AMD or Nvidia for my new GPU
I would like some help reviewing my decision process to see if I’m failing to take something into account.
I won a raffle for a 4070 ti/super or 4080 graphics card (I don’t know which quite yet). All the cards in the raffle will fit my case, but if I get a blower the thermals will be bad (vertical case). I currently have a 6700xt, am on linux. Case is NZXT H1 V2.
Use cases: gaming at 1440p, and potentially machine learning. If I stay AMD I’m not against buying GPU cluster time to run anything I need to on that.
Decision factors:
- I would like to do machine learning work, as I’m going into data science as a career. GPU cluster time is < $1 per card per hour, and it’s not like I’ll be doing anything big on my personal machine, just portfolio work. If I have an Nvidia GPU I can at least check for errors before i start trying run things remotely, or run them locally if I have GPU time on my end.
- Big point for Green.
- I’m on linux-wayland. Nvidia drivers are notoriously problematic here. The newest drivers should fix a lot of the graphical Nvidia-wayland issues but I’m skeptical since they dropped like, this last week or something. I haven’t seen a ton of complaining so that bodes well. It looks like they work well for AI. So AMD will be easier and less problematic, but Nvidia isn’t out of the picture.
- Point for Red
- I game a lot, 1440p. I will spend more time gaming than running machine learning algorithms, especially since my PC is in my room and I don’t like having it run while I am asleep. I don’t care for RTX. This tends me towards AMD + renting GPU cluster time.
- Point for Red
- I’ve got a 750W PSU and pcpartpicker specs my current system as 316W without the GPU. But even the 6700xt heats up my room a lot though. iirc it’s ideal to have ~20% headroom from system consumption to PSU max for transitory spikes and PSU max efficency? Higher end and newer GPUs are more efficient for a given performance, producing less heat in equivalent light workloads and only produce more heat under the higher/heavier workloads. Nvidia is better in this regard all around it looks like.
- Point for Green
- vram… I still believe that vram has the largest impact on GPU longevity since I like to play modded games, and game devs at large seem to assume everyone has infinite vram now, and don’t optimize for it. I don’t know how true this is anymore, but I bought a 4GB GTX 980 (my prev card) specifically bc it had 4GB ram over 2GB and that was a very good decision. This points to AMD, which would also be a long term bet on AMD building CUDA equivalent libraries and translation layers, which imo is a safe bet rn. But I also could easily be wrong about my vram bias, and the bet is long term, without an actual upside today.
- Small point for Red.
| GPU | vram | pcpp est. Watts | PSU headroom |
|---|---|---|---|
| 6700 XT | 12 GB | 546 | 27% |
| 7900 GRE | 16 GB | 576 | 25% |
| 4070 Super | 12 GB | 536 | 30% |
| 7900 XT | 20 GB | 647 | 15% |
| 4070 Ti | 16 GB | 601 | 20% |
Closing thoughts:
If I sell both I’ll probably be buying a 7900 GRE, because, for AMD’s offerings, the price to performance and thermals to performance are the best.
Nvidia has better performance to thermals overall, and the option to run local AI algs for testing, but potentially more problems getting to play nice with my system and the vram is lower.
If the card I get is too power hungry or a blower I’ll sell for sure (2/6 chance). But if I get a 4070 Super or Ti, I can not decide if I sell and buy AMD or if I keep it and learn to troubleshoot Nividia on wayland (I can also always hop to x11 if I need to)
… I can not decide. Sorry for the long post. Thanks for any thoughts. Esp about running Nvidia/AI work on KDE plasma + wayland. What would you do?
Resources
Tom’s Hardware GPU hierarchy, Tom’s Hardware Best GPUs of 2024