Check my specs for High-budget ML/CS PC

[Pascalos]

I am planning on getting my own PC built, after realising that computer parts are currently so expensive that it's actually cheaper to get it built for you.
This is my first time configuring a PC, so I would like some advice. I have been researching what parts to pick for almost 2 years now... and I feel like I have a decent understanding of what everything does and what I need for my build. Despite that, I am still hesitant, as I feel I may be over-spending a bit and am unsure on whether to go custom loop or AIO.

Total budget: €3000-4000 ex. VAT (for JUST the PC, excluding monitor, speakers, etc .)
monitor: 1440p 75Hz (Acer V277U bmiipx as example, this is a 10-bit RGB monitor, but I think the NVidia cards do support that)
Use-case:
I am planning on using this PC as my main machine, with the intend of hooking it up to my router and having remote access with a light-weight laptop. I do not need high-latency for this, as I don't intend to use it for gaming in this configuration.
I want to be able to train state-of-the-art Machine Learning algorithms, Genetic Algorithms (require lots of RAM), run physics simulations and do so while utilising high-performance multi-threading for long durations (think DAYS). I occasionally would like to use this system for 1440p gaming at 60+ fps as well, with games like "No Man's Sky", "Cities Skylines", "Civilisation VI" and heavily modded minecraft at high resolutions. But this is not the main focus.
As I like experimenting a lot with Genetic Algorithms, simulations and other advanced algorithms, I like to have a lot of memory and thread-count headroom. I was thinking about going for 128GB of RAM, but since I can't select RAM at that size for 3200MHz, I did not put that in here.
I am planning on expanding the SSD storage when I need more, but for now, 2TB seems plenty for me.

Main question:
Am I going over-kill with cooling?
I originally wanted to go for a H115i AIO cooler for the 5950X, but figured that the long-heavy-loads I expect from the PC might be better optimized using a custom loop and overclocked components.

This is the build I came up with:
Case
THERMALTAKE CORE X71 TEMPERED GLASS EDITION GAMING CASE (LS)
Custom Liquid Cooling Kit
Liquid Series RGB High Kit - Corsair Hydro X
Tubing
Clear Hardline Acrylic Tubing (Metallic Fittings)
Graphics Card Cooling
GPU Water Block - For One Graphics Card!
Coolant Colour
Mayhems X1 UV Clear Blue
LED Lighting
50cm UV LED Strip
Overclocked CPU
Overclocked AMD Ryzen 9 5950X 16 Core (3.4GHz @ up to 4.55GHz)
Motherboard
ASUS® TUF X570-PLUS GAMING WIFI (USB 3.2 Gen 2, PCIe 4.0, CrossFireX) - RGB Ready!
Memory (RAM)
64GB Corsair VENGEANCE DDR4 3200MHz (4 x 16GB)
Graphics Card
12GB NVIDIA GEFORCE RTX 3080 Ti - HDMI, DP
1st M.2 SSD Drive
2TB CORSAIR MP400 NVMe PCIe M.2 SSD (up to 3480 MB/R, 3000 MB/W)
Power Supply
CORSAIR 850W RMx SERIES™ MODULAR 80 PLUS® GOLD, ULTRA QUIET
Power Cable
1 x 1 Metre European Power Cable (Kettle Lead)
Braided Power Supply Cables
CORSAIR Premium Individually Sleeved PSU Cable Kit Pro - Black
Thermal Paste
ARCTIC MX-4 EXTREME THERMAL CONDUCTIVITY COMPOUND
Sound Card
ONBOARD 6 CHANNEL (5.1) HIGH DEF AUDIO (AS STANDARD)
Network Card
10/100/1000 GIGABIT LAN PORT
Wireless Network Card
NOT REQUIRED
USB/Thunderbolt Options
2 PORT (1 x TYPE A, 1 x TYPE C) USB 3.1 PCI-E CARD + STANDARD USB PORTS
Operating System
Windows 10 Professional 64 Bit - inc. Single Licence [MUP-00003]
Operating System Language
United Kingdom - English Language
Windows Recovery Media
Windows 10 Multi-Language Recovery Image - Unlimited Downloads from Online Account
Office Software
FREE 30 Day Trial of Microsoft 365® (Operating System Required)
Anti-Virus
NO ANTI-VIRUS SOFTWARE
Browser
Microsoft® Edge (Windows 10 Only)
Cable Management
3 x PCS 1.5M Zip Cable Tidy - Professional Cable Management
Warranty
3 Year Standard Warranty (1 Month Collect & Return, 1 Year Parts, 3 Year Labour)
Delivery
2 DAY DELIVERY TO NETHERLANDS
Build Time
Standard Build - Subject to stock availability on pre-order products
Price: €3.970,25 including VAT and Delivery

