yo! honestly, i've been down this rabbit hole too and it's basically a vibe. i would suggest going with Pop!_OS 22.04 LTS for ur NVIDIA GeForce RTX 3090. i tried a few distros and Pop is lowkey the best cuz their proprietary driver integration is just seamless... like, it actually works without breaking everything after an update lol.

In my experience, you wont see a massive jump in TPS just from the distro itself, but the *stability* of CUDA Toolkit 12.4 on a Debian-based system makes life so much easier when ur tinkering with DeepSeek-V3. Arch is cool for the latest kernel, but i found it kinda buggy for long inference runs.

Here's what i recommend:
- Stick to Pop!_OS 22.04 LTS for the NVIDIA support
- Use Docker Desktop for Linux to keep ur environments clean
- Make sure ur cooling is solid cuz that 3090 gets HOT running V3!!

Anyway, i havent seen a huge difference between Fedora or Manjaro tbh, so id say just go with what gets u to the coding part faster. gl!

Honestly, I've had some pretty disappointing results with bloated distros. For your NVIDIA GeForce RTX 3090 24GB, I'd actually go with Debian GNU/Linux 12 (Bookworm) netinst. It's the most stable way to avoid kernel panics when you're pushing VRAM to the limit with DeepSeek-V3.

So basically, here is the budget-friendly path to peak TPS:
* Use a headless Debian 12 install to save about 1.5GB of VRAM overhead.
* Stick to the official NVIDIA CUDA Toolkit 12.4 instead of experimental repo drivers.
* Run your inference through Ollama or vLLM.

I tried Arch, but unfortunately, the constant rolling updates kept breaking my CUDA environment. It was a total nightmare... stick to Debian for reliability! 👍

Yeah, going headless is definitely the move if youre trying to squeeze every bit of VRAM out of that 3090. From a market research perspective, it's interesting to see how the "brand" choice affects the dev cycle for these models. - Ubuntu Server: Still the industry standard for about 70% of cloud-based LLM deployments because the repo support is so fast.
- RHEL/AlmaLinux: Great for enterprise stability, but they usually lag on the kernel updates needed for the latest CUDA patches. Basically, if you look at how the big GPU cloud providers spec their inference instances, they almost always default to a stripped-down Ubuntu environment. Its less about the OS flavor and more about being in the ecosystem where NVIDIA and the DeepSeek devs are doing their primary testing.

sooo i totally get where youre coming from with the performance anxiety lol. i am still kinda a beginner with these deepseek models, but honestly, i am so happy with how my setup is running right now. i was really worried about breaking things or wasting money on parts that wouldnt play nice together, but i found that sticking to Fedora Workstation 40 has been a total lifesaver for me. it feels really professional and stable without being as scary or manual as arch, you know?

for your NVIDIA GeForce RTX 3090 24GB, my quick tip is basically this: focus on a distro that stays current with the kernel but isnt "bleeding edge" enough to break your cuda toolkit every Tuesday. i mean, i was terrified of kernel panics too! but i found that using a slightly more conservative approach with AlmaLinux 9.4 actually works really well if you want that enterprise-grade stability for long inference runs without any bloat slowing down your tps.

i havent done any crazy benchmarks, but i noticed my vram overhead stayed really low when i stopped using heavy desktop environments. maybe try a spin with a lightweight window manager? it feels safer for the hardware imo. anyway, i am super satisfied with how Fedora Workstation 40 handles the drivers automatically now. just be careful with the power draw when youre pushing deepseek-v3 hard... i always keep an eye on my temps just in case! good luck with the build!!

honestly man i have the exact same headache right now... been messing with this for over a month and still cant find a clear winner. i have swapped between like four different distros including stuff like NixOS just to see if the reproducibility helped, but the lack of solid benchmarks is just super disappointing. really sucks that we are still basically guessing which brand of linux actually squeezes the most out of a 3090 for these coder models.

Good to know!

Just saw this thread and honestly it drives me crazy how we're still dealing with this crap. I've been running Linux for years and trying to get a 3090 to play nice with DeepSeek or any LLM is still a total nightmare. It feels like such a scam that we spend all this money on hardware and then have to fight the OS every single day. Companies just dont care about the user experience once they have your money... its exhausting.

• Drivers breaking after every minor kernel update
• Random VRAM spikes that make no sense
• Repos that are always three versions behind what you actually need I've honestly had so many issues with this lately that I almost threw my rig out the window. It's not as good as it was supposed to be. But hey, dont let the frustration get to you too much. I'm here if you need to vent more or if you run into a specific error that makes you want to quit... we've all been there.