02/03/2026 3:25 pm
Topic starter
Notifications
Clear all
GPU Forum
5
Posts
6
Users
0
Reactions
56
Views
0
I am so sick and tired of these Out of Memory errors on my old rig. Honestly fed up with my 3060, it keeps crashing every time I try to run anything bigger than a tiny 7B model and im ready to switch to something else entirely. Ive been grinding and finally saved up about 5500 bucks to build a dedicated workstation for my startup here in Berlin and i need to get it done by mid-October.
So i was thinking... my logic was that I could just buy two 4090s and call it a day because they are so fast for the price. But then I started reading about the memory bandwidth and the vram limitations and now Im doubting everything. An A100 (even a used 40GB one) would be super stable and the drivers are better for linux, right? But the raw speed on the 4090 for training seems better? Or am I totally wrong?
I'm stuck between:
- going the consumer route with dual 4090s
- finding a refurbished A100 or maybe even an A6000
The A100 feels like the pro choice but the price tag for a new one is insane. If I go 4090 am I just gonna run into the same vram wall later? I just want to stop seeing that CUDA out of memory message it literally haunts my dreams at this point...
5 Answers
11
02/03/2026 3:26 pm
tbh if you want to end those OOM issues, dual NVIDIA GeForce RTX 4090 24GB cards are fast but a used NVIDIA RTX A6000 48GB is way more reliable. Itll fit your 5500 budget. Less headache.
- 48GB vram handles larger models easily
- Single card avoids pcie lane bottlenecking
- Better thermals for your workstation build
5
02/03/2026 4:20 pm
Omg me too! Those OOM messages are literal nightmares. Im diy-ing a budget setup to save cash:
- hunting used NVIDIA GeForce RTX 3090 24GB cards
- sniping local deals Saving money feels amazing!
3
02/03/2026 4:00 pm
Re: "tbh if you want to end those OOM..." - honestly that A6000 suggestion is the most logical route for your 5500 budget. People forget dual NVIDIA GeForce RTX 4090 24GB setups require expensive blower models or massive cases for thermals. Going with a single NVIDIA RTX A6000 48GB keeps power draw under 300W and works on a standard PCIe 4.0 slot. Its a way more cost-effective method to get 48GB addressable memory without the p2p headaches.
3
02/03/2026 4:00 pm
Noted!
3
05/03/2026 3:49 am
Works great for me