PCSPECIALIST - Top Spec Custom PCs & Laptops Built to Order

Manufacturer of performance custom computers and laptops. Customise your ideal gaming PC or laptop using our advanced configurator.

www.pcspecialist.co.uk

 

[DaelpixPhotos]

Be sure to add the reconfiguration link, then we can easily reconfig if neeeded

How to Post Your Specification to the Forum - Check this Spec

Just a first note, if you've been provided this link, don't amend your first post as it won't update the queue, please add an additional comment to the thread, that way it will go back to the front of the queue. Otherwise we don't know it's been updated. When creating a forum post looking to...

www.pcspecialist.co.uk

 

[Scott]

If you don't mind the budget difference on the cooling I think it's a fair shout for your uses. I'm not convinced on your GPU selection though. That's wayyyyyy past what you need to game at lower end 1440p on.

Can your software make good use of the CUDA cores from the GPU or something else? If the GPU is purely for gaming, save a ton and opt for the 3060ti. That would more than have you covered.

 

[Bigfoot]

I am not sure the liquid series is needed here. You will get as good cooling from an AIO. If you need a lot of cores and much RAM, would a Threadripper build suit you better? I don't have any expertise on those myself.

 

[Pascalos]

I'm not very sure about the GPU to be honest.. As I know you need a large amount of VRAM (at least +8GB, but 8GB would leave little expansion room) in order to load the state-of-the-art networks and datasets for training. But I don't really need the raw power of the 3090 that much.

And regarding threadripper, I'm not sure I'll be able to fully utilize the 32+ cores of such a system, and know that some algorithms are harder to parallelise, so the balance of high core count and excellent single core performance of the 5950X seemed best for me.

What GPU could I use instead? Would that require me to go for a threadripper due to the available options from PCS?

 

[Bigfoot]

It sounds like the Threadripper won’t suit your applications. If VRAM is the most important aspect of the GPU, the 3060 might be suitable, as it has 12GB. It will be much cheaper than the 3090. The 3060 is really a high end 1090p card rather than a 1440p one.

 

[Scott]

Just to make sure.... you're not confusing VRAM with System RAM? You're loading the datasets into the GPU? Is the GPU doing the processing then?

 

[Pascalos]

Just to make sure.... you're not confusing VRAM with System RAM? You're loading the datasets into the GPU? Is the GPU doing the processing then?

They are loaded into RAM at first, along with part of the data-set. But implementations of particularly deep neural networks are most efficiently ran on a GPU, as the operations needed to train it are all simple instructions executable by CUDA cores. Due to software limitations, this does mean you are gonna have a hard time doing this with an AMD GPU. And because state-of-the-art networks can be quite large, you need a high amount of VRAM in order to not let any of the training spill into main (RAM) memory.

 

[Scott]

Indeed, this is why I asked if your software used the GPU...... you said you weren't sure so I was trying to convey the importance of knowing the difference and the utilisation.

If you're using the GPU and the CUDA cores to do the calculations then you want a high end GPU, if you're using the RAM and the CPU... you want a high end CPU. If you're extensively using both (Quite unusual but not unheard of) you want both ends propped up.

