Which computer should I buy: Mac or custom-built 5090? [D]

A machine learning engineer's practical hardware dilemma has gone viral, framing a critical industry debate. The developer's workload is 70% fine-tuning large pretrained models (like Stable Diffusion or Llama) and 30% training models from scratch, primarily for image and video tasks. This split makes VRAM capacity—critical for loading large models—a top priority, challenging the conventional wisdom that NVIDIA GPUs are the only option for serious AI work. The post specifically questions if Apple's new MLX framework, designed to unify AI development across its Silicon chips, can make training on an M5 Max MacBook a viable alternative to a custom PC built around NVIDIA's anticipated RTX 5090.

The discussion highlights a shifting landscape. While NVIDIA's CUDA ecosystem remains the undisputed standard for peak training performance, Apple's unified memory architecture offers potentially massive VRAM advantages (up to 192GB in a Mac Studio) at a lower cost than professional NVIDIA cards like the RTX 6000 Ada. The community response is diving into real-world benchmarks for MLX, weighing the convenience and memory benefits of Apple Silicon against the raw speed, broader framework support (PyTorch, TensorFlow), and mature tooling of the NVIDIA stack. The cost barrier for high-end professional GPUs is a key driver, making the Mac a compelling 'good enough' option for many developers whose work hinges more on experimentation and fine-tuning than brute-force, large-scale training.