Radeon 7900 XTX

[-]

Boring_Stress_2019@reddit

Hey mate, Give-me the exact tags from the Ollama models page and I will try-it for you. I have a 7900xtx running on a AMD Ryzen 7 5700 with 64 gigs of ram PC. OS = Fedora 41. Ollama = 0.5.7 Here are some results for the models I have installed: gemma2:27b = 36.49 tps granite3.1-dense:8b = 71.03 tps deepseek-r1:32b = 25.79 tps deepseek-r1:7b = 85.49 tps

Reply

[-]

danishkirel@reddit

Late to the party here but could you try long context prompts and report prompt processing speed? Interested in 12-16k context with 14b qwen3 and maybe 27b Gemma2 if that fits.

Reply

[-]

Boring_Stress_2019@reddit

Hello, if you tell me how to do that, I will do-it with pleasure, pls give-me the exact ollama command.

Reply

[-]

danishkirel@reddit

`ollama run --verbose qwen3:14b` In the ollama shell `/set parameter num_ctx 16000` Then copy https://gist.githubusercontent.com/kirel/fd69f04bfe54eed888fdbe96307a67e8/raw/9818e21cd101359f6c707de271fe5005c5adcb35/medium.txt and paste as prompt. Leave ollama prompt and verify with `ollama ps` that it’s running 100% on gpu. If not same with qwen3:8b? Paste metrics here. Thank you! Appreciated!

Reply

[-]

PolGZ@reddit

Very useful info, thank you very much!

Reply

[-]

danishkirel@reddit

Late to the party here but could you try long context prompts and report prompt processing speed? Interested in 12-16k context with 14b qwen3 and maybe 27b Gemma if that fits.

Reply

[-]

LoafyLemon@reddit

7900 XTX is 42% slower during inference. I have both GPUs.

Reply

[-]

danishkirel@reddit

Is that for inference or prompt processing or both? Interested in long context prompt processing speed comparison (like 16k tokens)

Reply

[-]

LoafyLemon@reddit

Both. Prompt processing is stupidly slow in comparison.

Reply

[-]

Ill_Faithlessness368@reddit

I have two 7900xtx. Running a benchmark for one of them with llama.cpp I get 120t/s for a 8B model. With my 3090 I get 150t/s. Running on Ubuntu 24.04. I got the 3099 6 months ago for $640 on eBay. Now the cheapest one is around 900. In my case it was cheaper to buy the new 7900xtx than the used 3090. I am hoping performance will improve over time for AMD cards.

Reply

[-]

Theboyscampus@reddit (OP)

Damn. So no Radeon for me then.

Reply

[-]

darth_chewbacca@reddit

7900xtx works very well on Linux for both inference and image gen. It struggles in video generation. For "new things" it might take some time to get optimized. Eg. Kokoro TTS when I first tried it was actually faster on the cpu than on the card (kokoro was incredibly fast on the CPU though)... This is partially why video gen struggles on amd cards. But if your doing well optimized tasks like LLM and sdxl/flux, it's slower than the 3090, but no so slow as to be "bad"

Reply

[-]

danishkirel@reddit

It struggles means performance or support?

Reply

[-]

sarkarian@reddit

Hey mate, I am considering buying a Radeon 7900 XTX for running AI models on local, just like you. Did you buy this GPU? How did it go? Keen to hear your experience before I pull the triger.

Reply

[-]

Theboyscampus@reddit (OP)

Nah I got a 3090 for 650€.

Reply

[-]

ETERNALBLADE47@reddit

AMD claims 7900 XTX runs deep seek better than 4090, but I doubted that. Has anyone tested?

Reply

[-]

ITMSPGuy@reddit

I just got a 7900xtx, for this same reason, its suposed to work well on lmstudio

Reply

[-]

NNN_Throwaway2@reddit

If you're able to get yours working well on the latest lmstudio I'd love to hear about it, because I haven't been able to get lmstudio to utilize my 7900 XTX properly since version 0.35.

Reply

[-]

maddogawl@reddit

Mine works great make sure you configure LMStudio to use the GPU. I forget where it is, but you have to enable GPU use and select the 7900xtx

Reply

[-]

NNN_Throwaway2@reddit

Its using the GPU. But the utilization is low and unstable and the performance is down since 0.35. When you say yours works "great" how do you define that?

Reply

[-]

b3081a@reddit

Probably need a driver update to 25.1.1

Reply

[-]

Thrumpwart@reddit

Use 25.1.1 Adrenaline driver, and make sure you've installed [ROCM](https://www.amd.com/en/developer/resources/rocm-hub/hip-sdk.html).

Reply

[-]

ITMSPGuy@reddit

Also i was tol to use the Pro Drivers (Blue) not the Adrenaline(Red), it has worked for me.

Reply

[-]

Thrumpwart@reddit

I've never tried with Pro, but the 25.1.1 adrenaline drivers work great for me.

Reply

[-]

mnze_brngo_7325@reddit

If you mostly do inference and think you can live without proper flash attention support, if you like to hand-compile and adjust libraries all the time instead of simply installing them, generally like to tinker a lot to get stuff working, if you don't need these nice and fancy finetuning frameworks, like unsloth, if you run linux and if you are ok with it being substantially slower than "comparable" nvidia cards, ... yeah, it's ok. I bought it kind of as a challenge and I get stuff done with it. But I don't think I'd do it again.

Reply

[-]

Thrumpwart@reddit

Not quite as fast as a 3090, but uses less power and is still quite fast. VRAM is king - the 24GB is great. With LM Studio, make sure you're using the 25.1.1 Adrenaline driver, and the latest [](6.2.4 ROCM ) if you're using Windows.

Reply

[-]

everardproudfoot@reddit

I run one. It rips for games, ok luck on LLM.

Reply

[-]

NNN_Throwaway2@reddit

I run a 7900 XTX and I say just stick with Nvidia to avoid the compatibility headaches. The performance is good on paper, sadly the reality is hit and miss.

Reply

[-]

linh1987@reddit

i think for inteferencing, 7900 xtx is slower (as much as half) of 3090. I don't have the source at hand but I have been doing research for last weeks

Reply to Post

30 Comments

Boring_Stress_2019@reddit

danishkirel@reddit

Boring_Stress_2019@reddit

danishkirel@reddit

PolGZ@reddit

danishkirel@reddit

LoafyLemon@reddit

danishkirel@reddit

LoafyLemon@reddit

Ill_Faithlessness368@reddit

Theboyscampus@reddit (OP)

darth_chewbacca@reddit

danishkirel@reddit

sarkarian@reddit

Theboyscampus@reddit (OP)

ETERNALBLADE47@reddit

ITMSPGuy@reddit

NNN_Throwaway2@reddit

maddogawl@reddit

NNN_Throwaway2@reddit

b3081a@reddit

Thrumpwart@reddit

ITMSPGuy@reddit

Thrumpwart@reddit

mnze_brngo_7325@reddit

Thrumpwart@reddit

everardproudfoot@reddit

NNN_Throwaway2@reddit

linh1987@reddit

Billy462@reddit