I'm upgrading the GPU in my current system: a Ryzen 9 9950X3D, RX 9060 XT 16GB Gigabyte OC, 32GB DDR5-6000 CL38 memory, and an ASRock B850 RS Pro motherboard. I mainly use the PC for local LLMs, CPU-heavy games, and occasional AAA titles such as RDR2 and Cyberpunk at 1080p. My GPU options are 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 want at least 16GB of VRAM, which all three provide. I could buy any of them, but choosing the 5070 Ti or 9070 XT would leave enough money to upgrade my system memory to 64GB and buy other peripherals. My LLM workloads include inference, RAM offloading, and some fine-tuning or training, although training is still experimental for me. How much of a practical advantage does NVIDIA's CUDA ecosystem provide over AMD for these workloads? Would the RX 9070 XT be adequate, or would the RTX 5070 Ti—or even a used RTX 4090—be a better all-purpose choice? Prices are listed in Kazakhstani tenge, using roughly 470 tenge per US dollar.
3 Answers
For gaming alone, the RX 9070 XT is the value choice and should be more than sufficient at 1080p. For AI, however, ROCm and other AMD support can still be less consistent than CUDA depending on the tools and models you use. Since you’re doing inference, offloading, and some training rather than just gaming, the RTX 5070 Ti plus 64GB of RAM is a safer balanced setup.
If you’re doing inference through software such as LM Studio, the RX 9070 XT can be perfectly usable through Vulkan, and AMD support is improving. The situation is less straightforward for training, fine-tuning, and applications that expect CUDA, where NVIDIA generally has broader compatibility and fewer setup headaches. If you plan to offload part of the workload into system memory, 64GB of RAM would also be useful.
I do plan to offload larger models and I occasionally experiment with training. Most people I know recommend NVIDIA as the more reliable general-purpose AI option, which is why I’m reconsidering the extra cost.
The RTX 5070 Ti is probably the best compromise here. It gives you CUDA and NVIDIA’s AI-related features while costing considerably less than the 5080, leaving room for the 64GB memory upgrade. The RTX 5080 is excessive for 1080p gaming, and the 9070 XT is also more than enough for that resolution. Choose the 5080 only if its extra performance or specific software features will directly benefit your workloads.
That balance is what I’m leaning toward. I had not seriously considered a Mac with unified memory as a separate AI machine, but local availability and software compatibility would need to make sense first.