 

[DarTon]

I think your choice at that budget level is between a 5900X and 3090 and a 5950X and 3080 Ti. Personally, I'd go for the former to get the additional VRAM. If you still use the CPU (and the algorithms can actually use 16 cores really efficiently without precedence issues) and the data sets are smaller then the 3080 Ti is ok. It's just that these days, every $ spent on the CPU is a $ that could have been spent on the GPU.

I'd also consider changing your storage to PCIe4 drives. We use GPUDirectStorage now, which cuts out the CPU and system memory, so we put all the budget into the GPUs (multiple A6000, 3090, sometimes the odd single 3080 Ti depending on specifics) and fast storage and cut costs on the system memory and CPU. It reduces the time needed to train the models. Again depends on the specifics/size of your datasets.

 

[Pascalos]

What are you suggesting regarding storage? Replace M.2 with something else?
Okay I understand now, do you know which of the PCS supported M.2 drives are PCIe Gen4? It is hard to find tbh

In which cases would storage be the main bottleneck? As I see now going Gen4 is about €100 extra while doubling data transfer speeds..

 

[Pascalos]

Indeed, this is why I asked if your software used the GPU...... you said you weren't sure so I was trying to convey the importance of knowing the difference and the utilisation.

If you're using the GPU and the CUDA cores to do the calculations then you want a high end GPU, if you're using the RAM and the CPU... you want a high end CPU. If you're extensively using both (Quite unusual but not unheard of) you want both ends propped up.

Hmmm, thank you
I am enthusiast (and a student) in both Computer Science and Data Science, so I do think I can benefit from both a hefty CPU and GPU, regardless of the games I play.

 

[DarTon]

I don't know what software you are using so it's impossible for me to ascertain what is optimal. Storage is easy to add more later. Don't overbuy on it until you've experimented with what is best for you.

I would say though that the having two drives rather than one is probably sensible from a risk management perspective. SSDs fail very rarely but if they do it can be 100%. So having a 500Gb drive (say a M2 PCIe4 one like a Samsung 980 Pro, Firecuda 520/530) for OS, programs etc is a good idea. If you want to then save some money get a cheaper M2 PCIe3 SSD for your data. Division of risk etc.

If you're a student, I'd agree to just buy a balanced system since you don't actually know what you're going to need. In terms of deep learning though something like DGX-2 is a good example of a commercial system. It starts for about $100k and has 2 x Xeon 24 core processors. They have each about the same CPU processing capability as a 5900X (probably less actually). In contrast it does have 16 Tesla GPUs with over 80,000 CUDA total cores. So the direction of travel in terms of CPU vs. GPU balance is pretty clear.

 

[sck451]

What are you suggesting regarding storage? Replace M.2 with something else?
Okay I understand now, do you know which of the PCS supported M.2 drives are PCIe Gen4? It is hard to find tbh

In which cases would storage be the main bottleneck? As I see now going Gen4 is about €100 extra while doubling data transfer speeds..

If either read or write speeds are over 3500MB/s, it's gen 4. The Samsung 980 and the Seagate Firecuda 520 and 530 are the options available: the Seagate 530 has particularly awesome write endurance if you're going to be doing a lot of writes in your work.

 

[Pascalos]

(regarding software, I am not using anything specific; I guess you could mention openCV for Java and Python, as well as many different other Java and Python packages. I also intend on going more into c++ and will occasionally use OR tools like Gurobi. So far I only used my own Java code for simulations, so I don't know what software I'll use there, same counts for GA's)

Read endurance is probably more important for me, so would it then matter what drives I take?
And then I suppose I would like 500GB of gen4 storage and 1-2 TB of gen3 storage at first, as when the dataset is too large for RAM, I can still read from it at 7GB/s

I think I would lay a bit on the CPU side of things for the CPU vs GPU budget.. so 5950X and 3080Ti would be good for me (listening to all your feedback), as a threadripper is too much for me and I like the DL capabilities of the 3080Ti, but don't need more than 12GB VRAM.

I can always specialise my system later if I know more what direction of the field I want to go in
Thank ya'll for the great feedback

P.S.: why are the Braided Power Supply Cables so expensive? do they make a big difference in cable management?