MiniMax homelab plan tests MI50 limits

The appealing part is price-per-VRAM; the painful part is everything around it. Cheap MI50s can make local frontier-ish inference possible, but they are not a clean substitute for newer NVIDIA stacks.

• –MiniMax M2.7 is a 230B MoE model with 10B active parameters, so memory math depends heavily on quantization, KV cache size, context length, and inference framework support
• –Four MI50s give 128GB of VRAM on paper, but sharding across older AMD cards means PCIe traffic, ROCm compatibility, and kernel support become first-order performance constraints
• –Prompt processing and long-context runs will likely feel much worse than the raw VRAM number suggests, especially without modern interconnects and optimized attention kernels
• –The build also needs serious chassis airflow, motherboard slot spacing, PSU capacity, and enough EPYC PCIe lanes to avoid turning the GPUs into an expensive space heater
• –A used 3090/4090-class NVIDIA setup may cost more per GB, but the software path for vLLM, SGLang, quantized builds, and debugging is still materially easier