Chazore said:
Cyran said:
I was lucky and was able to get a 3090 basically at launch for $1800 so been running it for a bit over 4 years now. I leaning toward buying the 5090 assuming it not more then $2500 which with nvidia these days who knows. I am guessing the 5090 after launch only going to get more expensive not cheaper especially with the possibility of tariffs and I not sure I willing to wait till 2027 for a 6090. 6 years would be longer than I ever gone without upgrading. 4 years is already getting close to my record without a upgrade.
It don't seem worth it going to a 5080 as I been playing with running some LLMs locally lately and the 24GB in the 3090 get eaten up very quickly so don't really want to downgrade ram. Which leaves 5090 only viable option for a upgrade in 2025. Ideally would like something even more powerful for LLMs but those cards are way out of my budget and I would still need a gaming card since those would not be good for that so 5090 my only option for a upgrade for that also.
This going to be expensive year I think for me because I still prefer 27 inch monitors and not upgraded in a long time because been waiting for something like this that just been announce no price yet https://rog.asus.com/us/monitors/27-to-31-5-inches/rog-swift-oled-pg27ucdm/. Hopefully 5090 will have a DP 2.1 UHBR20 slot also I can't imagine Nvidia skipping it for another generation but who knows with nvidia. |
I know it may be rude to ask, but what type of LLM's have you been testing with recently?. Are they ones designed for visual generation/learning, or code learning ones?. |
Right now I been playing with text base models so similar to ChatGPT, OpenAI etc but running local (and a much smaller model). I find 33B MemLong model run ok on my 3090 but anything bigger get a bit slow. 70B takes over a min to respond so basically useless at that point. Right now mainly playing with it but eventually might try building code that can ask questions and take responses to automate stuff like a custom news feed. Obvious there easier ways to do that then with a AI but more for educational purposes for now. Coding if you a beginner it fine but generally speaking any journey level programer should be able to produce more efficient code then chatGPT etc can produce at moment. Now 5 years from now that could change but there would need to be a pretty big jump for that to happen. Problem with AI currently is it just wrong a lot and it present that wrong information as correct.