Mixed AMD GPUs power local coding LLMs

AMD's ROCm stack remains the "tinker's choice" for local LLMs, offering a high-performance alternative to NVIDIA if you're willing to handle the configuration overhead.

• –Combining disparate GPU architectures (Vega 64 and 6600 XT) requires specific environment overrides to ensure cross-compatibility in Ollama and PyTorch.
• –16GB of combined VRAM is the "sweet spot" for modern coding models like Qwen2.5-Coder 7B and DeepSeek-Coder-V2 16B, enabling low-latency autocomplete and complex logic.
• –Linux Mint 22.3 provides a stable base for the latest ROCm drivers, though using the built-in kernel driver via --no-dkms is often preferred for system stability.
• –Multi-GPU splitting in Ollama is seamless once drivers are configured, allowing developers to salvage older hardware for meaningful AI workloads.
• –The 64GB of system RAM provides a massive safety net for larger models, albeit at significantly lower token-per-second rates than dedicated VRAM.