I'm upgrading the GPU in my mid-range system, which currently has a Ryzen 9 9950X3D, an RX 9060 XT 16GB, 32GB of DDR5-6000 CL38 memory, and an ASRock B850 RS PRO motherboard. I mainly use the computer for local LLMs, CPU-heavy games, and occasional AAA titles such as RDR2 and Cyberpunk at 1080p. My minimum requirement is 16GB of VRAM. I'm considering an RTX 5080 for about 950,000 tenge ($2,000), an RTX 5070 Ti for 675,000 tenge ($1,400), or an RX 9070 XT for 520,000 tenge ($1,100). I can afford any of them, but choosing the 5070 Ti or 9070 XT would leave enough money to upgrade to 64GB of RAM and buy other peripherals. My LLM use includes inference, some training or fine-tuning, and offloading workloads to system memory. How much of a practical advantage does NVIDIA's CUDA ecosystem provide over AMD for this kind of work? Would the RX 9070 XT be sufficient, or is the 5070 Ti or a used RTX 4090 a better all-around choice?
4 Answers
For gaming at 1080p, all three cards are excessive, so the decision really comes down to your AI software. The RTX 5070 Ti is probably the safest balanced choice because CUDA and other NVIDIA features tend to have broader support, while still costing considerably less than the 5080. It should handle essentially any current game comfortably, even at higher resolutions.
The RX 9070 XT should be perfectly usable for local inference through Vulkan or compatible software such as LM Studio. AMD can run models, but ROCm support and general compatibility are still less consistent than CUDA. For training and fine-tuning, NVIDIA is the safer option because many tools and tutorials target CUDA first. If you plan to offload larger models, spending the savings on 64GB of system RAM would also be worthwhile.
I do plan to offload some workloads to system RAM, and I also experiment with training even though the results are inconsistent. That is why I’m leaning more toward NVIDIA despite the higher price.
The 9070 XT is attractive for gaming value, but CUDA remains a significant advantage for an all-purpose AI workstation. If you are doing both inference and training rather than only experimenting with small models, I would choose the RTX 5070 Ti and use the remaining budget for 64GB of RAM. The RTX 5080 is difficult to justify at that premium unless you specifically need its extra performance, since it still does not solve the 16GB VRAM limitation.
If local LLMs are a serious priority, moving from one 16GB card to another 16GB card is not a major capacity upgrade. A used RTX 4090 or a card with substantially more VRAM would provide more room for larger models and reduce the need to offload. Just inspect a used card carefully and account for its age, power draw, warranty situation, and possible previous mining use.
A higher-VRAM option would be ideal, but some models with large memory capacities are unavailable or heavily marked up locally. In that situation, the used 4090 may offer the best AI performance per dollar if the seller and card condition are trustworthy.
A used RTX 4090 is also worth considering if it costs about the same as the 5080. Its 24GB of VRAM can be much more useful for LLMs, although its age, power consumption, previous mining use, and potential need for new thermal pads make it a riskier purchase.