AMD Dual 7900 XTX Workstation Targets Local LLMs

This is a sensible cost-driven build for local AI, but it is still a compatibility project as much as a hardware purchase. My read is that Linux remains the safer default, while Windows is no longer the non-starter it used to be.

• –AMD now officially supports the RX 7900 XTX on ROCm for Linux, and its 2025 ROCm update added a public-preview Windows PyTorch stack for Radeon 7000 and 9000 series.
• –Two 24GB cards can make 35B-class inference viable, but the experience depends heavily on whether your serving stack can split work cleanly across GPUs.
• –The 128GB system RAM is useful for host-side buffering and larger contexts, but it does not replace VRAM or create a true pooled 48GB memory space for most workloads.
• –For coding models, the practical question is less raw specs and more software maturity: model loader, quantization format, kernel support, and multi-GPU behavior will determine whether this feels fast or fragile.
• –The post reflects the current local-LLM buying reality: use what is available and affordable, then choose the OS that minimizes setup and driver pain.