Consumer GPUs to run LLMs

Not sure if this is the right place, if not please let me know.

GPU prices in the US have been a horrific bloodbath with the scalpers recently. So for this discussion, let's keep it to MSRP and the lucky people who actually managed to afford those insane MSRPs + managed to actually find the GPU they wanted.

Which GPU are you using to run what LLMs? How is the performance of the LLMs you have selected? On an average, what size of LLMs are you able to run smoothly on your GPU (7B, 14B, 20-24B etc).

What GPU do you recommend for decent amount of VRAM vs price (MSRP)? If you're using the TOTL RX 7900XTX/4090/5090 with 24+ GB of RAM, comment below with some performance estimations too.

My use-case: code assistants for Terraform + general shell and YAML, plain chat, some image generation. And to be able to still pay rent after spending all my savings on a GPU with a pathetic amount of VRAM (LOOKING AT BOTH OF YOU, BUT ESPECIALLY YOU NVIDIA YOU JERK). I would prefer to have GPUs for under $600 if possible, but I want to also run models like Mistral small so I suppose I don't have a choice but spend a huge sum of money.

Thanks

You can probably tell that I'm not very happy with the current PC consumer market but I decided to post in case we find any gems in the wild.

40 comments

Hopefully once Trump crashes economy we will see some bankruptcies and markets flooded with commercial GPUs as AI companies go under.

- I don't mind multiple GPUs but my motherboard doesn't have 2+ electrically connected X16 slots. I could build a new homeserver (I've been thinking about it) but consumer platforms simply don't have the PCIE lanes for 2 actual x16 slots. I'd have to go back to Broadwell Xeons for that, which are really power hungry. Oh well, I don't think it matters considering how power hungry GPUs are now.
  
  I haven't looked into the issue of PCIe lanes and the GPU.
  I don't think it should matter with a smaller PCIe bus, in theory, if I understand correctly (unlikely). The only time a lot of data is transferred is when the model layers are initially loaded. Like with Oobabooga when I load a model, most of the time my desktop RAM monitor widget does not even have the time to refresh and tell me how much memory was used on the CPU side. What is loaded in the GPU is around 90% static. I have a script that monitors this so that I can tune the maximum number of layers. I leave overhead room for the context to build up over time but there are no major changes happening aside from initial loading. One just sets the number of layers to offload on the GPU and loads the model. However many seconds that takes is irrelevant startup delay that only happens once when initiating the server.
  So assuming the kernel modules and hardware support the more narrow bandwidth, it should work... I think. There are laptops that have options for an external FireWire GPU too, so I don't think the PCIe bus is too baked in.

Using 7900XTX with LMS. Speed are everwhere, driver dependent. With QwQ-32B-Q4_K_M, I got about 20 tok/s, with all VRAM filled. Phi-4 runs at about 30-40 tok/s. I can give more numbers if you can wait for a bit.
If you don't enjoy finding which driver works best, I strongly aginst running AMD for AI workload.
- I didn't know that. I thought just one ROCM binary to install, run Ollama and that's it. Thanks for the explanation

I know you said consumer GPU, but I run a used Tesla P40. It has 24 GB of vram. The price has gone up since I got it a couple years ago, there might be better options in the same price category. Still, it's going to be cheaper than a modern full fat consumer gpu, with a reasonable performance hit.
My use case is text generation, chat kind of things. In most cases, the inference is more than fast enough, but it can get slow when swapping out large context lengths.
Mostly I run quantized 8-20B models with the sweet spot being around 12. For specialized use cases outside of general language, you can run more compact models. The general output is quite good, and I would have never had thought it was possible 10 years ago.
ETA: I paid about $200 USD for the P40 a couple years ago, plus the price for a fan and 3d printed shroud.

I recommend a used 3090, as that has 24 GB of VRAM and generally can be found for $800ish or less (at least when I last checked, in February). It’s much cheaper than a 4090 and while admittedly more expensive than the inexpensive 24GB Nvidia Tesla card (the P40?) it also has much better performance and CUDA support.
I have dual 3090s so my performance won’t translate directly to what a single GPU would get, but it’s pretty easy to find stats on 3090 performance.
- The 7900XTX was $1000 when it launched, I wouldn't mind it used either.

I'm currently looking for this as well. As far as my investigation went right now I'll probably go for 2x AMD Instinct MI50. Each of them has equivalent to slightly higher performance than a P40, however usually only 16gb VRAM (If you're super lucky you might get one with 32gb, those are usually not labeled as such though; probably binned MI60). With two of them you got 32gb VRAM and quite the performance for, right now, 200€ / card. Alternatively you should be able to run quantized models on a single card as well.
If you don't mind running ROCm instead of CUDA this seems like a good bang for the buck. Alternatively you might look into AMDs new line of "AI" SoCs (for example Frameworks Desktop computer). They seem to be really good as well, and depending on your usecase might be more useful than an equally priced 4090.
- Do you have 2 PCIE X16 slots on your motherboard (speaking in terms of electrical connections)?
  
  They would run with 8x speed each. Should not be too much of a bottleneck though, I don't expect the performance to suffer noticeably more than 5% from this. Annoying, but getting a CPU+Board with 32 lanes or more would throw off the price/performance ratio.

3090 has 24gb and rolls the 38b models beautifully.
- I tried to run Gemma 3 27B Q4K and was surprised how quickly the VRAM requirements blew up proportional to context window, especially compared to other models (all quantized) at similar size like Qwq 32B.
- Wait how does that work? How is 24GB enough for a 38B model?
  
