4n 32gb 2rx4 3200 rdimm for 42.74€

 

X16 because 16 ram sticks to fill all the slots for 512gb, so 437.44€ for the 2133 lrdimm and 683.84€ for the 3200 rdimm

 

I suspect you can run the lrdimms at 3200c22 jedec since its samsung 8gbit d die based which is known to clock 4400+, but this is uncharted territory due to being lrdimms and nobody toying around with ecc overclocking so dont go buying those samsungs in the hopes of 3200+, doesnt hurt to try if you already intend on buying the samsungs anyway

 

 

Around 2200€/2000€ total depending on if you get the more expensive hynix 3200 rdimms or cheaper samsung 2133 lrdimms 

 

That leaves 8-13k for the rest of the system, so i guess you can go for a triple/quadruple 4090 setup depending on how much gpu compute you need, or you can opt for a dual socket board instead and add another epyc 7713 for 128c 256t, though youll need to run dual psus if you wanna run more than 2 4090s

 

 

But since your workload uses alot of ram im curious if it may benifit from extra cpu cache since the epyc 7773x exists, its about 2x the price of the 7713 with cheaper ones around 2200€ but comes with 3x the cache (256mb on the 7713 and 768mb on the 7773x) so that may be worth looking into if your workload benifits from extra cpu cache which i suspect it does due to the massive ram consumption

[Somerandomtechyboi]

Mz32 ar0 overview from the gigabyte website, interfaces to take note of are slimline and ocp mezzanine

 

Slimline is used for nvme with these adapters so you can get 6 nvme drives if you want, it is also used for sata with this adapter and is the only way to get a sata hdd on the board, you can also get 8 sata drives if you want, it can also be used for u.2 ssds but thats irrelevant due to their pricing

 

Ocp mezzanine is used for ethernet cards, Heres a 25gbit card for 300€ though you may find cheaper cards (atleast cheaper shipping)

 

 

https://de.pcpartpicker.com/list/Kzhdvj

Heres what i came up with for the rest of the build

 

Dual 4090 as the base, add more if needed, though keep in mind that youll need to run dual psu if you want more than 2 4090s unless you powerlimit and undervolt the 4090s (saves powerbills too)

 

Two 20tb 7200rpm hdds and two high speed 4tb gen4s also as a base so you can run raid 1 to prevent any data loss from a sudden drive failiure, you can add more drives upto 6 nvme m.2 and 8 sata with the slimline adapters alone

 

LS720 is one of the top aios atm next to the liquid freezer ii and ek aios, dont worry about the rgb as it shouldnt light up if it isnt connected to an argb header which the board obviously doesnt have, though if you do for some reason want rgb there are usb to rgb adapters

 

 

This is about as far as i can go as i dont specialize in server hardware, i highly reccomend checking out servethehome forums as they are very server/workstation oriented so im pretty sure theyll be of some help