PC Building

• PC Building in late 2025
  • The non-negotiables
  • Trade-offs
    • Why portability?
    • Non-negtiables
  • What’s available in the market?
    • Components
  • A bit more of my thought process -
• Ideas for the older laptop
• Readings

PC Building in late 2025

Historically, this is the time when RAM prices were out of the roof. A 32 GB (2 x 16 GB) kit of DDR5 6000MHz can cost anywhere from 300 to 400 euros. And I have decided to build my first custom PC at this time, what a great start!

So let’s set things up. I own an old gaming laptop, specifically a Dell G3, i5 8th gen 8300H, 1050Ti 4 GB VRAM, and upgraded 32 GB DDR4 2666 MHz (recently from 16 GB from 8 GB when bought in 2018), and running Linux Mint. It’s great for everyday use, but the processor is showing its age, and the thermals are not good, even after changing its fans three times and repasting both the CPU and GPU. I bought this one back in 2018 for my college, and it was really good and quite an expensive purchase back then, but it sure lasted a while.

I feel now is the time to upgrade and get a better machine for work and play. My idea with this is to have hardware and OS flexibility. This automatically rules out Apple with their MacBooks and Mini desktop computers. There’s also another reason to rule them out, and that comes from my work reason. So this narrows down to either an Intel or AMD laptop or desktop, or the newer mini-PCs.

The non-negotiables

1. OS Flexibility -> should run Linux.
2. Hardware flexibility -> should allow upgradability, RAM, CPU, storage, GPU
3. GPU computing - Support FP64 computations at the hardware level

Trade-offs

1. Portability - this is counterintuitive
2. Cost - RAM price uncertainty

Why portability?

• I initially thought of getting another gaming laptop. A similar price range as my old one, maybe a little upgrade, maybe a better screen, or softer fan, or less weight, or more professional aesthetic. Not a huge fan of gamer aesthetic either. But these were way out of my budget, all above 2k euros. And the ones in my budget - Legion, LOQs, Victus, etc.

• Second thing with portability and gaming laptops is that it’s kind contradiction. They always need a power connection to work beyond an hour or two, forget doing anything productive without draning battery (atleast in my experience).

• My use case does not want me to travel a lot and carry computing on the go. I will have a pretty stable living for at least the next year and hopefully for the next 2-3 years.

• For my portability needs, I’m well equipped with a work laptop (HP ProBook 445 14 inch G11, rocking AMD Ryzen 5 7535U with Radeon Graphics), and it is more than enough for writing code, reading papers, and using PowerPoint. Drawback is that it had HP and Microsoft bloatware.

• RAM price uncertainty and increase -> This is part of life, the last couple of years it was GPU shortage, now it’s RAM shortage. Can’t do anything about it, and I feel it’s better to get it now, already saved up money (read of satisfying the itch of building a PC for the first time, which may or may not be a bad financial decision, only time will tell!)

Non-negtiables

• I have used Linux (Linux Mint) for the last 2 years as my daily driver, initially dual booted with Windows on my Dell G3 and later on as a standalone. With Linux gamming becoming ever more accessible and user friendly, I see no reason to stick with Microsoft other than the limitation posed by certain software tools like AutoCAD, SolidWorks or ArcGIS. I do not do much CAD, but QGIS is a better alternative and works great on Linux (from experience). Linux saved me from upgrading sooner. Back in 2022, my old laptop started showing signs of slowing down to the point of being unusable, especially when running MATLAB. I then decided to upgrade my ram from measly 8 GB to plentiful 16 GB. It again slowed down after some windows update, and I have had enough. I decided to dual-boot my computer with Linux, and it was magic, my laptop felt like new and Linux Mint was as noob friendly as Windows. And never looked back since, had some issues with bluetooth and Discord still sucks on Linux. But it’s pretty great all around.

• I guess this leads me to the second part of hardware upgradability. When I bought the old laptop, I didn’t knew anything about my requirements, I just knew a couple of softwares - SolidWorks, AutoCAD, and PhotoShop and they should be supported, that’s it. Later on went with the recommendation of a family friend. Now I have a better understanding of PC hardware, and its limits (that’s more important). I feel a lot of advice online is geared towards max. specs, price-to-performance, and raw metrics, the point of hardware limitations is somewhere lost. I feel I have reached the hardware limits on my old laptop with my current work. It’s still great for a lot of other things, but not my work, and I have other ideas in mind to test with it. So, I want the flexibility for my PC to grow with me. After all, NASA sent bunch of humans to Moon with just 4 KB of RAM and 72 KB ROM (I just have a skill issue - need to learn Assembly!).

