Player not loading? Watch on YouTube
This video explores an AMD workstation for local AI, using a Ryzen Threadripper 9980X and Radeon AI Pro 9700 with 32 GB of VRAM. The speaker describes access arranged with Zidex and AMD, then tests chat models, image generation and Python workloads.
LM Studio is the first practical example. The speaker explains selecting its ROCm runtime and restarting to detect the GPU. He generally uses recommended 4-bit quantizations, with higher precision for smaller models. In the Qwen 3.6 mixture-of-experts example, he reports roughly 160 tokens per second and shows a context setting around 64K. These are results from his setup, rather than performance guarantees for other hardware or configurations. He also discusses Ollama for local LLM use and agent tasks.
The ComfyUI section covers its ROCm installation option and an image-generation example that changes a cat image's seed. The speaker also browses video, audio and image-to-3D model options.
For Python development, he recommends Linux and shows ROCm 7.2 with PyTorch. The Windows install selector shown does not offer that option; he also mentions WSL and dual boot. Examples include training a small ResNet on CIFAR-10 and running Gemma 4 through Transformers with a Gradio interface. Unsloth's AMD fine-tuning guide is discussed, but the video does not demonstrate an Unsloth training run.
