By using this site, you agree to our Privacy Policy and our Terms of Use. Close
sc94597 3 hours ago
Edit Thread
JEMC said:
sc94597 said:

I got the 4080 as a Christmas gift from my partner in December 2022 when used RTX 3090's were still over-priced from the crypto mining wave ($850-$950 used, $1100 new) and the 4090 wasn't in stock regularly. My partner found a 4080 in stock at MSRP. Was upgrading from a 3080ti. So the upgrade was pretty substantial. Bought the 4090 last year when I needed more VRAM to test local LLMs for a project I was doing for work. 

Could've probably sold the 4080 and bought a 3090 (or even another 4090), but didn't feel like going through the bother.

If the GPUs were in stock, I would've probably just had got a 4090 in December 2024 and then maybe a used 3090 later on when the prices came down to $700ish. 

I see. Well, can't complain with that kind of gifts.

The 3090 could have been better with yout AI work, but once you put that 4080 through some games, I'm sure you'll appreciate the extra horse power it has.

Getting a 4090 last month would have been a waste of money more likely than not. With Nvidia stopping production back in September or Octobert to get rid of all the 4000 cards in retail (they sure knew the new ones wouldn't be well received and make sure to get rid of other options), the price for them skyrocketed to the point where a 5090 would be a much better investment.

Good Lord, we're seeing lots of delistings in the last months. I don't recall having so many in prior years.

Anyway, thanks for sharing.

Oops, meant to say December 2022, not December 2024. I wanted either a 4080 or 4090 around then, but the 4090 was out of stock (or only the overpriced SKU's were in stock) and I was fine with the 4080's performance for the cheaper price ($1300 after tax.) 

Back then, I didn't really care as much about VRAM since the models I was training (not inferencing) were compute bottlenecked and not VRAM bottlenecked. The 4080's gaming performance was sufficient for me (4k 60fps-90fps native or DLSS Quality in most games) so I was fine with just getting that. Local LLMs started to get much more popular in Spring 2023 after GPT 4 released, and that's when VRAM became a bigger consideration for me as my job wanted us to implement an in-house API for local inference, and I thought I'd experiment on my personal PC first. Eventually this escalated to many personal and work projects, and I can't get enough VRAM, the more the better. 

I've used my 4080 and 4090 for gaming, but in most gaming use-cases my display is the limiter and the GPU's go under-utilized. If I didn't use them for ML/DL/RL and CUDA-accelerated scientific computing I would probably have just got an RTX 4070. Before becoming a Data Scientist/MLE I always would just get mid-ranged GPU's and use them for 3-4 years. For example, between 2017-2021 I had a GTX 1060 and before that I had an r9 280 from 2014.