I'm planning to build a new PC soon with a Ryzen 9 9950X3D, RTX 5080 Gaming Trio OC, 64GB of DDR5-6000 CL30 memory, and 6TB of Samsung 990 Pro storage. My previous system had a Ryzen 3 3200G, RX 550, and 32GB of DDR4, so this will be a major upgrade. I originally considered an RTX 5090 for experimenting with local LLMs, but its price is around €6,000 where I live, which is far beyond my budget. How capable is the RTX 5080 for local AI models, including model size, quantization, context length, and inference speed? I'd also like to know how it performs with Minecraft shaders and modpacks, Factorio, Just Cause 4, Subnautica 2, Helldivers 2, CS2, and DOOM Eternal. Finally, how suitable is it for video editing, 3D rendering, image generation, and image enhancement?
4 Answers
The 5080 can be very good for editing, but 16GB of VRAM may become noticeable in high-resolution DaVinci Resolve projects, especially with complex effects, large timelines, or 4K footage. You may occasionally need proxies or optimized media. More VRAM would help with those workflows, but the card’s encoders, CUDA support, and overall compute performance still make it a strong editing and rendering GPU.
The RTX 5090 is considerably better for AI because its 32GB of VRAM lets larger models and longer contexts stay on the GPU. A 4090 or another high-VRAM card could also make more sense than a 5080 if local AI becomes your main priority. However, for gaming the 5080 is already extremely powerful, and it should run games such as Minecraft with demanding shaders, Helldivers 2, CS2, DOOM Eternal, Factorio, and similar titles comfortably. Whether it reaches your preferred settings and resolution will depend on the game and monitor, but gaming is not where this card is likely to disappoint.
For local models, 16GB of VRAM is workable but restrictive. Smaller and medium quantized models should run well, while larger models will need to use system RAM, which reduces performance and can make long context windows expensive in terms of memory. A Qwen-class 27B model at Q4 is a useful example: it may run, but not necessarily at high token rates on a single 5080. Mixture-of-experts models can be a better fit because only part of the model is active at once. Your 64GB of system RAM is reasonable, although 96GB or 128GB would provide more headroom for larger models.
If the budget allows, I may add another graphics card with 16GB of VRAM later, although my current motherboard and case plan is mainly built around using one GPU.
For gaming, it’s one of Nvidia’s fastest consumer cards and should handle essentially any current game very well, especially at 1440p and often at 4K with DLSS. It should be excellent for video editing, rendering, image generation, and enhancement as well. The main limitation is that it has 16GB of VRAM, so it is much less attractive for local AI than a card with 24GB or 32GB.
Since I’m just getting serious about local LLMs and am still a beginner, I’m thinking the 5080 should be more than enough to start. I can upgrade later if I actually need more VRAM.
Extra VRAM can matter more than raw GPU speed for some editing and AI workloads. A slower card with 24GB or 32GB may avoid out-of-memory problems, even if it renders individual operations more slowly.