Nvidia GeForce GTX 1080

I wouldn't worry about hardware. I've run the latest local models over the last year, including the recent Qwen 3.6 30B A3B, on a 9yo GTX 1080 and 32G RAM I have lying around[0]. If I can do that I don't think hardware will be a problem for you in the near term. The only updates I've needed were to Llama.cpp when a new class of model was released. [0]: In my case, I want to see how local models perform on limited hardware, sacrificing context size and intelligence compared to SOTA models, so I have to really limit my expectations.

For me personally performance with Linux gaming is much less of a problem than it was with teenage me's Windows gaming. Back in the day I didn't have much money, but today I can just buy a slightly beefier GPU to compensate any possible performance loss from compatibility layers. With my 7 year old GTX 1080 and Ubuntu/Steam Proton I can still play most games at very high settings (except raytracing stuff).

Just to echo the other children of this comment; I'm pretty lost as to the reasoning here. I game at 1440p on a GTX 1080 from launch day in 2016. Cost is essentially no object for me, as gaming is my primary hobby and I can easily spend 60hrs/week playing video games, so the return would be more than justified in my mind. But I haven't upgraded in 6 years because there has been no reason to. I'm running my favorite games at 144fps, and the newer super-high fidelity stuff I can run at 60fps, which is more than reasonable. I will say that I was excited for the 40XX lineup, but at these prices I'm more inclined to hunt for a cheaper 2080 or 3080 solely for RTX functionality, but that's an enthusiast feature even for someone like me who spends a huge amount of time gaming. I couldn't imagine giving up my access to old titles, console emulators, mods, itch.io experiments, and low-budget indie games on Steam to save a few hundred bucks in the short-term to buy a console. YMMV if you're not paid well in a HCOL area, or if your location offers alternatives such as PC cafes.

"Of course it can't do anything useful for the amount of power that it consumes" Until about the top-end of the 3000 RTX series released, the GTX 1080 was still king of raster performance. I'm playing modern games like Marvel Rivals at 1440p no problem. Risk of Rain 2 4k no problem. STALKER 2 1080p no problem. It even runs most of my VR games perfectly fine, with me only needing to switch to my RTX 5060 machine for Into the Radius 2. Even today, it is a very capable card, despite the power draw.

"Nope. I watch a lotta youtube on a PC, FF runs vids like crap because Google changes something, and FF struggls to play well unlike Chrome or Edge." You must have a hardware problem, as I'm playing 4K videos on a 1st-gen Ryzen and GTX 1080 using Firefox and I've never had a single issue since I got the machine, around the 4th week the processor was available to purchase.

The ECC on the GTX 1080 let me figure out the optimal overclock with the only feedback mechanism being benchmarked throughout.

Funny enough, I have the 1080 and I tested qwen 3.5 9b on it today. I was positively surprised with the experience. It's no Opus but it did OK on the couple of queries and tasks I gave it. It's pretty impressive for a 10 year old gpu!

Agreed. Main box is 10+ years i5-4770K, I think. 16 GB RAM, GTX 1080. Runs Debian Trixie + KDE, Spotify, VS Code, docker, Steam, etc just fine.

I was surprised to find my old PC with a GTX1080 could transcribe/diarize about 10x faster than my m1 Mac. If anyone reading this is looking to transcribe 100s of hours of audio, do the extra work to get it set up on desktop with a dedicated graphics card.

Yes. I use https://heywillow.io/ and nothing ever leaves my LAN. It's entirely open, and relatively hackable - but also a pain to configure and requires 1 machine with a GPU (several years old is fine, I'm running mine on a nearly decade old Nvidia GTX 1080) and a couple satellite ESP32 devices (Espressif ESP32-S3-BOX-3). Home Assistant is also doing a more integrated version, but last I looked it was still "cloud-only" which defeats the purpose for me, since I was interested from a privacy standpoint. It does look like they finally released a local preview (at least for some languages) but I haven't yet tried them. By default, it still looks like they're using Google for TTS, which is a hard "NO!" from me. I've been pretty disappointed that they really seem to want to drive monetization with their voice features, and force people into the HA cloud services, but again - I haven't tried the latest release so take my frustrating with a grain of salt. They made a lot of noise in early 2023, and late last year still only supported cloud options... I was annoyed.