• GPU compute - So this is a good-to-have feature as it’s not a dedicated work machine, but my tool. Most of my research work is around computation, and I can get dedicated hardware if I ask for it, which is way more advanced than any personal computer. So a lot of scientific computing involves something called double precision computations (FP64 or float64). Any modern CPU will support FP64 operations. This cannot be said about a lot of consumer GPUs, though - I did not know about this fact until recently. A lot of Nvidia consumer-grade GPUs are great at FP32 (single precision) computing, as their main focus is gaming. Whereas the theoretical FP64 compute is a fraction of FP32 support (either 32nd or 64th or something). Some say this is intentional to sell their professional series of GPUs, which are way too expensive. There are also TPUs (tensor processing units) that have hardware optimized for matrix multiplication. Also, there are datacenter-grade GPUs. But anything other than consumer-grade is too expensive and way too overkill for my needs (for nuances and details check references below).

• So I thought, Apple must be the best as it’s super efficient with their integrated memory, but the shocker is Apple GPUs do not support FP64 (nuances see [4]), and support for Apple’s Metal API is limited. AMD GPUs are great for gaming and similar to Nvidia’s offering.

What’s available in the market?

I have been searching for over 2 weeks now. The possible options

1. Custom Linux computers - Framework, Tuxedo computers, Slimbook, Nova Customs -> way too expensive for raw performance, but support for good companies (not at this stage financially)

2. Big brand consumer offering - LOQ, Legion, Victus, Gigabyte Aero, MSI -> It’s good but hard to find something I like -> either too expensive and I lose upgradability or affordable but under-powered for my requirements.

3. Used market -> Not enough experience and fear of scams, possibility to make a Frankenstein with used enterprise-grade hardware, but again, skill issue.

4. Custom PC built -> Best alternative and itch to Lego assemble parts together.

Components

I hope this feels justified, when I inevitably regret spending 1300+ euros on this PC.

1. Intel Core Ultra 7 265K with 20 cores (8P + 12E cores) - massive upgrade from previous 4-core 8th gen and future support

2. Intel Arc B580 12 GB VRAM - best new budget GPU with good raw FP64 compute compared to other new alternatives from Nvidia or AMD(check GPU specs)

3. 32 GB (2 x 16 GB) DDR5 6000 MHz (super overpriced but no choice TT)

4. NVMe M.2 SSD, 2 TB from Samsung (overpriced)

• combined with a compatible wifi motherboard, ATX case, air cooler, and 850 W PSU.

A bit more of my thought process -

• With CPU selection, I was leaning towards i7-14700k as it’s the latest 14th gen and is the one in my work desktop (i7-14700 without GPU), and it works great with Windows and Linux, and handles my workload fine till now. But then there is the old issue with 13th and 14th gen CPUs, and also discontinuation of the socket, but it also supports DDR4. It was a decision that just felt right to get the latest Core Ultra series instead, even though the motherboard is a bit expensive. I guess it’s “future proof”?

• I went with a motherboard with WiFi, it’s just better - I have used a computer with a mobile hotspot a lot, and it’s really flexible.

• An air cooler is just better, fewer moving parts and no headache of a liquid cooler, considering a lower-end AIO vs a high-end vapour chamber-based air cooler tower. Performance is similar to what I have seen.

• ATX case - I debated with going for mATX form factor, but I feared thermal issues with my reservations with AIO. I saw some builds with mATX using more power hungry components, but I thought it will limit upgradability options in the future. Also mATX case that I liked (fractal silent) was super expensive. There’s something called “mATX Tax” in the PC building industry. Also, they require a special SFX power supply.

Ideas for the older laptop

1. Linux experimenting machine -> experiment with Arch Linux without hampering my work

2. using the unused storage on it as a backup drive, or using it as a NAS?

Readings

1. chapter 2, overview of GPU Architecture though for older gen Nvidia and AMD architecture FP32:FP64 (SP:DP) ratio table comparision of consumer grade (GTX) vs Pro grade cards. There are more nuances which I do not understand. Scalable scientific computing applications for GPU-accelerated heterogeneous systems, PhD thesis by Christoph Karl Riesinger, from TUM https://mediatum.ub.tum.de/doc/1360268/736774.pdf (as of 18.12.2025)

2. Taste of Actual Workstation/Scientific Computing recommendations that experts (Puget System) suggests. These are for the real pros with tens of thousands of dollars and need to perform complex CFD simulations, train or run AI/LLMs and other advanced stuff - https://www.pugetsystems.com/solutions/high-performance-computing/scientific-computing/hardware-recommendations/ (as of 18.12.2025)

3. Puget systems has interesting blog which I found later and echos similar stuff that I found elsewhere. This article also lays it flat the compute of consumer grade GPUs such as RTX 3080Ti for AI/ML or HPC tasks. https://www.pugetsystems.com/labs/hpc/nvidia-3080ti-compute-performance-ml-ai-hpc-2170/ (as of 18.12.2025)

4. Closer look at Apple Sillicon GPUs in scientific computing though only covers M1 series - Apple Silicon Performance in Scientific Computing by Connor Kenyon and Collin Capano https://pure.mpg.de/rest/items/item_3485053_3/component/file_3485054/content (as of 18.12.2025)