  Look up “LLM quantization.” The idea is that each parameter is a number; by default they use 16 bits of precision, but if you scale them into smaller sizes, you use less space and have less precision, but you still have the same parameters. There’s not much quality loss going from 16 bits to 8, but it gets more noticeable as you get lower and lower. (That said, there’s are ternary bit models being trained from scratch that use 1.58 bits per parameter and are allegedly just as good as fp16 models of the same parameter count.)
  If you’re using a 4-bit quantization, then you need about half that number in VRAM. Q4_K_M is better than Q4, but also a bit larger. Ollama generally defaults to Q4_K_M. If you can handle a higher quantization, Q6_K is generally best. If you can’t quite fit it, Q5_K_M is generally better than any other option, followed by Q5_K_S.
  For example, Llama3.3 70B, which has 70.6 billion parameters, has the following sizes for some of its quantizations:
  q4_K_M (the default): 43 GB
  fp16: 141 GB
  q8: 75 GB
  q6_K: 58 GB
  q5_k_m: 50 GB
  q4: 40 GB
  q3_K_M: 34 GB
  q2_K: 26 GB
  This is why I run a lot of Q4_K_M 70B models on two 3090s.
  Generally speaking, there’s not a perceptible quality drop going to Q6_K from 8 bit quantization (though I have heard this is less true with MoE models). Below Q6, there’s a bit of a drop between it and 5 and then 4, but the model’s still decent. Below 4-bit quantizations you can generally get better results from a smaller parameter model at a higher quantization.
  TheBloke on Huggingface has a lot of GGUF quantization repos, and most, if not all of them, have a blurb about the different quantization types and which are recommended. When Ollama.com doesn’t have a model I want, I’m generally able to find one there.

I got it working with my 6800XT. I'm running deep seek r1 14b (somewhere around there) and the deep seek coder V2. I have a link to a blog with those instructions
https://gotosocial.michaeldileo.org/@mdileo/statuses/01JQA4M4Q33PMCADH9M2AWQSS8
- Thank you. Are 14B models the biggest you can run comfortably?
  
  The coder model has only that one. The ones bigger than that are like 20GB+, and my GPU has 16GB. I've only tried two models, but it looked like the size balloons after that, so that may be the biggest models that I can run.

Thing is, you can trade off speed for quality. For coding support you can settle for Llama 3.2 or a smaller deepseek-r1 and still get most of what you need on a smaller GPU, then scale up to a bigger model that will run slower if you need something cleaner. I've had a small laptop with 16 GB of total memory and a 4060 mobile serving as a makeshift home server with a LLM and a few other things and... well, it's not instant, but I can get the sort of thing you need out of it.
Sure, if I'm digging in and want something faster I can run something else in my bigger PC GPU, but a lot of the time I don't have to.
Like I said below, though, I'm in the process of trying to move that to an Arc A770 with 16 GB of VRAM that I had just lying around because I saw it on sale for a couple hundred bucks and I needed a temporary GPU replacement for a smaller PC. I've tried running LLMs on it before and it's not... super fast, but it'll do what you want for 14B models just fine. That's going to be your sweet spot on home GPUs anyway, anything larger than 16GB and you're talking 3090, 4090 or 5090, pretty much exclusively.

It all depends on the size of the model you are running, if it cannot fit in GPU memory, then it has to go back and forth with the host (cpu memory or even disk) and the GPU. This is extremely slow. This is why some people are running LLMs on macs, as they can have a large amount of memory shared between the GPU and CPU, making it viable to fit some larger models in memory.
- This is... mostly right, but I have to say, macs with 16 gigs of shared memory aren't all that, you can get many other alternatives with similar memory distributions, although not as fast.
  A bunch of vendors are starting to lean on this by providing small, weaker PCs with a BIG cache of shared RAM. That new Framework desktop with an AMD APU specs up to 128 GB of shared memory, while the mac minis everybody is hyping up for this cap at 24 GB instead.
  I'd strongly recommend starting with a mid-sized GPU on a desktop PC. Intel ships the A770 with 16GB of RAM and the B580 with 12 and they're both dirt cheap. You can still get a 3060 with 12 GB for similar prices, too. I'm not sure how they benchmark relative to each other on LLM tasks, but I'm sure one can look it up. Cheap as the entry level mac mini is, all of those are cheaper if you already have a PC up and running, and the total amount of dedicated RAM you get is very comparable.
  
  Oh, I thought you could get 128gb ram or more, but I can see it does not make sense with the <24gb… sorry for spreading misinformation, I guess, in this case a GPU of the same GB ram would probably be better

I would prefer to have GPUs for under $600 if possible
Unfortunately not possible for a new nvidia card (you want CUDA) with 16GB VRAM. You can get them for ~$750 if you're patient. This deal was available for awhile earlier today:
https://us-store.msi.com/Graphics-Cards/NVIDIA-GPU/GeForce-RTX-50-Series/GeForce-RTX-5070-Ti-16G-SHADOW-3X-OC
Or you could try to find a 16GB 4070Ti Super like I got. It runs Deepseek 14B and stuff like Stable Diffusion no problem.
- I am OK with either Nvidia or AMD especially if Ollama supports it. With that said I have heard that AMD takes some manual effort whilst Nvidia is easier. Depends on how difficult ROCM is
  
  With Ollama, all you have do is copy an extra folder of ROCm files. Not hard at all.