> The GPU almost steals your 1k budget by itself You can play almost any game on ancient GeForce 1080 (i do) with decent fps, except Starfield and maybe some other poorly optimized titles. On windows, that is. You can also play almost anything on SteamDeck (15W TDP max) on linux ($400 before discounts). You can't really play 3/4 of game catalog on MacBook Pro M1Pro or newer ($2000+)

The GTX 1080 has GDDR5X, which has ECC. I was able to dial in my memory OC on that card by increasing the clock then running memory throughput benchmarks until the throughput was maxed out. At the point where errors start the theroughput would decrease, but nothing would fail. Assuming stationary processes this is pretty nice. Maybe bump back 5% for margin.

I'm running an nVidia GTX1080 in an eGPU enclosure connected to my framework and it works pretty well. Have to disable the Intel Iris card in device manager to get games to use the 1080 and then re-enable it before you undock, but other than that, it's seamless. I love it.

I got a GTX1080 in december 2016. It's still going strong.

I started out using my old GTX 1080 on Thursday, could generate 512x512 just fine. That's in 8G of VRAM. It worked well on the hlky branch using webui (built using gradio). Seeing that training etc. is much more memory intensive, and wanting to get faster results, I bought an RTX 3090, which has 24G of VRAM. However it maxes out at about 1024x512, only twice as many pixels. Observing the card with GPUZ, it never actually allocates more than 13.9G. Using the lstein branch, I can't get above 896x512. Similarly, GPUZ shows allocated VRAM never reaches 14G. The interface isn't as good as the webui on hlky either - never mind the web interface, a bigger problem is it doesn't save all the parameters alongside generated images. This is all running using Miniconda on Windows. On Linux it may be a different story, but my gaming PC is not dual-boot (yet).

Anecdotally, my main PC runs Linux Mint 21, has an Nvidia GTX 1080, and installing the proprietary driver for it took me all of 5 minutes in the GUI Driver Manager tool. The FOSS nouveau driver was pre-installed initially but the GUI tool handled disabling it and enabling the closed source driver, then prompted me to reboot. Never had a problem with it since then, but I'm also not trying to live on the cutting edge with Wayland.

I remain skeptical that Apple's best GPU silicon will match nvidia's premiere products (either the top-end desktop card, or a server monster) for training. It seems like this is ideal as an accelerator for already trained models; one can imagine Photoshop utilizing it for deep-learning based infill-painting. I was doing training on battery with a laptop that had a 1080 and could do training; I have trained models on the airplane while totalyl unplugged and still had enough power to websurf afterwards.

Somewhere around late 2016 or early 2017, at which point most machines I used had SSDs for system drives and a cheap SoC was running video from the media array to my TV (a Pi3 at the time, now a Pi4 is doing the job as the 3 struggled with x265 at 1080p). I'd just put together a new main PC with a GTX1080(6Gb). The biggest upgrade in that build was the 1080, which is hardly taxed ATM as I've stopped playing much by way of fancy games (due to a mix of having picked up other hobbies and the rest of life getting in the way). I have upgraded my main desktop recently, late 2020 IIRC, because I started doing a bit of video recording on it. The old CPU was fast enough but the new one does in less than an hour what would have been an overnight task for the old. The old kit was handed down to improve the home server, mainly to give it more RAM rather than for speed (the old motherboard in there wouldn't take more than 16G, at least not officially). The other “fast enough” question is home Internet. For me that was probably 2011 when FTTC became available with 10mbit upstream (more is useful, but 10mbit is fast enough ). For downstream I'd say even earlier with 17mbit down (the 1.4mbit up I had then was much more limiting) and anything that didn't happen quickly at that rate wasn't urgent enough to leave running while I did something else. Of course what I do has grown but the connection has grown too so stayed in the “fast enough” range. I'd like faster either way (currently up to ~7…

The GTX 1080 is currently your best bet. It has the best price:performance ratio. The 1070 might also work, although it has slower cores and memory clock. I use a GTX 1080 for TensorFlow, it's ~100X faster than the CPU for training. I started with a 1060, which was way cheaper, but upgraded to the 1080 for more RAM and faster/more CUDA cores. If you can increase your budget and your motherboard supports it, you can get two 1080s, many things will parallelize across multiple cards. I know people who put 8 cards in a single machine and barely use the CPUs.

Less energy usage, more vram, faster in games (see https://www.pc-kombo.com/us/benchmark/games/gpu/compare?ids%... ). But it is even more about timing. Nvidia shifted to RTX afterwards, opting to go for raytracing and upscaling instead of making strictly faster cards. Which made the last GTX cards very competitive for regular usage, raising its status.

That's because 1080 and whole 10xx generation was pinacle and is the best GPU nvidia ever made. Nvidia won't make the same mistake any time soon.