Over/under kill? Local LLM coding + 1440p gaming rig design
Posted by HeavyLight03@reddit | buildapc | View on Reddit | 3 comments
Looking for a bit of advice on a rig I'm throwing together soon (Australia-based, \~$4k AUD budget). Appreciate any views and critiques!
Have mucked around on design but still not sure if this is over/under kill for what I'm looking for as prices are just no longer a good guide on what's reasonable (I last built pre-2020). I know smart thing would be to wait for prices to drop at some point, but my laptop has officially cooked my secondary drive, so no longer any time but the present.
Use case is 1440p gaming (flight + space sims, Civ 6, shooters etc.) and tooling around with setting up a local LLM for coding assistance and text analysis. Some data visualization and analysis needs for work, but no video encoding/etc. in the background on the regular.
Hoping air cooling is sufficient in this case without adding too many additional case fans. Noise not a huge concern, but also ideally not setting up a turbine.
Rough plan is:
- CPU: 9800X3D (for future AM5 upgrade path)
- RAM: Patriot Viper Venom 64GB DDR5-6400 CL32 (I've accepted I'll eventually be disappointed with 32gb, but that's still ridiculous to me)
- GPU: RX 7900 XTX 24gb XFX Speedster MERC 310 (for the LLM VRAM needs)
- CPU Cooler: Thermalright Phantom Spirit 120 SE CPU cooler
- Mobo: Gigabyte B850 AORUS Elite WIFI7 ATX
- Drive 1: Kingston NV3 500GB PCIe 4.0 X4 (OS boot drive)
- Drive 2: Crucial P510 2 TB PCIe 5.0 X4 (work/handful of games)
- Case: Fractal Design North ATX w/glass panel (because fancy)
- Case fans: 1x120 rear exhaust fan (tbd) w/stock Fractal 140 fans (I've read others swapping out the stock fans and having a full setup for a more optimized fan/noise config, but wondering if that's necessary or a no-brainer in the case given the rest of the components)
- PSU: MSI MAG A850GL 850W 80+ Gold Certified Fully Modular ATX (I'm at about 640W with this config, and don't think I need 1000W, but curious if that headroom wise to avoid issues down the line?)
Born_Bad_1294@reddit
Why 2 drives? Also why would 32GB not be enough?
https://au.pcpartpicker.com/list/GZPrDw
HeavyLight03@reddit (OP)
That's a great comparison list, thanks!
On drives: tbh not sound rationale, have just done it that way over a couple builds to keep os drive less busy, and a bit of redundancy. Just so happened that's what's kept my current laptop going (secondary Kingston m.2 drive is fully toast, but primary os micron m.2 drive with key files and apps was fine so no problems yet). Part of me thinks I don't know it but the other part thinks might as well?
On ram: my impression has been some MoE LLMs capable of larger context (e.g. gpt-oss-120B) might be viable to run on 24gb vram, but then RAM capacity becomes important. I also have a couple games where high ram usage has been a thing - but more I figure that if ram prices don't crash soon, they'll keep rising, and I'll have a harder path to upgrade later if I do want to.
Born_Bad_1294@reddit
1)If you buy cheap QLC drives, they go toast. Buy QLC drives and you should be fine.
2)As for 64GB of RAM, I don't think any single game occupies 32GB of RAM alone. And if you are planning to run LLMs on RAM then it's a great mistake, as the bandwidth is less so you can expect 1-2 tokens per minute on large LLMs which is equivalent of trash and don't look feasible.
I would recommend you get the single QLC drive and you won't be able to run GPT OSS 120B parameter model or similar high end parameter models on 20-25GB VRAM. Even if you heavily quantize it, the speed will be very slow and not viable for day to day or coding use.
I would recommend that you get the RX 9070 XT 16GB and stick with 20-40B parameter models.
I myself own a 5070Ti and like to message around with Local LLMs and stuff